CS11-747 Neural Networks for NLP Adversarial...
Transcript of CS11-747 Neural Networks for NLP Adversarial...
![Page 1: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/1.jpg)
CS11-747 Neural Networks for NLP
Adversarial MethodsGraham Neubig
Sitehttps://phontron.com/class/nn4nlp2019/
With many slides by Zihang Dai & Qizhe Xie
![Page 2: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/2.jpg)
Generative Models
P (X) =X
Z
P (X | Z)P (Z)<latexit sha1_base64="gzwAYFR/DfB073PQH17WZuyNuCg=">AAACBnicbZBPS8MwGMbT+W/Of1WPggSHsF1GK4J6EIZePE6wbmwtJU3TLSxpS5IKo+zmxa/ixYOKVz+DN7+N2daDbj4Q+OV535fkfYKUUaks69soLS2vrK6V1ysbm1vbO+bu3r1MMoGJgxOWiE6AJGE0Jo6iipFOKgjiASPtYHg9qbcfiJA0ie/UKCUeR/2YRhQjpS3fPGzVOnV4CV2Zcb8L9Q26nIawW9fcrftm1WpYU8FFsAuogkIt3/xywwRnnMQKMyRlz7ZS5eVIKIoZGVfcTJIU4SHqk57GGHEivXy6xxgeayeEUSL0iRWcur8ncsSlHPFAd3KkBnK+NjH/q/UyFZ17OY3TTJEYzx6KMgZVAiehwJAKghUbaUBYUP1XiAdIIKx0dBUdgj2/8iI4J42Lhn17Wm1eFWmUwQE4AjVggzPQBDegBRyAwSN4Bq/gzXgyXox342PWWjKKmX3wR8bnD0Lile8=</latexit><latexit sha1_base64="gzwAYFR/DfB073PQH17WZuyNuCg=">AAACBnicbZBPS8MwGMbT+W/Of1WPggSHsF1GK4J6EIZePE6wbmwtJU3TLSxpS5IKo+zmxa/ixYOKVz+DN7+N2daDbj4Q+OV535fkfYKUUaks69soLS2vrK6V1ysbm1vbO+bu3r1MMoGJgxOWiE6AJGE0Jo6iipFOKgjiASPtYHg9qbcfiJA0ie/UKCUeR/2YRhQjpS3fPGzVOnV4CV2Zcb8L9Q26nIawW9fcrftm1WpYU8FFsAuogkIt3/xywwRnnMQKMyRlz7ZS5eVIKIoZGVfcTJIU4SHqk57GGHEivXy6xxgeayeEUSL0iRWcur8ncsSlHPFAd3KkBnK+NjH/q/UyFZ17OY3TTJEYzx6KMgZVAiehwJAKghUbaUBYUP1XiAdIIKx0dBUdgj2/8iI4J42Lhn17Wm1eFWmUwQE4AjVggzPQBDegBRyAwSN4Bq/gzXgyXox342PWWjKKmX3wR8bnD0Lile8=</latexit><latexit sha1_base64="gzwAYFR/DfB073PQH17WZuyNuCg=">AAACBnicbZBPS8MwGMbT+W/Of1WPggSHsF1GK4J6EIZePE6wbmwtJU3TLSxpS5IKo+zmxa/ixYOKVz+DN7+N2daDbj4Q+OV535fkfYKUUaks69soLS2vrK6V1ysbm1vbO+bu3r1MMoGJgxOWiE6AJGE0Jo6iipFOKgjiASPtYHg9qbcfiJA0ie/UKCUeR/2YRhQjpS3fPGzVOnV4CV2Zcb8L9Q26nIawW9fcrftm1WpYU8FFsAuogkIt3/xywwRnnMQKMyRlz7ZS5eVIKIoZGVfcTJIU4SHqk57GGHEivXy6xxgeayeEUSL0iRWcur8ncsSlHPFAd3KkBnK+NjH/q/UyFZ17OY3TTJEYzx6KMgZVAiehwJAKghUbaUBYUP1XiAdIIKx0dBUdgj2/8iI4J42Lhn17Wm1eFWmUwQE4AjVggzPQBDegBRyAwSN4Bq/gzXgyXox342PWWjKKmX3wR8bnD0Lile8=</latexit><latexit sha1_base64="gzwAYFR/DfB073PQH17WZuyNuCg=">AAACBnicbZBPS8MwGMbT+W/Of1WPggSHsF1GK4J6EIZePE6wbmwtJU3TLSxpS5IKo+zmxa/ixYOKVz+DN7+N2daDbj4Q+OV535fkfYKUUaks69soLS2vrK6V1ysbm1vbO+bu3r1MMoGJgxOWiE6AJGE0Jo6iipFOKgjiASPtYHg9qbcfiJA0ie/UKCUeR/2YRhQjpS3fPGzVOnV4CV2Zcb8L9Q26nIawW9fcrftm1WpYU8FFsAuogkIt3/xywwRnnMQKMyRlz7ZS5eVIKIoZGVfcTJIU4SHqk57GGHEivXy6xxgeayeEUSL0iRWcur8ncsSlHPFAd3KkBnK+NjH/q/UyFZ17OY3TTJEYzx6KMgZVAiehwJAKghUbaUBYUP1XiAdIIKx0dBUdgj2/8iI4J42Lhn17Wm1eFWmUwQE4AjVggzPQBDegBRyAwSN4Bq/gzXgyXox342PWWjKKmX3wR8bnD0Lile8=</latexit>
![Page 3: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/3.jpg)
Generative Models
• Model a data distribution P(X) or a conditional one P(X|Y)
• Latent variable models: introduce another variable Z, and model
P (X) =X
Z
P (X | Z)P (Z)<latexit sha1_base64="gzwAYFR/DfB073PQH17WZuyNuCg=">AAACBnicbZBPS8MwGMbT+W/Of1WPggSHsF1GK4J6EIZePE6wbmwtJU3TLSxpS5IKo+zmxa/ixYOKVz+DN7+N2daDbj4Q+OV535fkfYKUUaks69soLS2vrK6V1ysbm1vbO+bu3r1MMoGJgxOWiE6AJGE0Jo6iipFOKgjiASPtYHg9qbcfiJA0ie/UKCUeR/2YRhQjpS3fPGzVOnV4CV2Zcb8L9Q26nIawW9fcrftm1WpYU8FFsAuogkIt3/xywwRnnMQKMyRlz7ZS5eVIKIoZGVfcTJIU4SHqk57GGHEivXy6xxgeayeEUSL0iRWcur8ncsSlHPFAd3KkBnK+NjH/q/UyFZ17OY3TTJEYzx6KMgZVAiehwJAKghUbaUBYUP1XiAdIIKx0dBUdgj2/8iI4J42Lhn17Wm1eFWmUwQE4AjVggzPQBDegBRyAwSN4Bq/gzXgyXox342PWWjKKmX3wR8bnD0Lile8=</latexit><latexit sha1_base64="gzwAYFR/DfB073PQH17WZuyNuCg=">AAACBnicbZBPS8MwGMbT+W/Of1WPggSHsF1GK4J6EIZePE6wbmwtJU3TLSxpS5IKo+zmxa/ixYOKVz+DN7+N2daDbj4Q+OV535fkfYKUUaks69soLS2vrK6V1ysbm1vbO+bu3r1MMoGJgxOWiE6AJGE0Jo6iipFOKgjiASPtYHg9qbcfiJA0ie/UKCUeR/2YRhQjpS3fPGzVOnV4CV2Zcb8L9Q26nIawW9fcrftm1WpYU8FFsAuogkIt3/xywwRnnMQKMyRlz7ZS5eVIKIoZGVfcTJIU4SHqk57GGHEivXy6xxgeayeEUSL0iRWcur8ncsSlHPFAd3KkBnK+NjH/q/UyFZ17OY3TTJEYzx6KMgZVAiehwJAKghUbaUBYUP1XiAdIIKx0dBUdgj2/8iI4J42Lhn17Wm1eFWmUwQE4AjVggzPQBDegBRyAwSN4Bq/gzXgyXox342PWWjKKmX3wR8bnD0Lile8=</latexit><latexit sha1_base64="gzwAYFR/DfB073PQH17WZuyNuCg=">AAACBnicbZBPS8MwGMbT+W/Of1WPggSHsF1GK4J6EIZePE6wbmwtJU3TLSxpS5IKo+zmxa/ixYOKVz+DN7+N2daDbj4Q+OV535fkfYKUUaks69soLS2vrK6V1ysbm1vbO+bu3r1MMoGJgxOWiE6AJGE0Jo6iipFOKgjiASPtYHg9qbcfiJA0ie/UKCUeR/2YRhQjpS3fPGzVOnV4CV2Zcb8L9Q26nIawW9fcrftm1WpYU8FFsAuogkIt3/xywwRnnMQKMyRlz7ZS5eVIKIoZGVfcTJIU4SHqk57GGHEivXy6xxgeayeEUSL0iRWcur8ncsSlHPFAd3KkBnK+NjH/q/UyFZ17OY3TTJEYzx6KMgZVAiehwJAKghUbaUBYUP1XiAdIIKx0dBUdgj2/8iI4J42Lhn17Wm1eFWmUwQE4AjVggzPQBDegBRyAwSN4Bq/gzXgyXox342PWWjKKmX3wR8bnD0Lile8=</latexit><latexit sha1_base64="gzwAYFR/DfB073PQH17WZuyNuCg=">AAACBnicbZBPS8MwGMbT+W/Of1WPggSHsF1GK4J6EIZePE6wbmwtJU3TLSxpS5IKo+zmxa/ixYOKVz+DN7+N2daDbj4Q+OV535fkfYKUUaks69soLS2vrK6V1ysbm1vbO+bu3r1MMoGJgxOWiE6AJGE0Jo6iipFOKgjiASPtYHg9qbcfiJA0ie/UKCUeR/2YRhQjpS3fPGzVOnV4CV2Zcb8L9Q26nIawW9fcrftm1WpYU8FFsAuogkIt3/xywwRnnMQKMyRlz7ZS5eVIKIoZGVfcTJIU4SHqk57GGHEivXy6xxgeayeEUSL0iRWcur8ncsSlHPFAd3KkBnK+NjH/q/UyFZ17OY3TTJEYzx6KMgZVAiehwJAKghUbaUBYUP1XiAdIIKx0dBUdgj2/8iI4J42Lhn17Wm1eFWmUwQE4AjVggzPQBDegBRyAwSN4Bq/gzXgyXox342PWWjKKmX3wR8bnD0Lile8=</latexit>
![Page 4: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/4.jpg)
What do we want from generative models?
• A “perfect” generative model
• Evaluate likelihood: P(x)
• e.g. Perplexity in language modeling
• Generate samples: x ~ P(X)
• e.g. Generate a sentence randomly from P(X) or conditioned on some other information using P(X|Y)
• Infer latent attributes: P(Z|X)
• e.g. Infer the “topic” of a sentence in topic models
![Page 5: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/5.jpg)
No Generative Model is Perfect (so far)
Likelihood
Generation (image)
Inference
Non-Latent VAE GAN
• Mostly rely on MLE (Lower bound) based training
• GANs are particularly good at generating continuous samples
![Page 6: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/6.jpg)
MLE vs. GAN
Image Credit: Lotter et al. 2015
![Page 7: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/7.jpg)
MLE vs. GAN• Over-emphasis of common outputs, fuzziness
Image Credit: Lotter et al. 2015
![Page 8: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/8.jpg)
MLE vs. GAN• Over-emphasis of common outputs, fuzziness
Real MLE Adversarial
Image Credit: Lotter et al. 2015
![Page 9: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/9.jpg)
MLE vs. GAN• Over-emphasis of common outputs, fuzziness
• Note: this is probably a good idea if you are doing maximum likelihood!
Real MLE Adversarial
Image Credit: Lotter et al. 2015
![Page 10: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/10.jpg)
Adversarial Training
![Page 11: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/11.jpg)
Adversarial Training
• Basic idea: create a “discriminator” that criticizes some aspect of the generated output
![Page 12: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/12.jpg)
Adversarial Training
• Basic idea: create a “discriminator” that criticizes some aspect of the generated output
• Generative adversarial networks: criticize the generated output
![Page 13: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/13.jpg)
Adversarial Training
• Basic idea: create a “discriminator” that criticizes some aspect of the generated output
• Generative adversarial networks: criticize the generated output
• Adversarial feature learning: criticize the generated features to find some trait
![Page 14: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/14.jpg)
Generative Adversarial Networks
![Page 15: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/15.jpg)
Basic Paradigm• Two players: generator and discriminator
• Discriminator: given an image, try to tell whether it is real or not → P(image is real)
• Generator: try to generate an image that fools the discriminator into answering “real”
• Desired result at convergence
• Generator: generate perfect image
• Discriminator: cannot tell the difference
![Page 16: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/16.jpg)
Training Method
xreal
sample minibatch
sample latent vars.
z
xfake
convert w/ generator
yreal
discriminator loss (higher if fail predictions)
generator loss (higher if correct predictions)
predict w/ discriminator
yfake
D gradientG gradient
![Page 17: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/17.jpg)
In Equations• Discriminator loss function:
• Make generated data “less fake” → Zero sum loss:
• Make generated data “more real” → Heuristic non-saturating loss:
• Latter gives better gradients when discriminator accurate
`D(✓D, ✓G) = �1
2Ex⇠Pdata logD(x)� 1
2Ez log(1�D(G(z)))
Predict fake for fake data
`G(✓D, ✓G) = �1
2Ez logD(G(z))
`G(✓D, ✓G) = �`D(✓D, ✓G)
• Generator loss function:
P(fake) = 1 - P(real)
Predict real for real data
![Page 18: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/18.jpg)
Interpretation: Distribution Matching
Process
• [Step1] Z ~ P(Z), P(Z) can be any distribution
• [Step2] X = F(Z), F is a deterministic function
Result
• X is a random variable with an implicit distribution P(X), which decided by both P(Z) and F
• The process can produce any complicated distribution P(X) with a reasonable P(Z) and a powerful enough F
Image Credit: He et al. 2018
z1<latexit sha1_base64="jbtr+W9vpRfg2saI7YImOblkNik=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcqMCOqu6MZlRccW2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7GN9m8/USVZjJ6gElMfYGHEQsZwWCo++e+26/WnIaTl70M3ALUUFGtfvW3N5AkETQCwrHWXdeJwU+xAkY4nVZ6iaYxJmM8pF0DIyyo9tPc6tQ+MczADqUyJwI7Z/9vpFhogWFklFnTc7OMASm5rhsVjETWsmfyu56IoB6IeiZSOtQLRiC89FMWxQnQiMx8hAm3QdpZLPaAKUqATwzARDHzFZuMsMIETHgVk5G7mMgy8M4aVw3n7rzWvC7CKqMjdIxOkYsuUBPdohbyEEFD9Ire0Lv1Yn1Yn9bXTFqyip1DNFfW9x/uTZqw</latexit><latexit sha1_base64="jbtr+W9vpRfg2saI7YImOblkNik=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcqMCOqu6MZlRccW2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7GN9m8/USVZjJ6gElMfYGHEQsZwWCo++e+26/WnIaTl70M3ALUUFGtfvW3N5AkETQCwrHWXdeJwU+xAkY4nVZ6iaYxJmM8pF0DIyyo9tPc6tQ+MczADqUyJwI7Z/9vpFhogWFklFnTc7OMASm5rhsVjETWsmfyu56IoB6IeiZSOtQLRiC89FMWxQnQiMx8hAm3QdpZLPaAKUqATwzARDHzFZuMsMIETHgVk5G7mMgy8M4aVw3n7rzWvC7CKqMjdIxOkYsuUBPdohbyEEFD9Ire0Lv1Yn1Yn9bXTFqyip1DNFfW9x/uTZqw</latexit><latexit sha1_base64="jbtr+W9vpRfg2saI7YImOblkNik=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcqMCOqu6MZlRccW2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7GN9m8/USVZjJ6gElMfYGHEQsZwWCo++e+26/WnIaTl70M3ALUUFGtfvW3N5AkETQCwrHWXdeJwU+xAkY4nVZ6iaYxJmM8pF0DIyyo9tPc6tQ+MczADqUyJwI7Z/9vpFhogWFklFnTc7OMASm5rhsVjETWsmfyu56IoB6IeiZSOtQLRiC89FMWxQnQiMx8hAm3QdpZLPaAKUqATwzARDHzFZuMsMIETHgVk5G7mMgy8M4aVw3n7rzWvC7CKqMjdIxOkYsuUBPdohbyEEFD9Ire0Lv1Yn1Yn9bXTFqyip1DNFfW9x/uTZqw</latexit>
z2<latexit sha1_base64="ZNKuRe23lzCJMNesh6cRgweKXJI=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3Lio4W2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DBy0TRahHJJeqE2BNOYuoBww47cSKYhFw+hiMr7P54xNVmsnoHiYx9QUeRixkBIOh7p77zX615jScvOxl4Baghopq96u/vYEkiaAREI617rpODH6KFTDC6bTSSzSNMRnjIe0aGGFBtZ/mVqf2iWEGdiiVORHYOft/I8VCCwwjo8yanptlDEjJdd2oYCSylj2T3/VEBPVA1DOR0qFeMALhhZ+yKE6ARmTmI0y4DdLOYrEHTFECfGIAJoqZr9hkhBUmYMKrmIzcxUSWgddsXDac27Na66oIq4yO0DE6RS46Ry10g9rIQwQN0St6Q+/Wi/VhfVpfM2nJKnYO0VxZ33/v75qx</latexit><latexit sha1_base64="ZNKuRe23lzCJMNesh6cRgweKXJI=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3Lio4W2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DBy0TRahHJJeqE2BNOYuoBww47cSKYhFw+hiMr7P54xNVmsnoHiYx9QUeRixkBIOh7p77zX615jScvOxl4Baghopq96u/vYEkiaAREI617rpODH6KFTDC6bTSSzSNMRnjIe0aGGFBtZ/mVqf2iWEGdiiVORHYOft/I8VCCwwjo8yanptlDEjJdd2oYCSylj2T3/VEBPVA1DOR0qFeMALhhZ+yKE6ARmTmI0y4DdLOYrEHTFECfGIAJoqZr9hkhBUmYMKrmIzcxUSWgddsXDac27Na66oIq4yO0DE6RS46Ry10g9rIQwQN0St6Q+/Wi/VhfVpfM2nJKnYO0VxZ33/v75qx</latexit><latexit sha1_base64="ZNKuRe23lzCJMNesh6cRgweKXJI=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3Lio4W2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DBy0TRahHJJeqE2BNOYuoBww47cSKYhFw+hiMr7P54xNVmsnoHiYx9QUeRixkBIOh7p77zX615jScvOxl4Baghopq96u/vYEkiaAREI617rpODH6KFTDC6bTSSzSNMRnjIe0aGGFBtZ/mVqf2iWEGdiiVORHYOft/I8VCCwwjo8yanptlDEjJdd2oYCSylj2T3/VEBPVA1DOR0qFeMALhhZ+yKE6ARmTmI0y4DdLOYrEHTFECfGIAJoqZr9hkhBUmYMKrmIzcxUSWgddsXDac27Na66oIq4yO0DE6RS46Ry10g9rIQwQN0St6Q+/Wi/VhfVpfM2nJKnYO0VxZ33/v75qx</latexit>
z3<latexit sha1_base64="VwYPdoeVjXkUU912kwmjdymaS+E=">AAACCHicZVBLTsMwFHT4lvIrsGQTUSGxqKoEkIBdBRuWRRBaqY0qx3Vaq3Yc2S9IJeoJEFs4ByvElltwDG6Ak2ZB2ydZbzxvnjWeIOZMg+P8WEvLK6tr66WN8ubW9s5uZW//UctEEeoRyaVqB1hTziLqAQNO27GiWASctoLRTTZvPVGlmYweYBxTX+BBxEJGMBjq/rl31qtUnbqTl70I3AJUUVHNXuW325ckETQCwrHWHdeJwU+xAkY4nZS7iaYxJiM8oB0DIyyo9tPc6sQ+NkzfDqUyJwI7Z/9vpFhogWFolFnTM7OMASm5rhkVDEXWsmfyux6LoBaIWiZSOtRzRiC89FMWxQnQiEx9hAm3QdpZLHafKUqAjw3ARDHzFZsMscIETHhlk5E7n8gi8E7rV3Xn7rzauC7CKqFDdIROkIsuUAPdoibyEEED9Ire0Lv1Yn1Yn9bXVLpkFTsHaKas7z/xkZqy</latexit><latexit sha1_base64="VwYPdoeVjXkUU912kwmjdymaS+E=">AAACCHicZVBLTsMwFHT4lvIrsGQTUSGxqKoEkIBdBRuWRRBaqY0qx3Vaq3Yc2S9IJeoJEFs4ByvElltwDG6Ak2ZB2ydZbzxvnjWeIOZMg+P8WEvLK6tr66WN8ubW9s5uZW//UctEEeoRyaVqB1hTziLqAQNO27GiWASctoLRTTZvPVGlmYweYBxTX+BBxEJGMBjq/rl31qtUnbqTl70I3AJUUVHNXuW325ckETQCwrHWHdeJwU+xAkY4nZS7iaYxJiM8oB0DIyyo9tPc6sQ+NkzfDqUyJwI7Z/9vpFhogWFolFnTM7OMASm5rhkVDEXWsmfyux6LoBaIWiZSOtRzRiC89FMWxQnQiEx9hAm3QdpZLHafKUqAjw3ARDHzFZsMscIETHhlk5E7n8gi8E7rV3Xn7rzauC7CKqFDdIROkIsuUAPdoibyEEED9Ire0Lv1Yn1Yn9bXVLpkFTsHaKas7z/xkZqy</latexit><latexit sha1_base64="VwYPdoeVjXkUU912kwmjdymaS+E=">AAACCHicZVBLTsMwFHT4lvIrsGQTUSGxqKoEkIBdBRuWRRBaqY0qx3Vaq3Yc2S9IJeoJEFs4ByvElltwDG6Ak2ZB2ydZbzxvnjWeIOZMg+P8WEvLK6tr66WN8ubW9s5uZW//UctEEeoRyaVqB1hTziLqAQNO27GiWASctoLRTTZvPVGlmYweYBxTX+BBxEJGMBjq/rl31qtUnbqTl70I3AJUUVHNXuW325ckETQCwrHWHdeJwU+xAkY4nZS7iaYxJiM8oB0DIyyo9tPc6sQ+NkzfDqUyJwI7Z/9vpFhogWFolFnTM7OMASm5rhkVDEXWsmfyux6LoBaIWiZSOtRzRiC89FMWxQnQiEx9hAm3QdpZLHafKUqAjw3ARDHzFZsMscIETHhlk5E7n8gi8E7rV3Xn7rzauC7CKqFDdIROkIsuUAPdoibyEEED9Ire0Lv1Yn1Yn9bXVLpkFTsHaKas7z/xkZqy</latexit>
x3<latexit sha1_base64="r7RTsVKDbYXIRCyPj9szuInKMVA=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcqMCuqu6MZlBccW2qFk0kwbmkyG5I5Yhv6DuNXvcCVu/QY/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7Gt9m8/USVZjJ6gElMfYGHEQsZwWCoTi8Q6fO0f96v1pyGk5e9DNwC1FBRrX71tzeQJBE0AsKx1l3XicFPsQJGOJ1WeommMSZjPKRdAyMsqPbT3O/UPjHMwA6lMicCO2f/b6RYaIFhZJRZ03OzjAEpua4bFYxE1rJn8rueiKAeiHomUjrUC0YgvPJTFsUJ0IjMfIQJt0HaWTb2gClKgE8MwEQx8xWbjLDCBEyCFZORu5jIMvDOGtcN5/6i1rwpwiqjI3SMTpGLLlET3aEW8hBBHL2iN/RuvVgf1qf1NZOWrGLnEM2V9f0HLzudBQ==</latexit><latexit sha1_base64="r7RTsVKDbYXIRCyPj9szuInKMVA=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcqMCuqu6MZlBccW2qFk0kwbmkyG5I5Yhv6DuNXvcCVu/QY/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7Gt9m8/USVZjJ6gElMfYGHEQsZwWCoTi8Q6fO0f96v1pyGk5e9DNwC1FBRrX71tzeQJBE0AsKx1l3XicFPsQJGOJ1WeommMSZjPKRdAyMsqPbT3O/UPjHMwA6lMicCO2f/b6RYaIFhZJRZ03OzjAEpua4bFYxE1rJn8rueiKAeiHomUjrUC0YgvPJTFsUJ0IjMfIQJt0HaWTb2gClKgE8MwEQx8xWbjLDCBEyCFZORu5jIMvDOGtcN5/6i1rwpwiqjI3SMTpGLLlET3aEW8hBBHL2iN/RuvVgf1qf1NZOWrGLnEM2V9f0HLzudBQ==</latexit><latexit sha1_base64="r7RTsVKDbYXIRCyPj9szuInKMVA=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcqMCuqu6MZlBccW2qFk0kwbmkyG5I5Yhv6DuNXvcCVu/QY/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7Gt9m8/USVZjJ6gElMfYGHEQsZwWCoTi8Q6fO0f96v1pyGk5e9DNwC1FBRrX71tzeQJBE0AsKx1l3XicFPsQJGOJ1WeommMSZjPKRdAyMsqPbT3O/UPjHMwA6lMicCO2f/b6RYaIFhZJRZ03OzjAEpua4bFYxE1rJn8rueiKAeiHomUjrUC0YgvPJTFsUJ0IjMfIQJt0HaWTb2gClKgE8MwEQx8xWbjLDCBEyCFZORu5jIMvDOGtcN5/6i1rwpwiqjI3SMTpGLLlET3aEW8hBBHL2iN/RuvVgf1qf1NZOWrGLnEM2V9f0HLzudBQ==</latexit>
x2<latexit sha1_base64="OUe9NQ3ny9IilmHPU3ONQ4TtfRo=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3LCo4ttEPJpJk2NJkMyR2xDP0Hcavf4Urc+g1+hn9gZjoL214I9+Tcc8PJCWLONDjOj1VaW9/Y3CpvV3Z29/YPqodHj1omilCPSC5VN8CachZRDxhw2o0VxSLgtBNMbrN554kqzWT0ANOY+gKPIhYygsFQ3X4g0ufZoDmo1pyGk5e9CtwC1FBR7UH1tz+UJBE0AsKx1j3XicFPsQJGOJ1V+ommMSYTPKI9AyMsqPbT3O/MPjPM0A6lMicCO2f/b6RYaIFhbJRZ0wuzjAEpua4bFYxF1rJn8rueiqAeiHomUjrUS0YgvPJTFsUJ0IjMfYQJt0HaWTb2kClKgE8NwEQx8xWbjLHCBEyCFZORu5zIKvCajeuGc39Ra90UYZXRCTpF58hFl6iF7lAbeYggjl7RG3q3XqwP69P6mktLVrFzjBbK+v4DLZmdBA==</latexit><latexit sha1_base64="OUe9NQ3ny9IilmHPU3ONQ4TtfRo=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3LCo4ttEPJpJk2NJkMyR2xDP0Hcavf4Urc+g1+hn9gZjoL214I9+Tcc8PJCWLONDjOj1VaW9/Y3CpvV3Z29/YPqodHj1omilCPSC5VN8CachZRDxhw2o0VxSLgtBNMbrN554kqzWT0ANOY+gKPIhYygsFQ3X4g0ufZoDmo1pyGk5e9CtwC1FBR7UH1tz+UJBE0AsKx1j3XicFPsQJGOJ1V+ommMSYTPKI9AyMsqPbT3O/MPjPM0A6lMicCO2f/b6RYaIFhbJRZ0wuzjAEpua4bFYxF1rJn8rueiqAeiHomUjrUS0YgvPJTFsUJ0IjMfYQJt0HaWTb2kClKgE8NwEQx8xWbjLHCBEyCFZORu5zIKvCajeuGc39Ra90UYZXRCTpF58hFl6iF7lAbeYggjl7RG3q3XqwP69P6mktLVrFzjBbK+v4DLZmdBA==</latexit><latexit sha1_base64="OUe9NQ3ny9IilmHPU3ONQ4TtfRo=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3LCo4ttEPJpJk2NJkMyR2xDP0Hcavf4Urc+g1+hn9gZjoL214I9+Tcc8PJCWLONDjOj1VaW9/Y3CpvV3Z29/YPqodHj1omilCPSC5VN8CachZRDxhw2o0VxSLgtBNMbrN554kqzWT0ANOY+gKPIhYygsFQ3X4g0ufZoDmo1pyGk5e9CtwC1FBR7UH1tz+UJBE0AsKx1j3XicFPsQJGOJ1V+ommMSYTPKI9AyMsqPbT3O/MPjPM0A6lMicCO2f/b6RYaIFhbJRZ0wuzjAEpua4bFYxF1rJn8rueiqAeiHomUjrUS0YgvPJTFsUJ0IjMfYQJt0HaWTb2kClKgE8NwEQx8xWbjLHCBEyCFZORu5zIKvCajeuGc39Ra90UYZXRCTpF58hFl6iF7lAbeYggjl7RG3q3XqwP69P6mktLVrFzjBbK+v4DLZmdBA==</latexit>
x1<latexit sha1_base64="wPlw8y/0IqfHJrazcSQaHuwQ07k=">AAACDXicZVDLSgMxFM34rPVVdekmWAQXpcyIoO6KblxWcGyhLSWTZtrQZDIkd8Qy9B/ErX6HK3HrN/gZ/oGZ6SxseyHck3PPDScniAU34Lo/zsrq2vrGZmmrvL2zu7dfOTh8NCrRlPlUCaXbATFM8Ij5wEGwdqwZkYFgrWB8m81bT0wbrqIHmMSsJ8kw4iGnBCzV7gYyfZ72vX6l6tbdvPAy8ApQRUU1+5Xf7kDRRLIIqCDGdDw3hl5KNHAq2LTcTQyLCR2TIetYGBHJTC/N/U7xqWUGOFTanghwzv7fSIk0ksDIKrNm5mYZA0oJU7MqGMmsZc/kdzORQS2QtUykTWgWjEB41Ut5FCfAIjrzESYCg8JZNnjANaMgJhYQqrn9CqYjogkFm2DZZuQtJrIM/PP6dd29v6g2boqwSugYnaAz5KFL1EB3qIl8RJFAr+gNvTsvzofz6XzNpCtOsXOE5sr5/gMr950D</latexit><latexit sha1_base64="wPlw8y/0IqfHJrazcSQaHuwQ07k=">AAACDXicZVDLSgMxFM34rPVVdekmWAQXpcyIoO6KblxWcGyhLSWTZtrQZDIkd8Qy9B/ErX6HK3HrN/gZ/oGZ6SxseyHck3PPDScniAU34Lo/zsrq2vrGZmmrvL2zu7dfOTh8NCrRlPlUCaXbATFM8Ij5wEGwdqwZkYFgrWB8m81bT0wbrqIHmMSsJ8kw4iGnBCzV7gYyfZ72vX6l6tbdvPAy8ApQRUU1+5Xf7kDRRLIIqCDGdDw3hl5KNHAq2LTcTQyLCR2TIetYGBHJTC/N/U7xqWUGOFTanghwzv7fSIk0ksDIKrNm5mYZA0oJU7MqGMmsZc/kdzORQS2QtUykTWgWjEB41Ut5FCfAIjrzESYCg8JZNnjANaMgJhYQqrn9CqYjogkFm2DZZuQtJrIM/PP6dd29v6g2boqwSugYnaAz5KFL1EB3qIl8RJFAr+gNvTsvzofz6XzNpCtOsXOE5sr5/gMr950D</latexit><latexit sha1_base64="wPlw8y/0IqfHJrazcSQaHuwQ07k=">AAACDXicZVDLSgMxFM34rPVVdekmWAQXpcyIoO6KblxWcGyhLSWTZtrQZDIkd8Qy9B/ErX6HK3HrN/gZ/oGZ6SxseyHck3PPDScniAU34Lo/zsrq2vrGZmmrvL2zu7dfOTh8NCrRlPlUCaXbATFM8Ij5wEGwdqwZkYFgrWB8m81bT0wbrqIHmMSsJ8kw4iGnBCzV7gYyfZ72vX6l6tbdvPAy8ApQRUU1+5Xf7kDRRLIIqCDGdDw3hl5KNHAq2LTcTQyLCR2TIetYGBHJTC/N/U7xqWUGOFTanghwzv7fSIk0ksDIKrNm5mYZA0oJU7MqGMmsZc/kdzORQS2QtUykTWgWjEB41Ut5FCfAIjrzESYCg8JZNnjANaMgJhYQqrn9CqYjogkFm2DZZuQtJrIM/PP6dd29v6g2boqwSugYnaAz5KFL1EB3qIl8RJFAr+gNvTsvzofz6XzNpCtOsXOE5sr5/gMr950D</latexit>
{<latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit><latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit><latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit>
Symbo
lic
Mod
el
{<latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit><latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit><latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit>
Neu
ral
Projec
tor
e1<latexit sha1_base64="WgitCLnES1r4TeJXeVDWd7cTvpM=">AAACDXicZVDLSgMxFM3UV62vqks3wSK4KGVGBHVXdOOygmMLbSmZNNOGJpMhuSOUof8gbvU7XIlbv8HP8A/MTGdh2wvhnpx7bjg5QSy4Adf9cUpr6xubW+Xtys7u3v5B9fDoyahEU+ZTJZTuBMQwwSPmAwfBOrFmRAaCtYPJXTZvPzNtuIoeYRqzviSjiIecErBUpxfIlM0G3qBacxtuXngVeAWooaJag+pvb6hoIlkEVBBjup4bQz8lGjgVbFbpJYbFhE7IiHUtjIhkpp/mfmf4zDJDHCptTwQ4Z/9vpEQaSWBslVkzC7OMAaWEqVsVjGXWsmfyu5nKoB7IeibSJjRLRiC87qc8ihNgEZ37CBOBQeEsGzzkmlEQUwsI1dx+BdMx0YSCTbBiM/KWE1kF/kXjpuE+XNaat0VYZXSCTtE58tAVaqJ71EI+okigV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAMuJzw</latexit><latexit sha1_base64="WgitCLnES1r4TeJXeVDWd7cTvpM=">AAACDXicZVDLSgMxFM3UV62vqks3wSK4KGVGBHVXdOOygmMLbSmZNNOGJpMhuSOUof8gbvU7XIlbv8HP8A/MTGdh2wvhnpx7bjg5QSy4Adf9cUpr6xubW+Xtys7u3v5B9fDoyahEU+ZTJZTuBMQwwSPmAwfBOrFmRAaCtYPJXTZvPzNtuIoeYRqzviSjiIecErBUpxfIlM0G3qBacxtuXngVeAWooaJag+pvb6hoIlkEVBBjup4bQz8lGjgVbFbpJYbFhE7IiHUtjIhkpp/mfmf4zDJDHCptTwQ4Z/9vpEQaSWBslVkzC7OMAaWEqVsVjGXWsmfyu5nKoB7IeibSJjRLRiC87qc8ihNgEZ37CBOBQeEsGzzkmlEQUwsI1dx+BdMx0YSCTbBiM/KWE1kF/kXjpuE+XNaat0VYZXSCTtE58tAVaqJ71EI+okigV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAMuJzw</latexit><latexit sha1_base64="WgitCLnES1r4TeJXeVDWd7cTvpM=">AAACDXicZVDLSgMxFM3UV62vqks3wSK4KGVGBHVXdOOygmMLbSmZNNOGJpMhuSOUof8gbvU7XIlbv8HP8A/MTGdh2wvhnpx7bjg5QSy4Adf9cUpr6xubW+Xtys7u3v5B9fDoyahEU+ZTJZTuBMQwwSPmAwfBOrFmRAaCtYPJXTZvPzNtuIoeYRqzviSjiIecErBUpxfIlM0G3qBacxtuXngVeAWooaJag+pvb6hoIlkEVBBjup4bQz8lGjgVbFbpJYbFhE7IiHUtjIhkpp/mfmf4zDJDHCptTwQ4Z/9vpEQaSWBslVkzC7OMAaWEqVsVjGXWsmfyu5nKoB7IeibSJjRLRiC87qc8ihNgEZ37CBOBQeEsGzzkmlEQUwsI1dx+BdMx0YSCTbBiM/KWE1kF/kXjpuE+XNaat0VYZXSCTtE58tAVaqJ71EI+okigV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAMuJzw</latexit>
e2<latexit sha1_base64="wBcNk6euWGMrg+Da+mnoLSs6/Fg=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcq0COqu6MZlBUcL7VAyaaYNzWNI7ghl6D+IW/0OV+LWb/Az/AMz01nY9kK4J+eeG05OGHNmwPN+nNLa+sbmVnm7srO7t39QPTx6NCrRhPpEcaW7ITaUM0l9YMBpN9YUi5DTp3Bym82fnqk2TMkHmMY0EHgkWcQIBkt1+6FI6WzQGlRrXsPLy10FzQLUUFGdQfW3P1QkEVQC4diYXtOLIUixBkY4nVX6iaExJhM8oj0LJRbUBGnud+aeWWboRkrbI8HN2f8bKRZGYBhbZdbMwixjQClu6lYFY5G17Jn8bqYirIeinom0icySEYiugpTJOAEqydxHlHAXlJtl4w6ZpgT41AJMNLNfcckYa0zAJlixGTWXE1kFfqtx3fDuL2rtmyKsMjpBp+gcNdElaqM71EE+IoijV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAOWpzx</latexit><latexit sha1_base64="wBcNk6euWGMrg+Da+mnoLSs6/Fg=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcq0COqu6MZlBUcL7VAyaaYNzWNI7ghl6D+IW/0OV+LWb/Az/AMz01nY9kK4J+eeG05OGHNmwPN+nNLa+sbmVnm7srO7t39QPTx6NCrRhPpEcaW7ITaUM0l9YMBpN9YUi5DTp3Bym82fnqk2TMkHmMY0EHgkWcQIBkt1+6FI6WzQGlRrXsPLy10FzQLUUFGdQfW3P1QkEVQC4diYXtOLIUixBkY4nVX6iaExJhM8oj0LJRbUBGnud+aeWWboRkrbI8HN2f8bKRZGYBhbZdbMwixjQClu6lYFY5G17Jn8bqYirIeinom0icySEYiugpTJOAEqydxHlHAXlJtl4w6ZpgT41AJMNLNfcckYa0zAJlixGTWXE1kFfqtx3fDuL2rtmyKsMjpBp+gcNdElaqM71EE+IoijV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAOWpzx</latexit><latexit sha1_base64="wBcNk6euWGMrg+Da+mnoLSs6/Fg=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcq0COqu6MZlBUcL7VAyaaYNzWNI7ghl6D+IW/0OV+LWb/Az/AMz01nY9kK4J+eeG05OGHNmwPN+nNLa+sbmVnm7srO7t39QPTx6NCrRhPpEcaW7ITaUM0l9YMBpN9YUi5DTp3Bym82fnqk2TMkHmMY0EHgkWcQIBkt1+6FI6WzQGlRrXsPLy10FzQLUUFGdQfW3P1QkEVQC4diYXtOLIUixBkY4nVX6iaExJhM8oj0LJRbUBGnud+aeWWboRkrbI8HN2f8bKRZGYBhbZdbMwixjQClu6lYFY5G17Jn8bqYirIeinom0icySEYiugpTJOAEqydxHlHAXlJtl4w6ZpgT41AJMNLNfcckYa0zAJlixGTWXE1kFfqtx3fDuL2rtmyKsMjpBp+gcNdElaqM71EE+IoijV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAOWpzx</latexit>
e3<latexit sha1_base64="LgiizHRmhHOKEEbwqVR4tWBS3lw=">AAACDXicZVDLTgIxFL2DL8QXauLGzURi4oKQQRfqjujGJSQiJECwUzrQ0E4n7R0TMuEfjFv9DlfGrd/gxn/wD+wAC4GbNPf03HOb0+NHghv0vG8ns7K6tr6R3cxtbe/s7uX3Dx6MijVldaqE0k2fGCZ4yOrIUbBmpBmRvmANf3ibzhtPTBuuwnscRawjST/kAacELdVs+zJh4+5FN1/wSt6k3GVQnoFC5aj28wgA1W7+t91TNJYsRCqIMa2yF2EnIRo5FWyca8eGRYQOSZ+1LAyJZKaTTPyO3VPL9NxAaXtCdCfs/42ESCMJDqwybWZuljKolDBFq8KBTFv6zORuRtIv+rKYirQJzIIRDK46CQ+jGFlIpz6CWLio3DQbt8c1oyhGFhCquf2KSwdEE4o2wZzNqLyYyDKon5euS17NZnUD08rCMZzAGZThEipwB1WoAwUBL/AKb86z8+58OJ9TacaZ7RzCXDlff9YqntM=</latexit><latexit sha1_base64="RMNkU799kfoHwXHS6vi9XiRQ2tw=">AAACDXicZVDLSgMxFM3UV62vquDGzWARXJQy1YW6K3XjsgXHFtqhZNJMG5pMhuSOUIb+g7jVX3DrSlwJfoMb/8E/MDPtwrYXwj0599xwcvyIMw2O823lVlbX1jfym4Wt7Z3dveL+wb2WsSLUJZJL1faxppyF1AUGnLYjRbHwOW35o5t03nqgSjMZ3sE4op7Ag5AFjGAwVLvri4ROehe9YsmpOFnZy6A6A6XaUfOHvdY/G73ib7cvSSxoCIRjrTtVJwIvwQoY4XRS6MaaRpiM8IB2DAyxoNpLMr8T+9QwfTuQypwQ7Iz9v5FgoQWGoVGmTc/NUgak5LpsVDAUaUufye56LPyyL8qpSOlALxiB4MpLWBjFQEMy9RHE3AZpp9nYfaYoAT42ABPFzFdsMsQKEzAJFkxG1cVEloF7XrmuOE2TVR1NK4+O0Qk6Q1V0iWroFjWQiwji6Ak9oxfr0Xqz3q2PqTRnzXYO0VxZX39bgqCP</latexit><latexit sha1_base64="RMNkU799kfoHwXHS6vi9XiRQ2tw=">AAACDXicZVDLSgMxFM3UV62vquDGzWARXJQy1YW6K3XjsgXHFtqhZNJMG5pMhuSOUIb+g7jVX3DrSlwJfoMb/8E/MDPtwrYXwj0599xwcvyIMw2O823lVlbX1jfym4Wt7Z3dveL+wb2WsSLUJZJL1faxppyF1AUGnLYjRbHwOW35o5t03nqgSjMZ3sE4op7Ag5AFjGAwVLvri4ROehe9YsmpOFnZy6A6A6XaUfOHvdY/G73ib7cvSSxoCIRjrTtVJwIvwQoY4XRS6MaaRpiM8IB2DAyxoNpLMr8T+9QwfTuQypwQ7Iz9v5FgoQWGoVGmTc/NUgak5LpsVDAUaUufye56LPyyL8qpSOlALxiB4MpLWBjFQEMy9RHE3AZpp9nYfaYoAT42ABPFzFdsMsQKEzAJFkxG1cVEloF7XrmuOE2TVR1NK4+O0Qk6Q1V0iWroFjWQiwji6Ak9oxfr0Xqz3q2PqTRnzXYO0VxZX39bgqCP</latexit>
f�(e)<latexit sha1_base64="8oxgnabmEamXCM6ELWdhWTrrt8c=">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</latexit><latexit sha1_base64="JbCPJ2fHAGZa8Ebw6DTj/vvFjWE=">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</latexit><latexit sha1_base64="JbCPJ2fHAGZa8Ebw6DTj/vvFjWE=">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</latexit>
zi ⇠ Symbolic Model<latexit sha1_base64="qRL/AxTFpgdIbBxTWc/weaeer2U=">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</latexit><latexit sha1_base64="Dnc6JOWSxrJAoiQOjwnmnyt3/Tw=">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</latexit><latexit sha1_base64="Dnc6JOWSxrJAoiQOjwnmnyt3/Tw=">AAACJXicZVA9SwNBEN3zM8avqKUii0GwCOFio3ZBGxsholHBhLC3mTOLu7fH7pwYj3T6X8RWC3+FlQhW/gAr/4F7iYXRgWXevnkzzLwglsKi7797I6Nj4xOTuan89Mzs3HxhYfHE6sRwqHMttTkLmAUpIqijQAlnsQGmAgmnweVeVj+9AmOFjo6xG0NTsYtIhIIzdFSrsHrTErRhhaINhGtMj7oq0FJweqDbIHutQtEv+/2g/0HlBxSrK8+Hn3erz7VW4avR1jxRECGXzNrzih9jM2UGBZfQyzcSCzHjl+wCzh2MmALbTPt39Oi6Y9o01Ma9CGmf/d2RMmUVw45TZskO1TIGtZa25FTYUVnKxvT/1p1VClQpExkb2j+LYLjdTEUUJwgRH+wRJpKipplntC0McJRdBxg3wp1CeYcZxtE5m3ceVf468h/UN8s7Zf/QebVLBpEjy2SNbJAK2SJVsk9qpE44uSUP5JE8effei/fqvQ2kI95PzxIZCu/jG4oiqh8=</latexit>
ei ⇠ N (µzi ,⌃zi)<latexit sha1_base64="8CLMr+o0dZvNiJKf88rFwAP//jQ=">AAACQ3icZVBNa9tAEB0laZO4H3GaYy9LQiGlxki5JIUcQnvJKaS0bgyWEKv1yl6yqxW7o4Ir9F9Kfkzptb31P+TS0Guhu7IPiT2wzJs3b5aZl5VSWAzD38Ha+sajx5tb250nT5893+nuvvhsdWUYHzAttRlm1HIpCj5AgZIPS8OpyiS/yq7f+/7VF26s0MUnnJU8UXRSiFwwio5Ku6dxpmrepILEVigSK4pTRmV90Rz6TqyqJq2/pqLpkbb+KCaKLqjXafcg7IdtkFUQLcDB2X785hsAXKbdP/FYs0rxApmk1o6isMSkpgYFk7zpxJXlJWXXdMJHDhZUcZvU7ZUNeeWYMcm1ca9A0rL3J2qqrF/fKX2yD3qeQa2l7TkVTpVP/pu2tjOV9TLV8yJjc7u0COYnSS2KskJesPkeeSUJauIdJWNhOEM5c4AyI9wphE2poQyd7x3nUbTsyCoYHPXf9sMPzqt3MI8teAn7cAgRHMMZnMMlDIDBDfyAn/Ar+B7cBnfB37l0LVjM7MGDCP79BwcMtOc=</latexit><latexit sha1_base64="N3CijJIqFBowULdbrh+WNrhPgX8=">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</latexit><latexit sha1_base64="N3CijJIqFBowULdbrh+WNrhPgX8=">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</latexit>
xi = f�(ei)<latexit sha1_base64="270VJLD7gt0vH7AtNf4cVJpk6VA=">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</latexit><latexit sha1_base64="PuNsspGjpcc2SeVmkDWfr5ab2W0=">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</latexit><latexit sha1_base64="PuNsspGjpcc2SeVmkDWfr5ab2W0=">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</latexit>
xi ⇠ Point mass at f�(ei)<latexit sha1_base64="bmmUzqeCYB8Tl/BJoP0d5EVDo5U=">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</latexit><latexit sha1_base64="ZRpvVrSQojZgWGWUaf4g+64Qgc4=">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</latexit><latexit sha1_base64="ZRpvVrSQojZgWGWUaf4g+64Qgc4=">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</latexit>
z1<latexit sha1_base64="jbtr+W9vpRfg2saI7YImOblkNik=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcqMCOqu6MZlRccW2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7GN9m8/USVZjJ6gElMfYGHEQsZwWCo++e+26/WnIaTl70M3ALUUFGtfvW3N5AkETQCwrHWXdeJwU+xAkY4nVZ6iaYxJmM8pF0DIyyo9tPc6tQ+MczADqUyJwI7Z/9vpFhogWFklFnTc7OMASm5rhsVjETWsmfyu56IoB6IeiZSOtQLRiC89FMWxQnQiMx8hAm3QdpZLPaAKUqATwzARDHzFZuMsMIETHgVk5G7mMgy8M4aVw3n7rzWvC7CKqMjdIxOkYsuUBPdohbyEEFD9Ire0Lv1Yn1Yn9bXTFqyip1DNFfW9x/uTZqw</latexit><latexit sha1_base64="jbtr+W9vpRfg2saI7YImOblkNik=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcqMCOqu6MZlRccW2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7GN9m8/USVZjJ6gElMfYGHEQsZwWCo++e+26/WnIaTl70M3ALUUFGtfvW3N5AkETQCwrHWXdeJwU+xAkY4nVZ6iaYxJmM8pF0DIyyo9tPc6tQ+MczADqUyJwI7Z/9vpFhogWFklFnTc7OMASm5rhsVjETWsmfyu56IoB6IeiZSOtQLRiC89FMWxQnQiMx8hAm3QdpZLPaAKUqATwzARDHzFZuMsMIETHgVk5G7mMgy8M4aVw3n7rzWvC7CKqMjdIxOkYsuUBPdohbyEEFD9Ire0Lv1Yn1Yn9bXTFqyip1DNFfW9x/uTZqw</latexit><latexit sha1_base64="jbtr+W9vpRfg2saI7YImOblkNik=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcqMCOqu6MZlRccW2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7GN9m8/USVZjJ6gElMfYGHEQsZwWCo++e+26/WnIaTl70M3ALUUFGtfvW3N5AkETQCwrHWXdeJwU+xAkY4nVZ6iaYxJmM8pF0DIyyo9tPc6tQ+MczADqUyJwI7Z/9vpFhogWFklFnTc7OMASm5rhsVjETWsmfyu56IoB6IeiZSOtQLRiC89FMWxQnQiMx8hAm3QdpZLPaAKUqATwzARDHzFZuMsMIETHgVk5G7mMgy8M4aVw3n7rzWvC7CKqMjdIxOkYsuUBPdohbyEEFD9Ire0Lv1Yn1Yn9bXTFqyip1DNFfW9x/uTZqw</latexit>
z2<latexit sha1_base64="ZNKuRe23lzCJMNesh6cRgweKXJI=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3Lio4W2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DBy0TRahHJJeqE2BNOYuoBww47cSKYhFw+hiMr7P54xNVmsnoHiYx9QUeRixkBIOh7p77zX615jScvOxl4Baghopq96u/vYEkiaAREI617rpODH6KFTDC6bTSSzSNMRnjIe0aGGFBtZ/mVqf2iWEGdiiVORHYOft/I8VCCwwjo8yanptlDEjJdd2oYCSylj2T3/VEBPVA1DOR0qFeMALhhZ+yKE6ARmTmI0y4DdLOYrEHTFECfGIAJoqZr9hkhBUmYMKrmIzcxUSWgddsXDac27Na66oIq4yO0DE6RS46Ry10g9rIQwQN0St6Q+/Wi/VhfVpfM2nJKnYO0VxZ33/v75qx</latexit><latexit sha1_base64="ZNKuRe23lzCJMNesh6cRgweKXJI=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3Lio4W2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DBy0TRahHJJeqE2BNOYuoBww47cSKYhFw+hiMr7P54xNVmsnoHiYx9QUeRixkBIOh7p77zX615jScvOxl4Baghopq96u/vYEkiaAREI617rpODH6KFTDC6bTSSzSNMRnjIe0aGGFBtZ/mVqf2iWEGdiiVORHYOft/I8VCCwwjo8yanptlDEjJdd2oYCSylj2T3/VEBPVA1DOR0qFeMALhhZ+yKE6ARmTmI0y4DdLOYrEHTFECfGIAJoqZr9hkhBUmYMKrmIzcxUSWgddsXDac27Na66oIq4yO0DE6RS46Ry10g9rIQwQN0St6Q+/Wi/VhfVpfM2nJKnYO0VxZ33/v75qx</latexit><latexit sha1_base64="ZNKuRe23lzCJMNesh6cRgweKXJI=">AAACCHicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3Lio4W2qFk0kwbmkyG5I5Qh36BuNXvcCVu/Qs/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DBy0TRahHJJeqE2BNOYuoBww47cSKYhFw+hiMr7P54xNVmsnoHiYx9QUeRixkBIOh7p77zX615jScvOxl4Baghopq96u/vYEkiaAREI617rpODH6KFTDC6bTSSzSNMRnjIe0aGGFBtZ/mVqf2iWEGdiiVORHYOft/I8VCCwwjo8yanptlDEjJdd2oYCSylj2T3/VEBPVA1DOR0qFeMALhhZ+yKE6ARmTmI0y4DdLOYrEHTFECfGIAJoqZr9hkhBUmYMKrmIzcxUSWgddsXDac27Na66oIq4yO0DE6RS46Ry10g9rIQwQN0St6Q+/Wi/VhfVpfM2nJKnYO0VxZ33/v75qx</latexit>
z3<latexit sha1_base64="VwYPdoeVjXkUU912kwmjdymaS+E=">AAACCHicZVBLTsMwFHT4lvIrsGQTUSGxqKoEkIBdBRuWRRBaqY0qx3Vaq3Yc2S9IJeoJEFs4ByvElltwDG6Ak2ZB2ydZbzxvnjWeIOZMg+P8WEvLK6tr66WN8ubW9s5uZW//UctEEeoRyaVqB1hTziLqAQNO27GiWASctoLRTTZvPVGlmYweYBxTX+BBxEJGMBjq/rl31qtUnbqTl70I3AJUUVHNXuW325ckETQCwrHWHdeJwU+xAkY4nZS7iaYxJiM8oB0DIyyo9tPc6sQ+NkzfDqUyJwI7Z/9vpFhogWFolFnTM7OMASm5rhkVDEXWsmfyux6LoBaIWiZSOtRzRiC89FMWxQnQiEx9hAm3QdpZLHafKUqAjw3ARDHzFZsMscIETHhlk5E7n8gi8E7rV3Xn7rzauC7CKqFDdIROkIsuUAPdoibyEEED9Ire0Lv1Yn1Yn9bXVLpkFTsHaKas7z/xkZqy</latexit><latexit sha1_base64="VwYPdoeVjXkUU912kwmjdymaS+E=">AAACCHicZVBLTsMwFHT4lvIrsGQTUSGxqKoEkIBdBRuWRRBaqY0qx3Vaq3Yc2S9IJeoJEFs4ByvElltwDG6Ak2ZB2ydZbzxvnjWeIOZMg+P8WEvLK6tr66WN8ubW9s5uZW//UctEEeoRyaVqB1hTziLqAQNO27GiWASctoLRTTZvPVGlmYweYBxTX+BBxEJGMBjq/rl31qtUnbqTl70I3AJUUVHNXuW325ckETQCwrHWHdeJwU+xAkY4nZS7iaYxJiM8oB0DIyyo9tPc6sQ+NkzfDqUyJwI7Z/9vpFhogWFolFnTM7OMASm5rhkVDEXWsmfyux6LoBaIWiZSOtRzRiC89FMWxQnQiEx9hAm3QdpZLHafKUqAjw3ARDHzFZsMscIETHhlk5E7n8gi8E7rV3Xn7rzauC7CKqFDdIROkIsuUAPdoibyEEED9Ire0Lv1Yn1Yn9bXVLpkFTsHaKas7z/xkZqy</latexit><latexit sha1_base64="VwYPdoeVjXkUU912kwmjdymaS+E=">AAACCHicZVBLTsMwFHT4lvIrsGQTUSGxqKoEkIBdBRuWRRBaqY0qx3Vaq3Yc2S9IJeoJEFs4ByvElltwDG6Ak2ZB2ydZbzxvnjWeIOZMg+P8WEvLK6tr66WN8ubW9s5uZW//UctEEeoRyaVqB1hTziLqAQNO27GiWASctoLRTTZvPVGlmYweYBxTX+BBxEJGMBjq/rl31qtUnbqTl70I3AJUUVHNXuW325ckETQCwrHWHdeJwU+xAkY4nZS7iaYxJiM8oB0DIyyo9tPc6sQ+NkzfDqUyJwI7Z/9vpFhogWFolFnTM7OMASm5rhkVDEXWsmfyux6LoBaIWiZSOtRzRiC89FMWxQnQiEx9hAm3QdpZLHafKUqAjw3ARDHzFZsMscIETHhlk5E7n8gi8E7rV3Xn7rzauC7CKqFDdIROkIsuUAPdoibyEEED9Ire0Lv1Yn1Yn9bXVLpkFTsHaKas7z/xkZqy</latexit>
x3<latexit sha1_base64="r7RTsVKDbYXIRCyPj9szuInKMVA=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcqMCuqu6MZlBccW2qFk0kwbmkyG5I5Yhv6DuNXvcCVu/QY/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7Gt9m8/USVZjJ6gElMfYGHEQsZwWCoTi8Q6fO0f96v1pyGk5e9DNwC1FBRrX71tzeQJBE0AsKx1l3XicFPsQJGOJ1WeommMSZjPKRdAyMsqPbT3O/UPjHMwA6lMicCO2f/b6RYaIFhZJRZ03OzjAEpua4bFYxE1rJn8rueiKAeiHomUjrUC0YgvPJTFsUJ0IjMfIQJt0HaWTb2gClKgE8MwEQx8xWbjLDCBEyCFZORu5jIMvDOGtcN5/6i1rwpwiqjI3SMTpGLLlET3aEW8hBBHL2iN/RuvVgf1qf1NZOWrGLnEM2V9f0HLzudBQ==</latexit><latexit sha1_base64="r7RTsVKDbYXIRCyPj9szuInKMVA=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcqMCuqu6MZlBccW2qFk0kwbmkyG5I5Yhv6DuNXvcCVu/QY/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7Gt9m8/USVZjJ6gElMfYGHEQsZwWCoTi8Q6fO0f96v1pyGk5e9DNwC1FBRrX71tzeQJBE0AsKx1l3XicFPsQJGOJ1WeommMSZjPKRdAyMsqPbT3O/UPjHMwA6lMicCO2f/b6RYaIFhZJRZ03OzjAEpua4bFYxE1rJn8rueiKAeiHomUjrUC0YgvPJTFsUJ0IjMfIQJt0HaWTb2gClKgE8MwEQx8xWbjLDCBEyCFZORu5jIMvDOGtcN5/6i1rwpwiqjI3SMTpGLLlET3aEW8hBBHL2iN/RuvVgf1qf1NZOWrGLnEM2V9f0HLzudBQ==</latexit><latexit sha1_base64="r7RTsVKDbYXIRCyPj9szuInKMVA=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcqMCuqu6MZlBccW2qFk0kwbmkyG5I5Yhv6DuNXvcCVu/QY/wz8wM52FbS+Ee3LuueHkBDFnGhznxyqtrK6tb5Q3K1vbO7t71f2DRy0TRahHJJeqE2BNOYuoBww47cSKYhFw2g7Gt9m8/USVZjJ6gElMfYGHEQsZwWCoTi8Q6fO0f96v1pyGk5e9DNwC1FBRrX71tzeQJBE0AsKx1l3XicFPsQJGOJ1WeommMSZjPKRdAyMsqPbT3O/UPjHMwA6lMicCO2f/b6RYaIFhZJRZ03OzjAEpua4bFYxE1rJn8rueiKAeiHomUjrUC0YgvPJTFsUJ0IjMfIQJt0HaWTb2gClKgE8MwEQx8xWbjLDCBEyCFZORu5jIMvDOGtcN5/6i1rwpwiqjI3SMTpGLLlET3aEW8hBBHL2iN/RuvVgf1qf1NZOWrGLnEM2V9f0HLzudBQ==</latexit>
x2<latexit sha1_base64="OUe9NQ3ny9IilmHPU3ONQ4TtfRo=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3LCo4ttEPJpJk2NJkMyR2xDP0Hcavf4Urc+g1+hn9gZjoL214I9+Tcc8PJCWLONDjOj1VaW9/Y3CpvV3Z29/YPqodHj1omilCPSC5VN8CachZRDxhw2o0VxSLgtBNMbrN554kqzWT0ANOY+gKPIhYygsFQ3X4g0ufZoDmo1pyGk5e9CtwC1FBR7UH1tz+UJBE0AsKx1j3XicFPsQJGOJ1V+ommMSYTPKI9AyMsqPbT3O/MPjPM0A6lMicCO2f/b6RYaIFhbJRZ0wuzjAEpua4bFYxF1rJn8rueiqAeiHomUjrUS0YgvPJTFsUJ0IjMfYQJt0HaWTb2kClKgE8NwEQx8xWbjLHCBEyCFZORu5zIKvCajeuGc39Ra90UYZXRCTpF58hFl6iF7lAbeYggjl7RG3q3XqwP69P6mktLVrFzjBbK+v4DLZmdBA==</latexit><latexit sha1_base64="OUe9NQ3ny9IilmHPU3ONQ4TtfRo=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3LCo4ttEPJpJk2NJkMyR2xDP0Hcavf4Urc+g1+hn9gZjoL214I9+Tcc8PJCWLONDjOj1VaW9/Y3CpvV3Z29/YPqodHj1omilCPSC5VN8CachZRDxhw2o0VxSLgtBNMbrN554kqzWT0ANOY+gKPIhYygsFQ3X4g0ufZoDmo1pyGk5e9CtwC1FBR7UH1tz+UJBE0AsKx1j3XicFPsQJGOJ1V+ommMSYTPKI9AyMsqPbT3O/MPjPM0A6lMicCO2f/b6RYaIFhbJRZ0wuzjAEpua4bFYxF1rJn8rueiqAeiHomUjrUS0YgvPJTFsUJ0IjMfYQJt0HaWTb2kClKgE8NwEQx8xWbjLHCBEyCFZORu5zIKvCajeuGc39Ra90UYZXRCTpF58hFl6iF7lAbeYggjl7RG3q3XqwP69P6mktLVrFzjBbK+v4DLZmdBA==</latexit><latexit sha1_base64="OUe9NQ3ny9IilmHPU3ONQ4TtfRo=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcpMEdRd0Y3LCo4ttEPJpJk2NJkMyR2xDP0Hcavf4Urc+g1+hn9gZjoL214I9+Tcc8PJCWLONDjOj1VaW9/Y3CpvV3Z29/YPqodHj1omilCPSC5VN8CachZRDxhw2o0VxSLgtBNMbrN554kqzWT0ANOY+gKPIhYygsFQ3X4g0ufZoDmo1pyGk5e9CtwC1FBR7UH1tz+UJBE0AsKx1j3XicFPsQJGOJ1V+ommMSYTPKI9AyMsqPbT3O/MPjPM0A6lMicCO2f/b6RYaIFhbJRZ0wuzjAEpua4bFYxF1rJn8rueiqAeiHomUjrUS0YgvPJTFsUJ0IjMfYQJt0HaWTb2kClKgE8NwEQx8xWbjLHCBEyCFZORu5zIKvCajeuGc39Ra90UYZXRCTpF58hFl6iF7lAbeYggjl7RG3q3XqwP69P6mktLVrFzjBbK+v4DLZmdBA==</latexit>
x1<latexit sha1_base64="wPlw8y/0IqfHJrazcSQaHuwQ07k=">AAACDXicZVDLSgMxFM34rPVVdekmWAQXpcyIoO6KblxWcGyhLSWTZtrQZDIkd8Qy9B/ErX6HK3HrN/gZ/oGZ6SxseyHck3PPDScniAU34Lo/zsrq2vrGZmmrvL2zu7dfOTh8NCrRlPlUCaXbATFM8Ij5wEGwdqwZkYFgrWB8m81bT0wbrqIHmMSsJ8kw4iGnBCzV7gYyfZ72vX6l6tbdvPAy8ApQRUU1+5Xf7kDRRLIIqCDGdDw3hl5KNHAq2LTcTQyLCR2TIetYGBHJTC/N/U7xqWUGOFTanghwzv7fSIk0ksDIKrNm5mYZA0oJU7MqGMmsZc/kdzORQS2QtUykTWgWjEB41Ut5FCfAIjrzESYCg8JZNnjANaMgJhYQqrn9CqYjogkFm2DZZuQtJrIM/PP6dd29v6g2boqwSugYnaAz5KFL1EB3qIl8RJFAr+gNvTsvzofz6XzNpCtOsXOE5sr5/gMr950D</latexit><latexit sha1_base64="wPlw8y/0IqfHJrazcSQaHuwQ07k=">AAACDXicZVDLSgMxFM34rPVVdekmWAQXpcyIoO6KblxWcGyhLSWTZtrQZDIkd8Qy9B/ErX6HK3HrN/gZ/oGZ6SxseyHck3PPDScniAU34Lo/zsrq2vrGZmmrvL2zu7dfOTh8NCrRlPlUCaXbATFM8Ij5wEGwdqwZkYFgrWB8m81bT0wbrqIHmMSsJ8kw4iGnBCzV7gYyfZ72vX6l6tbdvPAy8ApQRUU1+5Xf7kDRRLIIqCDGdDw3hl5KNHAq2LTcTQyLCR2TIetYGBHJTC/N/U7xqWUGOFTanghwzv7fSIk0ksDIKrNm5mYZA0oJU7MqGMmsZc/kdzORQS2QtUykTWgWjEB41Ut5FCfAIjrzESYCg8JZNnjANaMgJhYQqrn9CqYjogkFm2DZZuQtJrIM/PP6dd29v6g2boqwSugYnaAz5KFL1EB3qIl8RJFAr+gNvTsvzofz6XzNpCtOsXOE5sr5/gMr950D</latexit><latexit sha1_base64="wPlw8y/0IqfHJrazcSQaHuwQ07k=">AAACDXicZVDLSgMxFM34rPVVdekmWAQXpcyIoO6KblxWcGyhLSWTZtrQZDIkd8Qy9B/ErX6HK3HrN/gZ/oGZ6SxseyHck3PPDScniAU34Lo/zsrq2vrGZmmrvL2zu7dfOTh8NCrRlPlUCaXbATFM8Ij5wEGwdqwZkYFgrWB8m81bT0wbrqIHmMSsJ8kw4iGnBCzV7gYyfZ72vX6l6tbdvPAy8ApQRUU1+5Xf7kDRRLIIqCDGdDw3hl5KNHAq2LTcTQyLCR2TIetYGBHJTC/N/U7xqWUGOFTanghwzv7fSIk0ksDIKrNm5mYZA0oJU7MqGMmsZc/kdzORQS2QtUykTWgWjEB41Ut5FCfAIjrzESYCg8JZNnjANaMgJhYQqrn9CqYjogkFm2DZZuQtJrIM/PP6dd29v6g2boqwSugYnaAz5KFL1EB3qIl8RJFAr+gNvTsvzofz6XzNpCtOsXOE5sr5/gMr950D</latexit>
{<latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit><latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit><latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit>
Symbo
lic
Mod
el
{<latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit><latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit><latexit sha1_base64="/RVgraER0C9MMfFvKD5SjE6r/Oo=">AAACB3icZVDLSgMxFM34rPVVdelmsAguSpmKoO6KblxWcWyhHUomzbSheQzJHaEM/QFxq9/hStz6GX6Gf2BmOgvbXgj35Nxzw8kJY84MeN6Ps7K6tr6xWdoqb+/s7u1XDg6fjEo0oT5RXOlOiA3lTFIfGHDaiTXFIuS0HY5vs3n7mWrDlHyESUwDgYeSRYxgsNRDL+1Xql7dy8tdBo0CVFFRrX7ltzdQJBFUAuHYmG7DiyFIsQZGOJ2We4mhMSZjPKRdCyUW1ARp7nTqnlpm4EZK2yPBzdn/GykWRmAYWWXWzNwsY0ApbmpWBSORteyZ/G4mIqyFopaJtInMghGIroKUyTgBKsnMR5RwF5SbpeIOmKYE+MQCTDSzX3HJCGtMwGZXthk1FhNZBv55/bru3V9UmzdFWCV0jE7QGWqgS9REd6iFfERQhF7RG3p3XpwP59P5mklXnGLnCM2V8/0HdQyacw==</latexit>
Neu
ral
Projec
tor
e1<latexit sha1_base64="WgitCLnES1r4TeJXeVDWd7cTvpM=">AAACDXicZVDLSgMxFM3UV62vqks3wSK4KGVGBHVXdOOygmMLbSmZNNOGJpMhuSOUof8gbvU7XIlbv8HP8A/MTGdh2wvhnpx7bjg5QSy4Adf9cUpr6xubW+Xtys7u3v5B9fDoyahEU+ZTJZTuBMQwwSPmAwfBOrFmRAaCtYPJXTZvPzNtuIoeYRqzviSjiIecErBUpxfIlM0G3qBacxtuXngVeAWooaJag+pvb6hoIlkEVBBjup4bQz8lGjgVbFbpJYbFhE7IiHUtjIhkpp/mfmf4zDJDHCptTwQ4Z/9vpEQaSWBslVkzC7OMAaWEqVsVjGXWsmfyu5nKoB7IeibSJjRLRiC87qc8ihNgEZ37CBOBQeEsGzzkmlEQUwsI1dx+BdMx0YSCTbBiM/KWE1kF/kXjpuE+XNaat0VYZXSCTtE58tAVaqJ71EI+okigV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAMuJzw</latexit><latexit sha1_base64="WgitCLnES1r4TeJXeVDWd7cTvpM=">AAACDXicZVDLSgMxFM3UV62vqks3wSK4KGVGBHVXdOOygmMLbSmZNNOGJpMhuSOUof8gbvU7XIlbv8HP8A/MTGdh2wvhnpx7bjg5QSy4Adf9cUpr6xubW+Xtys7u3v5B9fDoyahEU+ZTJZTuBMQwwSPmAwfBOrFmRAaCtYPJXTZvPzNtuIoeYRqzviSjiIecErBUpxfIlM0G3qBacxtuXngVeAWooaJag+pvb6hoIlkEVBBjup4bQz8lGjgVbFbpJYbFhE7IiHUtjIhkpp/mfmf4zDJDHCptTwQ4Z/9vpEQaSWBslVkzC7OMAaWEqVsVjGXWsmfyu5nKoB7IeibSJjRLRiC87qc8ihNgEZ37CBOBQeEsGzzkmlEQUwsI1dx+BdMx0YSCTbBiM/KWE1kF/kXjpuE+XNaat0VYZXSCTtE58tAVaqJ71EI+okigV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAMuJzw</latexit><latexit sha1_base64="WgitCLnES1r4TeJXeVDWd7cTvpM=">AAACDXicZVDLSgMxFM3UV62vqks3wSK4KGVGBHVXdOOygmMLbSmZNNOGJpMhuSOUof8gbvU7XIlbv8HP8A/MTGdh2wvhnpx7bjg5QSy4Adf9cUpr6xubW+Xtys7u3v5B9fDoyahEU+ZTJZTuBMQwwSPmAwfBOrFmRAaCtYPJXTZvPzNtuIoeYRqzviSjiIecErBUpxfIlM0G3qBacxtuXngVeAWooaJag+pvb6hoIlkEVBBjup4bQz8lGjgVbFbpJYbFhE7IiHUtjIhkpp/mfmf4zDJDHCptTwQ4Z/9vpEQaSWBslVkzC7OMAaWEqVsVjGXWsmfyu5nKoB7IeibSJjRLRiC87qc8ihNgEZ37CBOBQeEsGzzkmlEQUwsI1dx+BdMx0YSCTbBiM/KWE1kF/kXjpuE+XNaat0VYZXSCTtE58tAVaqJ71EI+okigV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAMuJzw</latexit>
e2<latexit sha1_base64="wBcNk6euWGMrg+Da+mnoLSs6/Fg=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcq0COqu6MZlBUcL7VAyaaYNzWNI7ghl6D+IW/0OV+LWb/Az/AMz01nY9kK4J+eeG05OGHNmwPN+nNLa+sbmVnm7srO7t39QPTx6NCrRhPpEcaW7ITaUM0l9YMBpN9YUi5DTp3Bym82fnqk2TMkHmMY0EHgkWcQIBkt1+6FI6WzQGlRrXsPLy10FzQLUUFGdQfW3P1QkEVQC4diYXtOLIUixBkY4nVX6iaExJhM8oj0LJRbUBGnud+aeWWboRkrbI8HN2f8bKRZGYBhbZdbMwixjQClu6lYFY5G17Jn8bqYirIeinom0icySEYiugpTJOAEqydxHlHAXlJtl4w6ZpgT41AJMNLNfcckYa0zAJlixGTWXE1kFfqtx3fDuL2rtmyKsMjpBp+gcNdElaqM71EE+IoijV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAOWpzx</latexit><latexit sha1_base64="wBcNk6euWGMrg+Da+mnoLSs6/Fg=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcq0COqu6MZlBUcL7VAyaaYNzWNI7ghl6D+IW/0OV+LWb/Az/AMz01nY9kK4J+eeG05OGHNmwPN+nNLa+sbmVnm7srO7t39QPTx6NCrRhPpEcaW7ITaUM0l9YMBpN9YUi5DTp3Bym82fnqk2TMkHmMY0EHgkWcQIBkt1+6FI6WzQGlRrXsPLy10FzQLUUFGdQfW3P1QkEVQC4diYXtOLIUixBkY4nVX6iaExJhM8oj0LJRbUBGnud+aeWWboRkrbI8HN2f8bKRZGYBhbZdbMwixjQClu6lYFY5G17Jn8bqYirIeinom0icySEYiugpTJOAEqydxHlHAXlJtl4w6ZpgT41AJMNLNfcckYa0zAJlixGTWXE1kFfqtx3fDuL2rtmyKsMjpBp+gcNdElaqM71EE+IoijV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAOWpzx</latexit><latexit sha1_base64="wBcNk6euWGMrg+Da+mnoLSs6/Fg=">AAACDXicZVDLSgMxFM3UV62vqks3g0VwUcq0COqu6MZlBUcL7VAyaaYNzWNI7ghl6D+IW/0OV+LWb/Az/AMz01nY9kK4J+eeG05OGHNmwPN+nNLa+sbmVnm7srO7t39QPTx6NCrRhPpEcaW7ITaUM0l9YMBpN9YUi5DTp3Bym82fnqk2TMkHmMY0EHgkWcQIBkt1+6FI6WzQGlRrXsPLy10FzQLUUFGdQfW3P1QkEVQC4diYXtOLIUixBkY4nVX6iaExJhM8oj0LJRbUBGnud+aeWWboRkrbI8HN2f8bKRZGYBhbZdbMwixjQClu6lYFY5G17Jn8bqYirIeinom0icySEYiugpTJOAEqydxHlHAXlJtl4w6ZpgT41AJMNLNfcckYa0zAJlixGTWXE1kFfqtx3fDuL2rtmyKsMjpBp+gcNdElaqM71EE+IoijV/SG3p0X58P5dL7m0pJT7ByjhXK+/wAOWpzx</latexit>
e3<latexit sha1_base64="LgiizHRmhHOKEEbwqVR4tWBS3lw=">AAACDXicZVDLTgIxFL2DL8QXauLGzURi4oKQQRfqjujGJSQiJECwUzrQ0E4n7R0TMuEfjFv9DlfGrd/gxn/wD+wAC4GbNPf03HOb0+NHghv0vG8ns7K6tr6R3cxtbe/s7uX3Dx6MijVldaqE0k2fGCZ4yOrIUbBmpBmRvmANf3ibzhtPTBuuwnscRawjST/kAacELdVs+zJh4+5FN1/wSt6k3GVQnoFC5aj28wgA1W7+t91TNJYsRCqIMa2yF2EnIRo5FWyca8eGRYQOSZ+1LAyJZKaTTPyO3VPL9NxAaXtCdCfs/42ESCMJDqwybWZuljKolDBFq8KBTFv6zORuRtIv+rKYirQJzIIRDK46CQ+jGFlIpz6CWLio3DQbt8c1oyhGFhCquf2KSwdEE4o2wZzNqLyYyDKon5euS17NZnUD08rCMZzAGZThEipwB1WoAwUBL/AKb86z8+58OJ9TacaZ7RzCXDlff9YqntM=</latexit><latexit sha1_base64="RMNkU799kfoHwXHS6vi9XiRQ2tw=">AAACDXicZVDLSgMxFM3UV62vquDGzWARXJQy1YW6K3XjsgXHFtqhZNJMG5pMhuSOUIb+g7jVX3DrSlwJfoMb/8E/MDPtwrYXwj0599xwcvyIMw2O823lVlbX1jfym4Wt7Z3dveL+wb2WsSLUJZJL1faxppyF1AUGnLYjRbHwOW35o5t03nqgSjMZ3sE4op7Ag5AFjGAwVLvri4ROehe9YsmpOFnZy6A6A6XaUfOHvdY/G73ib7cvSSxoCIRjrTtVJwIvwQoY4XRS6MaaRpiM8IB2DAyxoNpLMr8T+9QwfTuQypwQ7Iz9v5FgoQWGoVGmTc/NUgak5LpsVDAUaUufye56LPyyL8qpSOlALxiB4MpLWBjFQEMy9RHE3AZpp9nYfaYoAT42ABPFzFdsMsQKEzAJFkxG1cVEloF7XrmuOE2TVR1NK4+O0Qk6Q1V0iWroFjWQiwji6Ak9oxfr0Xqz3q2PqTRnzXYO0VxZX39bgqCP</latexit><latexit sha1_base64="RMNkU799kfoHwXHS6vi9XiRQ2tw=">AAACDXicZVDLSgMxFM3UV62vquDGzWARXJQy1YW6K3XjsgXHFtqhZNJMG5pMhuSOUIb+g7jVX3DrSlwJfoMb/8E/MDPtwrYXwj0599xwcvyIMw2O823lVlbX1jfym4Wt7Z3dveL+wb2WsSLUJZJL1faxppyF1AUGnLYjRbHwOW35o5t03nqgSjMZ3sE4op7Ag5AFjGAwVLvri4ROehe9YsmpOFnZy6A6A6XaUfOHvdY/G73ib7cvSSxoCIRjrTtVJwIvwQoY4XRS6MaaRpiM8IB2DAyxoNpLMr8T+9QwfTuQypwQ7Iz9v5FgoQWGoVGmTc/NUgak5LpsVDAUaUufye56LPyyL8qpSOlALxiB4MpLWBjFQEMy9RHE3AZpp9nYfaYoAT42ABPFzFdsMsQKEzAJFkxG1cVEloF7XrmuOE2TVR1NK4+O0Qk6Q1V0iWroFjWQiwji6Ak9oxfr0Xqz3q2PqTRnzXYO0VxZX39bgqCP</latexit>
f�(e)<latexit sha1_base64="8oxgnabmEamXCM6ELWdhWTrrt8c=">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</latexit><latexit sha1_base64="JbCPJ2fHAGZa8Ebw6DTj/vvFjWE=">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</latexit><latexit sha1_base64="JbCPJ2fHAGZa8Ebw6DTj/vvFjWE=">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</latexit>
zi ⇠ Symbolic Model<latexit sha1_base64="qRL/AxTFpgdIbBxTWc/weaeer2U=">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</latexit><latexit sha1_base64="Dnc6JOWSxrJAoiQOjwnmnyt3/Tw=">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</latexit><latexit sha1_base64="Dnc6JOWSxrJAoiQOjwnmnyt3/Tw=">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</latexit>
ei ⇠ N (µzi ,⌃zi)<latexit sha1_base64="8CLMr+o0dZvNiJKf88rFwAP//jQ=">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</latexit><latexit sha1_base64="N3CijJIqFBowULdbrh+WNrhPgX8=">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</latexit><latexit sha1_base64="N3CijJIqFBowULdbrh+WNrhPgX8=">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</latexit>
xi = f�(ei)<latexit sha1_base64="270VJLD7gt0vH7AtNf4cVJpk6VA=">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</latexit><latexit sha1_base64="PuNsspGjpcc2SeVmkDWfr5ab2W0=">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</latexit><latexit sha1_base64="PuNsspGjpcc2SeVmkDWfr5ab2W0=">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</latexit>
xi ⇠ Point mass at f�(ei)<latexit sha1_base64="bmmUzqeCYB8Tl/BJoP0d5EVDo5U=">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</latexit><latexit sha1_base64="ZRpvVrSQojZgWGWUaf4g+64Qgc4=">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</latexit><latexit sha1_base64="ZRpvVrSQojZgWGWUaf4g+64Qgc4=">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</latexit>
P(Z)
x = F(z)
P(X)
![Page 19: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/19.jpg)
In Pseudo-Code• xreal ~ Training data
• z ~ P(Z) → Normal(0, 1) or Uniform(-1, 1)
• xfake = G(z)
• yreal = D(xreal) → P(xreal is real)
• yfake = D(xfake) → P(xfake is real)
• Train D: minD - log yreal - log (1 - yfake)
• Train G: minG - log yfake → non-saturating loss
![Page 20: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/20.jpg)
Why are GANs good?
• Discriminator is a “learned metric” parameterized by powerful neural networks
• Can easily pick up any kind of discrepancy, e.g. blurriness, global inconsistency
• Generator has fine-grained (gradient) signals to inform it what and how to improve
![Page 21: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/21.jpg)
Problems in GAN Training• GANs are great, but training is notoriously difficult
• Known problems • Convergence & Stability:
• WGAN (Arjovsky et al., 2017) • Gradient-Based Regularization (Roth et al., 2017)
• Mode collapse/dropping: • Mini-batch Discrimination (Salimans et al. 2016) • Unrolled GAN (Metz et al. 2016)
• Overconfident discriminator: • One-side label smoothing (Salimans et al. 2016)
![Page 22: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/22.jpg)
Applying GANs to Text
![Page 23: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/23.jpg)
Applications of GAN Objectives to Language
![Page 24: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/24.jpg)
Applications of GAN Objectives to Language
• GANs for Language Generation (Yu et al. 2017)
![Page 25: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/25.jpg)
Applications of GAN Objectives to Language
• GANs for Language Generation (Yu et al. 2017)
• GANs for MT (Yang et al. 2017, Wu et al. 2017, Gu et al. 2017)
![Page 26: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/26.jpg)
Applications of GAN Objectives to Language
• GANs for Language Generation (Yu et al. 2017)
• GANs for MT (Yang et al. 2017, Wu et al. 2017, Gu et al. 2017)
• GANs for Dialogue Generation (Li et al. 2016)
![Page 27: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/27.jpg)
Problem! Can’t Backprop through Sampling
xreal
sample minibatch
sample latent vars.
z
xfake
convert w/ generator
y
predict w/ discriminator
![Page 28: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/28.jpg)
Problem! Can’t Backprop through Sampling
xreal
sample minibatch
sample latent vars.
z
xfake
convert w/ generator
y
predict w/ discriminator Discrete! Can’t backprop
![Page 29: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/29.jpg)
Solution: Use Learning Methods for Latent Variables
![Page 30: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/30.jpg)
Solution: Use Learning Methods for Latent Variables
• Policy gradient reinforcement learning methods (e.g. Yu et al. 2016)
![Page 31: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/31.jpg)
Solution: Use Learning Methods for Latent Variables
• Policy gradient reinforcement learning methods (e.g. Yu et al. 2016)
• Reparameterization trick for latent variables using Gumbel softmax (Gu et al. 2017)
![Page 32: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/32.jpg)
Discriminators for Sequences
![Page 33: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/33.jpg)
Discriminators for Sequences
• Decide whether a particular generated output is true or not
![Page 34: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/34.jpg)
Discriminators for Sequences
• Decide whether a particular generated output is true or not
• Commonly use CNNs as discriminators, either on sentences (e.g. Yu et al. 2017), or pairs of sentences (e.g. Wu et al. 2017)
![Page 35: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/35.jpg)
Discriminators for Sequences
• Decide whether a particular generated output is true or not
• Commonly use CNNs as discriminators, either on sentences (e.g. Yu et al. 2017), or pairs of sentences (e.g. Wu et al. 2017)
![Page 36: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/36.jpg)
GANs for Text are Hard! (Yang et al. 2017)
![Page 37: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/37.jpg)
GANs for Text are Hard! (Yang et al. 2017)
Type of Discriminator
![Page 38: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/38.jpg)
GANs for Text are Hard! (Yang et al. 2017)
Type of Discriminator
Strength of Discriminator
![Page 39: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/39.jpg)
GANs for Text are Hard! (Wu et al. 2017)
![Page 40: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/40.jpg)
GANs for Text are Hard! (Wu et al. 2017)
Learning Rate for Generator
![Page 41: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/41.jpg)
GANs for Text are Hard! (Wu et al. 2017)
Learning Rate for GeneratorLearning Rate for Discriminator
![Page 42: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/42.jpg)
Stabilization Trick: Assigning Reward to Specific Actions
![Page 43: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/43.jpg)
Stabilization Trick: Assigning Reward to Specific Actions• Getting a reward at the end of the sentence gives a
credit assignment problem
![Page 44: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/44.jpg)
Stabilization Trick: Assigning Reward to Specific Actions• Getting a reward at the end of the sentence gives a
credit assignment problem
• Solution: assign reward for partial sequences (Yu et al. 2016, Li et al. 2017)
![Page 45: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/45.jpg)
Stabilization Trick: Assigning Reward to Specific Actions• Getting a reward at the end of the sentence gives a
credit assignment problem
• Solution: assign reward for partial sequences (Yu et al. 2016, Li et al. 2017)
D(this)
![Page 46: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/46.jpg)
Stabilization Trick: Assigning Reward to Specific Actions• Getting a reward at the end of the sentence gives a
credit assignment problem
• Solution: assign reward for partial sequences (Yu et al. 2016, Li et al. 2017)
D(this)D(this is)
![Page 47: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/47.jpg)
Stabilization Trick: Assigning Reward to Specific Actions• Getting a reward at the end of the sentence gives a
credit assignment problem
• Solution: assign reward for partial sequences (Yu et al. 2016, Li et al. 2017)
D(this)D(this is)
D(this is a)
![Page 48: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/48.jpg)
Stabilization Trick: Assigning Reward to Specific Actions• Getting a reward at the end of the sentence gives a
credit assignment problem
• Solution: assign reward for partial sequences (Yu et al. 2016, Li et al. 2017)
D(this)D(this is)
D(this is a)D(this is a fake)
![Page 49: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/49.jpg)
Stabilization Trick: Assigning Reward to Specific Actions• Getting a reward at the end of the sentence gives a
credit assignment problem
• Solution: assign reward for partial sequences (Yu et al. 2016, Li et al. 2017)
D(this)D(this is)
D(this is a)D(this is a fake)
D(this is a fake sentence)
![Page 50: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/50.jpg)
Stabilization Tricks: Performing Multiple Rollouts
![Page 51: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/51.jpg)
Stabilization Tricks: Performing Multiple Rollouts
• Like other methods using discrete samples, instability is a problem
![Page 52: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/52.jpg)
Stabilization Tricks: Performing Multiple Rollouts
• Like other methods using discrete samples, instability is a problem
• This can be helped somewhat by doing multiple rollouts (Yu et al. 2016)
![Page 53: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/53.jpg)
Stabilization Tricks: Performing Multiple Rollouts
• Like other methods using discrete samples, instability is a problem
• This can be helped somewhat by doing multiple rollouts (Yu et al. 2016)
![Page 54: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/54.jpg)
Discrimination over Softmax Results (Hu et al. 2017)
• Attempt to generate outputs with a specific trait (e.g. tense, sentiment)
• Discriminator over the softmax results
x h yP(y)Adversary!
![Page 55: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/55.jpg)
Adversarial Feature Learning
![Page 56: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/56.jpg)
Adversaries over Features vs. Over Outputs
![Page 57: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/57.jpg)
Adversaries over Features vs. Over Outputs
• Generative adversarial networks
![Page 58: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/58.jpg)
Adversaries over Features vs. Over Outputs
• Generative adversarial networks
x h y
![Page 59: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/59.jpg)
Adversaries over Features vs. Over Outputs
• Generative adversarial networks
x h yAdversary!
![Page 60: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/60.jpg)
Adversaries over Features vs. Over Outputs
• Generative adversarial networks
• Adversarial feature learning
x h yAdversary!
![Page 61: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/61.jpg)
Adversaries over Features vs. Over Outputs
• Generative adversarial networks
• Adversarial feature learning
x h y
x h y
Adversary!
![Page 62: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/62.jpg)
Adversaries over Features vs. Over Outputs
• Generative adversarial networks
• Adversarial feature learning
x h y
x h y
Adversary!
Adversary!
![Page 63: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/63.jpg)
Adversaries over Features vs. Over Outputs
• Generative adversarial networks
• Adversarial feature learning
x h y
x h y
Adversary!
Adversary!
• Why adversaries over features?
![Page 64: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/64.jpg)
Adversaries over Features vs. Over Outputs
• Generative adversarial networks
• Adversarial feature learning
x h y
x h y
Adversary!
Adversary!
• Why adversaries over features?• Non-generative tasks
![Page 65: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/65.jpg)
Adversaries over Features vs. Over Outputs
• Generative adversarial networks
• Adversarial feature learning
x h y
x h y
Adversary!
Adversary!
• Why adversaries over features?• Non-generative tasks• Continuous features easier than discrete outputs
![Page 66: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/66.jpg)
Learning Domain-invariant Representations (Ganin et al. 2016)
![Page 67: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/67.jpg)
Learning Domain-invariant Representations (Ganin et al. 2016)
• Learn features that cannot be distinguished by domain
![Page 68: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/68.jpg)
Learning Domain-invariant Representations (Ganin et al. 2016)
• Learn features that cannot be distinguished by domain
![Page 69: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/69.jpg)
Learning Domain-invariant Representations (Ganin et al. 2016)
• Learn features that cannot be distinguished by domain
• Interesting application to synthetically generated or stale data (Kim et al. 2017)
![Page 70: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/70.jpg)
Learning Language-invariant Representations
![Page 71: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/71.jpg)
Learning Language-invariant Representations
• Chen et al. (2016) learn language-invariant representations for text classification
![Page 72: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/72.jpg)
Learning Language-invariant Representations
• Chen et al. (2016) learn language-invariant representations for text classification
![Page 73: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/73.jpg)
Learning Language-invariant Representations
• Chen et al. (2016) learn language-invariant representations for text classification
• Also on multi-lingual machine translation (Xie et al. 2017)
![Page 74: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/74.jpg)
Adversarial Multi-task Learning (Liu et al. 2017)
![Page 75: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/75.jpg)
Adversarial Multi-task Learning (Liu et al. 2017)
• Basic idea: want some features in a shared space across tasks, others separate
![Page 76: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/76.jpg)
Adversarial Multi-task Learning (Liu et al. 2017)
• Basic idea: want some features in a shared space across tasks, others separate
![Page 77: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/77.jpg)
Adversarial Multi-task Learning (Liu et al. 2017)
• Basic idea: want some features in a shared space across tasks, others separate
• Method: adversarial discriminator on shared features, orthogonality constraints on separate features
![Page 78: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/78.jpg)
Implicit Discourse Connection Classification w/ Adversarial Objective
(Qin et al. 2017)
![Page 79: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/79.jpg)
Implicit Discourse Connection Classification w/ Adversarial Objective
(Qin et al. 2017)• Idea: implicit discourse relations are not explicitly
marked, but would like to detect them if they are
![Page 80: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/80.jpg)
Implicit Discourse Connection Classification w/ Adversarial Objective
(Qin et al. 2017)• Idea: implicit discourse relations are not explicitly
marked, but would like to detect them if they are
• Text with explicit discourse connectives should be the same as text without!
![Page 81: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/81.jpg)
Implicit Discourse Connection Classification w/ Adversarial Objective
(Qin et al. 2017)• Idea: implicit discourse relations are not explicitly
marked, but would like to detect them if they are
• Text with explicit discourse connectives should be the same as text without!
![Page 82: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/82.jpg)
Professor Forcing (Lamb et al. 2016)
![Page 83: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/83.jpg)
Professor Forcing (Lamb et al. 2016)
• Halfway in between a discriminator on discrete outputs and feature learning
![Page 84: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/84.jpg)
Professor Forcing (Lamb et al. 2016)
• Halfway in between a discriminator on discrete outputs and feature learning
• Generate output sequence according to model
![Page 85: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/85.jpg)
Professor Forcing (Lamb et al. 2016)
• Halfway in between a discriminator on discrete outputs and feature learning
• Generate output sequence according to model
• But train discriminator on hidden states
![Page 86: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/86.jpg)
Professor Forcing (Lamb et al. 2016)
• Halfway in between a discriminator on discrete outputs and feature learning
• Generate output sequence according to model
• But train discriminator on hidden states
x h y
(sampled or true output sequence)
![Page 87: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/87.jpg)
Professor Forcing (Lamb et al. 2016)
• Halfway in between a discriminator on discrete outputs and feature learning
• Generate output sequence according to model
• But train discriminator on hidden states
Adversary!x h y
(sampled or true output sequence)
![Page 88: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/88.jpg)
Unsupervised Distribution Matching
![Page 89: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/89.jpg)
Unsupervised Style Transfer for Text (Shen et al. 2017)
• Task: transfer sentences with one style to another style • Decipherment: Translate ciphered sentences to natural sentences (A simpler case of
unsupervised MT) • Transfer sentences with positive sentiment to negative sentiment. • Word reordering
• Impressive performance on decipherment
![Page 90: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/90.jpg)
Unsupervised Alignment of Word Embeddings (Lample et al. 2018)• We have two word embedding spaces (A) in different languages
• Define a function (e.g. orthogonal transform) to map between the spaces
• Use adversarial loss to try to align (B), further find closest words (C), use supervised objective (D)
![Page 91: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/91.jpg)
Unsupervised Machine Translation (Lample et al. 2017, Artetxe et al. 2017)
• Methods:
• Cycle consistency (dual learning) (He et al. 2016, Zhu et al. 2017)
• Employing denoising auto-encoder to refine translated sentence
![Page 92: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/92.jpg)
Adversarial Robustness
![Page 93: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/93.jpg)
Problem!Networks Sensitive to Small Perturbations
(e.g. Belinkov et al. 2018)
![Page 94: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/94.jpg)
Adversarial Noise: Noise Specifically Designed to Break
Systems• Relatively simple to perform attacks on image
classification systems: calculate gradient to maximize loss
• More difficult for text because input is discrete, but still some success (e.g. Ebrahimi et al. 2018)
![Page 95: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/95.jpg)
What is an Adversarial Example? (Michel et al. 2019)
• It should be "meaning preserving" on the source side, and "meaning destroying" on the target side
• Meaning defined by semantic similarity (whatever that means)
![Page 96: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/96.jpg)
Adversarial Training
• We'd like to train our models to be robust to attacks!
• Simplest idea: sample adversarial examples at training time and make sure that they are also classified correctly
• Lots of theory, but little for NLP taskshttps://adversarial-ml-tutorial.org
![Page 97: CS11-747 Neural Networks for NLP Adversarial Methodsphontron.com/class/nn4nlp2019/assets/slides/nn4nlp-17-adversarial.pdf · • Adversarial feature learning: criticize the generated](https://reader033.fdocuments.net/reader033/viewer/2022060606/605c3f74c8bf111a21557001/html5/thumbnails/97.jpg)
Questions?