Nonnegative Tensor Factorization for EEG Pattern ... · PDF fileNonnegative Tensor...

76
Nonnegative Tensor Factorization for EEG Pattern Classification Seungjin Choi Department of Computer Science POSTECH, Korea [email protected] Abstract Learning fruitful representation from data is one of fundamental problems in machine learning and pattern recognition. Various methods have been developed, including factor analysis, principal component analysis (PCA), independent component analysis (ICA), manifold learning, and so on. Among those, nonnegative matrix factorization (NMF) has recently drawn extensive attention, since promising results were reported in handling nonnegative data such as document, image data, spectrograms of audio. NMF seeks a decomposition of a nonnegative data matrix into a product of two factor matrices (basis matrix and encoding matrix) such that all factor matrices are forced to be nonnegative. In this talk, I will begin with the useful behavior of NMF in EEG pattern classification, which plays a critical role in noninvasive brain computer interface (BCI). Next, I will introduce multiway extension of NMF, what is called, “nonnegative tensor factorization” and will emphasize its useful behavior in EEG pattern classification. A tensor is nothing but ‘multiway array’, generalizing vector and matrix in order to accommodate higher-order representations. For instance, a vector is a 1-way tensor, a matrix is a 2-way tensor, and a cube is a 3-way tensor, etc. Through this talk, I will stress why tensor is useful in learning fruitful representation, compared to existing matrix-based methods. Proc. of the 8th POSTECH-KYUTECH Joint Workshop on Neuroinformatics, Kitakyushu, Japan, 2008 - 1 -

Transcript of Nonnegative Tensor Factorization for EEG Pattern ... · PDF fileNonnegative Tensor...

  • Nonnegative Tensor Factorization for EEG Pattern Classification

    Seungjin Choi

    Department of Computer Science

    POSTECH, Korea

    [email protected]

    Abstract

    Learning fruitful representation from data is one of fundamental problems in machine learning and

    pattern recognition. Various methods have been developed, including factor analysis, principal component

    analysis (PCA), independent component analysis (ICA), manifold learning, and so on. Among those,

    nonnegative matrix factorization (NMF) has recently drawn extensive attention, since promising results

    were reported in handling nonnegative data such as document, image data, spectrograms of audio. NMF

    seeks a decomposition of a nonnegative data matrix into a product of two factor matrices (basis matrix

    and encoding matrix) such that all factor matrices are forced to be nonnegative. In this talk, I will begin

    with the useful behavior of NMF in EEG pattern classification, which plays a critical role in noninvasive

    brain computer interface (BCI).

    Next, I will introduce multiway extension of NMF, what is called, nonnegative tensor factorization and

    will emphasize its useful behavior in EEG pattern classification. A tensor is nothing but multiway array,

    generalizing vector and matrix in order to accommodate higher-order representations. For instance, a

    vector is a 1-way tensor, a matrix is a 2-way tensor, and a cube is a 3-way tensor, etc. Through this talk, I

    will stress why tensor is useful in learning fruitful representation, compared to existing matrix-based

    methods.

    Proc. of the 8th POSTECH-KYUTECH Joint Workshop on Neuroinformatics, Kitakyushu, Japan, 2008

    - 1 -

  • Sequence learning: a network model of the entorhinal cortex layer II with hippocampal-entorhinal loop circuit

    Hatsuo Hayashi

    Department of Brain Science and Engineering, Graduate School of Life Science and Systems Engineering,

    Kyushu Institute of Technology Email: [email protected]

    Finding of place cells encoding place information in phase of the theta rhythm has revealed that the hippocampus

    has ability to represent spaces. The firing phase with respect to the theta rhythm advances within a theta cycle when the animal traverses a place field [1, 2]. Although the phase precession was found in the hippocampal CA3 and CA1, it has also been found in the dentate gyrus [2], and then in grid cells in the medial entorhinal cortex [3]. So, a model of the entorhinal-hippocampal system accounting for the phase precession has been proposed [4]. This model is rather abstract, and the several layers of the hippocampus and the entorhinal cortex were modeled by simple oscillator units to describe local field theta rhythms. Another model of the entorhinal cortex layer II network that has the entorhinal-hippocampal loop circuit has also been proposed [5]. This model consisted of conductance-based multi-compartmental models of the stellate cell and the inhibitory interneuron, and the entorhinal-hippocampal loop circuit was modeled by delay lines whose signal transmission delay was 20 ms, based on physiological data. In this lecture, we will review this model having mechanisms of theta rhythm generation and theta phase coding, and show how successfully sequence learning is done.

    We also tested how the shape of STDP rule affected the performance of the present model amid background noise [6]. STDP rules observed in the hippocampus and the entorhinal cortex are classified into two types: temporally symmetric and asymmetric STDP rules. The symmetric STDP rule that has LTD windows on both sides of the central LTP window prevented irrelevant enhancement of loop connections caused by noise in cooperation with loop connections having a larger signal transmission delay and the theta rhythm pacing the activity of the stellate cells. However, the asymmetric STDP rule that has no LTD window in the range of positive-spike timing did not prevent such irrelevant enhancement, and the memory pattern was blurred by noise.

    Important elements of the present model for sequence learning are (1) symmetric STDP rule, (2) loop connections having a large signal transmission delay, and (3) theta rhythm pacing the activity of stellate cells. These elements also contribute to robust sequence learning amid background noise. Above all, the LTD window in the range of positive spike-timing is important to prevent influences of noise with the progress of sequence learning.

    REFERENCES [1] OKeefe J and Recce ML, Phase relationship between hippocampal place units and the EEG theta rhythm,

    Hippocampus, vol. 3, pp. 317-330, 1993 [2] Skaggs WE, McNaughton BL, Wilson MA, and Barnes CA, Theta phase precession in hippocampal neuronal

    populations and the compression of temporal sequence, Hippocampus, vol. 6, pp. 149-172, 1996 [3] Hafting T, Fyhn M, Bonnevie T, Moser M-B, and Moser E, Hippocampus-independent phase precession in

    entorhinal grid cells, Nature, vol. 453, pp. 1248-1252, 2008 [4] Yamaguchi Y, A theory of hippocampal memory based on theta phase precession, Biol. Cybern., vol. 89, pp.

    1-9, 2003 [5] Igarashi J, Hayashi H, and Tateno K, Theta phase coding in a network model of the entorhinal cortex layer II

    with entorhinal-hippocampal loop connections, Cogn. Neurodyn., vol. 1, pp. 169-184, 2007 [6] Igarashi J and Hayashi H, Theta phase coding and suppression of irrelevant plastic change through STDP in the

    entorhino-hippocampal system amid background noise, in Advances in Cognitive Neurodynamics, (Eds) Wang R, Gu F, and Shen E, Springer, in press, 2008

    Proc. of the 8th POSTECH-KYUTECH Joint Workshop on Neuroinformatics, Kitakyushu, Japan, 2008

    - 2 -

  • TRAINING AND FARTIGUE OF POLYMER ARTIFICIAL MUSCLES

    Keiichi Kaneto, Hirotaka Suematsu, and Kentaro Yamato Department of Biological Function and Engineering, Kyushu Institute of Technology,

    Email: [email protected]

    1. INTRODUCTION

    Present robots are heavy and noisy, because they are driven by motors made with copper and iron. In order to build light weight and human friendly robots, artificial muscles or soft actuators are demanded to replace motors. Artificial muscles are required not only to drive robots but also to mimic biological motions like a snake, fishes and insects for medical equipments and variety of application. The research on soft actuators [1-7] began 20 years ago based on ionic polymers (membrane), polymer gels, dielectric elastic polymers and conducting polymers. Among them conducting polymers are the most prospective material from the view point of low operating voltage, large strain and stress[8]. The conducting polymers are actuated or deformed by electrochemical oxidation and reduction. The deformation results from the insertion of bulky ions as well as the change of polymer conformation due to the delocalization of -electron upon oxidation. The deformation is named as the electrochemomechanical strain (ECMS).

    In various conducting polymers, polypyrrole (PPy) based ECMS is currently studied, since PPy shows excellent performance in strain, stress and cycle life compared with those of polyaniline [9] and polythiophene [10]. Efforts have been mostly paid to improve the strain, stress and response time [3,4], however, characteristics of ECMS under tensile loads have not been investigated from the view points of training, fatigue and aging.

    Muscles can be strengthen by training through heavy tasks, though it feels fatigue. One may consider training of artificial muscle is impossible, however, we have found that a training of an artificial muscle fabricated with conducting polymers [11] is feasible. The training effect was observed as the increased strain after an experience of heavier tensile stress. This is resulted from a relaxation of anisotropic deformation, which was induced along the tensile direction as realignment of polymeric molecules during a creeping under higher stresses.

    In this paper, ECMS responses of cation driven PPy actuators [11-14] operated under various tensile stresses are mentioned. After the application of large tensile stresses, the recovery of strain due to creeping was studied to demonstrate the effects of training and fatigue of artificial muscles. The results are discussed taking a model of relaxation of anisotropic deformation induced by high tensile stress into consideration. The energy conversion efficiency from electrical input energy to mechanical output work as the function of tensile stress is also reported.

    2. EXPERIMENTAL

    Polypyrrole films were electrodeposited on a Ti plate in an aqueous electrolyte solution of 0.15 M pyrrole and 0.25 M of DBS acid at a constant current of 1 mA/cm2 for 1500s vs. a Ni counter electrode as described in previous paper [11,14]. The thickness and conductivity of typical film (named as PPy/DBS) was approximately 18 m and 30-50 S/cm, respectively. Young modulus of film was 140 and 70 MPa at oxidized and reduced states, respectively. 3. RESULTS AND DISCUSSION

    Cyclic voltammograms (CV) of PPy/DBS film in 1 M NaCl aqueous solution are shown in Fig.1 for various tensile stresses. The film was cycled for 3 times at given tensile stress from 0 to 5 MPa by a step of 1MPa sequentially, and the CVs obtained at 3rd cycle are shown in Fig.1. The film was expanded by the negative voltage application (reduction), and contracted by the positive voltage application (oxidation) as observed in cation driven films [11,14]. It was found that with increasing the tensile stress a shoulder at around 0V in the oxidation cycle increased and shifted to higher potential. Similarly, the peak at around -400 mV in the reduction cycle also shifted to highe