OCNC---2004 Statistical Approach to Neural Learning and Population Coding ---- Introduction to Mathematical Neuroscience Shun-ichi Amari Laboratory for Mathematical Neuroscience RIKEN Brain Science Institute
BRAIN biological science information science Computational neuroscience Neurocomputing Mathematical Neuroscience
II. Population Coding I. Mathematical Neuroscience ---- modern topics ----classical theories II. Population Coding ---- modern topics III. Bayesian Inference ---- its merits and critique
Dynamics of Neuro-Ensembles Dynamics of Neuro-Fields 1. Mathematical Neurons Dynamics of Neuro-Ensembles Dynamics of Neuro-Fields Learning and Self-Organization 5. Self-Organization of Neuro-Fields
I Mathematical Neurons Simple model
output function u
spiking neuron integration-and-fire neuron rate coding
synchrony : spatial correlations firing probability
rate coding ensemble coding
1-layer network
Ensemble of networks macroscopic state macroscopic law
stability = =
Associative memory m pairs
Randomly generated Random matrix
II Dynamics of Neuro-Ensembles spiking neurons : stochastic point process synchronization Ensemble coding : macrodynamics
Simple examples Bistable S S Multi-stable
oscillation Amari (1971); Wilson-Cowan (1972)
competitive model (winner-take-all) ・・・ (winner-share-some)
multistable associative memory decision process (Anderson, Amari, Nakano, Kohonen Hopfield) decision process (Hopfield) travelling salesman problem
General Theory Transient Attractors stable state limit cycle chaos (strange attractors)
Chaotic behavior random stable states chaos Chaotic memory search
Associative memory (content-addressable memory) dynamics random attractor
Theory 1 =
=
Theory 2 …..
Macroscopic state Amari & Maginu, 1998
Dynamics of recalling processes Direction cosine Correct pattern 1 time simulations
Direction cosine 1 theory time
simulation Threshold of recalling Spurious memory
Dynamics of temporal sequence (Amari, 1972) non-monotonic output function Morita model
Nonmonotonic model non-monotonic
memory capacity : sparse exact : no spurious memories chaotic oscillation inhibitory connection
Biology hippocumpus, Rolls et al Chaotic associative memory Tonegawa et al CA3 Chaotic associative memory Aihara et al Chaotic search
Associative Memory Dynamics of a Chaotic Neural Networks Each neuron model shows chaotic dynamics Synaptic weights are determined by an auto-correlation matrix of the stored patterns Stored Patterns t=0 t=1 t=2 t=3 t=4
t=5 t=6 t=7 t=8 t=9 t=10 t=11 t=12 t=13 t=14 t=15 t=16 t=17 t=18 t=19
t=20 t=21 t=22 t=23 t=24 t=25 t=26 t=27 t=28 t=29 t
III Field Dynamics of Neural Excitation timing local excitations: travelling wave: oscillatory: memory decision Amari, Biol. Cybern,1978
Dynamics of Neural Fields
unstable stable
excitatory and inhibitory fields traveling wave oscillation
Neural Learning (Hebbian) classic theory ……… Information source I
Amari, Biol,Cybern,1978 Hebbian correlation generalized inverse principal component analyzer Perceptron ….
Neural learning (STDP) Spike-time dependent plasticity …. ……. emergence of synchrony LTP LTD
Learning Potential ………
1. Hebbian … 2. correlation associative memory …
3. generalized inverse least square
4. principal component analyzer Amari (1978), Oja (1980) 5. perceptron
Theory of Learning Networks Amari, IEEE Trans.C,1967 PDP: backprop; natural gradient
Learning algorithm
outer world
Self-organization …. …….
Proof RF of a neuron :
Special case Theorem: each : receptive field size of receptive field
1.resolution 2.topological property Self-Organizing Nerve Field signal space neural field 1.resolution 2.topological property
higher-dimensional 2- dimensional Topology Signal space Neural field higher-dimensional 2- dimensional position×orientation
orientation Signal space Neural field position
Self-organizing nerve field
dynamical stability patch structure variational equation stability : Takeuchi&Amari, Biol.Cybern, 35, 63-72, 1979
Topological properties emergence of block structure
Bayesian vs Fisherian. Any confrontation. --- histrical New framework Bayesian vs Fisherian? Any confrontation? --- histrical New framework? ---Amari, 1967 New in neuroscience?
Bayesian framework mle vs map information and decision asymptotically equivalent regularization theory predictive distribution
Priors: uniform, Jeffreys, smooth Hierarchical (empirical) Bayes decision of prior
Singular statistical model Singular model and prior