Presentation on theme: "Spike timing-dependent plasticity: Rules and use of synaptic adaptation Rudy Guyonneau Rufin van Rullen and Simon J. Thorpe Rétroaction lors de l‘ Intégration."— Presentation transcript:
Spike timing-dependent plasticity: Rules and use of synaptic adaptation Rudy Guyonneau Rufin van Rullen and Simon J. Thorpe Rétroaction lors de l‘ Intégration Visuelle: vers une Architecture GEnérique
What is STDP anyway? - Biological evidence - Theoretical studies - Synthesis Learning with STDP and asynchronously spiking neurons - The « controlled » learning paradigm (theory and biologically plausible implications) - The « autonomous » learning paradigm (growing filters and connectivity design) Overview
I. Biological evidence of synaptic adaptation « When an axon of cell A is near enough to excite a cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells such that A’s efficiency as one of the cell firing B, is increased » [Hebb, 1949] Which fires together wires together (taken from Markram 1997) (taken from Bi & Poo 1998)
I. Experimental to functional (taken from Abbott & Nelson, NatureN, 2000) STDP… … regulatory mechanism for rate and variability of post-synaptic firing … coincidence detector … introduces competition between synapses to control the firing rate (correlation factor) … suppresses strong recurrent excitatory loops « In general, STDP greatly expands the capability of Hebbian learning to address temporally sensitive computational tasks. »
I.Modelling stuff STDP can act as a learning mechanism for generating neuronal responses selective to input timing, order and sequence. sequence learningand prediction [Abbott & Blum, 96] spatial path learning in navigation [Blum & Abbott, 96; Mehta, Neuron, 00] direction selectivity in visual responses [Mehta, Neuron, 00] STDP also as « temporal difference learning » [Rao & Sejnowski, NeuralComp, 01] (Taken from Song et al., NatureN, 2000) (Taken from Gerstner & Kistler, BioCyb, 2002)
I. Spike timing-dependent plasticity So what have we got? The neuron’s activity, its spikes history, gives rise to modifications of the efficacy of its excitatory synapses that in turn affects its own spiking behaviour. That means we have synaptic adaptation in the autopoeitic sense.
The retina provides evidence for highly reproducible firing events... [Berry et al., 97]... And individual spikes with high temporal precision have been reported throughout the ventral stream. [from the retina (Sestokas, 91) to IT (Nakamura, 98); MT (Bair and Koch, 96)] Importantly, the latency of the first spike in the spike train is the most reliable. [Mainen & Sejnowski, 95] stimulus-locked spike timings are temporally (quite) precise! Spikes reproducibility What does that mean when we consider STDP in the context of reproducible spike trains?
II. What happens when a neuron repeatedly receives the same stimulus? 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Spike time OUTPUT 1/3 + 1/3 + 1/3 = 1 t3 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Step 1 Spike time STDP Synapses get modified at t3 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Step 1 Spike time OUTPUT 1/3 + 2/3 = 1 t2 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Spike time Step 2 STDP Synapses get modified at t2 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Step 2 Spike time OUTPUT 2/3 + 3/3 = 5/3 t2 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Spike time Step 3 STDP Synapses get modified at t2 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Step 3 Spike time OUTPUT 3/3 = 1 t1 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Spike time Step 4 STDP Synapses get modified at t1 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Step 4 Spike time OUTPUT 3/3 = 1 t1 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Spike time Step 5 STDP Synapses get modified at t1 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Step 5 Spike time OUTPUT 3/3 = 1 t1 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Spike time Step 6 STDP Synapses get modified at t1 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Step 6 Spike time OUTPUT 3/3 = 1 t1 1/3 3/3 2/3 0/3 t0t1t2t3 Weight INPUT Spike time Step 7 Step n a wave of spikes elicits a post-synaptic response and triggers STDP-like learning rule synapses carrying spikes just preceding the post-synaptic one are potentiated. Step n+1 Re-propagating the input spike-wave, the latency of the post-synaptic spike is slightly decreased. synapses carrying even earlier spikes are potentiated and later ones depressed. Reiterate By iterating the processus on and on from an(y) initial state allowing a post-synaptic response, it follows that the earliest synapses get maximally potentiated and later ones are depressed. Main assumption Reproducibility of the spike wave (saccades, oscillations).
II. Results Stimulus is composed of … 20Hz reproducible spike trains (Poisson process) by 1000 input neurons … to which a 5ms jitter is added 5Hz spontaneous activity Typical input spike trainsDynamics of the simulation
II. Latency versus… … firing rate (Gerstner, 97) no spontaneous activity – 5ms jitter … synchronicity (Abeles) 5 Hz spontaneous activity - no jitter
II. Neuron population case Retina to V1 (preliminary results) V1 to V2? (preliminary results) « Autonomous learning » Given the reliability and form of retinal input, what happens when a reductive, simili, but still biologically plausible, visual system « experiments » a simili natural world? Extension It may be possible that through STDP, areas build their selectivities from correlated afferents.
The end. General conclusions STDP is accountable for adaptation at the cellular level. Definition of a learning paradigm for asynchronously spiking neuron networks. Connectivity… (taken from Felleman & vanEssen, 91) (taken from vanRullen & Thorpe, 02)