Report

Dendritic Computation Group Project Review 19 July 2013 Projects • Modelling dragonfly attention switching • Dendritic auditory processing – Mesgarani and Chang, in silicio – The auditory pathway • Processing images with spikes • Dendritic computation with memristors • Computation in RATSLAM – Image processing – SKIM on Spinnaker • Dendritic computation on Nengo • SKIM model on FPAA • Spike based cross-correlation Auditory Pathway Audio Signal to Spikes Poisson Spike Trains for Hair Cell Stimulated at 200Hz 0.5 0 -0.5 Neuron firing rate limited by spike delay Rectified by the volley principle and phase-locking Poisson spike train generated for each fiber for hair cell Promotes parallelism and simplicity in processing through stochastic computation 1 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 0.5 0 1 0.5 0 1 0.5 0 1 0.5 0 1 0.5 0 Poisson Spike Trains for Hair Cell Stimulated at 2000Hz 0.5 0 -0.5 1 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 0.5 0 1 0.5 0 1 0.5 0 1 0.5 0 1 0.5 0 Dendritic computation with memristors Jens Burger, Greg Cohen Memristors for Alpha Functions ● ● Use tunable resistance of memristor to control time constants for charging and discharging of capacitor Use memristor under 2 conditions – – With fixed resistances With changing resistances caused by exceeding threshold Implementation ● Matlab code rewritten in C++ and interfaced to Ngspice – – Compute each synaptic function in Ngspice and return data to C++ code Use multi-threading to compute synaptic kernels in parallel Results ● Can reproduce results by using RC circuits as alpha functions – Worked with identical RC circtuits (resistive) and different RC circuits (memristive) •Comments ● A lot of the computational power lies within the mapping between inputs and synaptic kernels – ● Requirements of synaptic kernels was rather low and impact of different setups on overall performance is hard to evaluate Proof-of-Concept successful – For parameter and setup exploration we need more computational resources Dendritic computation with Nengo Daniel Rasmussen FPAA Implementation for the SKIM model Suma George, Georgia Institute of Technology Atlanta Replacing SKIM hidden layer neurons with a dendrite Spiking patterns for different Input delays Spiking pattern for different patterns: Dendrite with varying diameter Generating random weights SKIM model hidden layer with a single ncompartment dendrite Spiking pattern for random input weights Stochastic Electronics: cross-correlation with neurons Tara Julia Hamilton, Jonathan Tapson, and others Calibration with square wave inputs gives phase delay in histogram i.e. it works! Autocorrelation with a single neuron Block diagram of chip Crossorrelation with two neurons