2013/SkimFPAA

Examining how to implement the SKIM model on FPAA hardware

Project members: Suma George ,Tara Julia Hamilton,Greg Cohen,

Background

Field Programmable Analog Arrays (FPAAs) is a mixed-signal CMOS chip which allows analog components to be connected together in an arbitrary fashion. Re-configurable Analog Signal Processor(RASP) was one of the first large scale FPAAs. It allowed us to build multiple complex circuits. The specific chip used from the family of RASP chips for this project is RASP 2.9ab. It is a powerful and re-configurable analog computing platform that can be used to build neuromorphic models. The dendrite model to be used looks as follows:

Bio-physically modeled CMOS dendrites

Project Details

SKIM Framework

The project goal was to implement the SKIM framework using bio-physically modeled dendrites to substitute the alpha function. In the SKIM model, we replaced the hidden layer with a dendrites on the hardware and experimented with two architectures for the same. One replacing each hidden layer neuron with a n-compartment dendrite and the second being a n-compartment dendrite to represent the hidden layer.

Architecture 1

SKIM dendrite architecture1

The objective of this experiment was to test if we got different spike responses for different patterns which would then help in classification. Instead of setting the random weights for the inputs I varied the diameter of the dendrite by varying the axial conductance. This would inherently model different weights.

Results

Studied the outputs from the two different architectures to see which would be a better implementation.

arch1 result2

arch1 result3

Also, tested for different random weights as well to compare while still having a varying diameter dendrite.

arch1 varying weights arch1 varying weights 2

Architecture 2

In this architecture instead of n dendrites for the hidden layer, we tried using a single n-compartment dendrite with all the inputs coming into every stage of the dendrite. We haven't tested this architecture for its feasibility but it seemed like an interesting alternative and would be a much smaller design.

SKIM dendrite architecture 2

Results

arch1 result1

Now varying the input synaptic weights:

arch 2 varying weights

Discussion

This project demonstrates how SKIM can be implemented using the FPAA. So far we haven't incorporated learning of weights which can be implemented using STDP synapses modeled on the FPAA. We have dedicated neuron chips that can be used to implement SKIM efficiently. This project served as a good proof of concept if you will.

Attachments