Transcript Investigating shape representation in human visual cortex

What can spiking neurons compute?
Kendrick Kay
(with Greg Corrado)
October 5, 2009
The big picture
• Given neuromorphic spiking hardware, what useful
things can we compute?
• Think about computation in terms of input/output
• Literature on spiking neurons not large; hard to subject
spiking neurons to theoretical analysis
• Approach: Simulate neuromorphic spiking hardware
(in MATLAB) and see what happens
Example of a result
• Oja’s rule applied to a perceptron leads to
computation of the first principal component of the
data (Oja 1982)
Let’s get started: FNS
• FNS: “Framework for Neural Simulation”
– a modular code architecture (in MATLAB) that allows
different flavors of synapses, membranes, and spikers to be
interchanged and reused
– includes a variety of neuronal models as special cases
What kind of neural network to build?
• Can we design a general-purpose neural network that
can learn appropriate behaviors when embedded in a
closed-loop system (input: sensory, output: action)?
• Example:
– learn to play Pong based on raw video input
– or maybe more vision-type stuff
Some initial results
• Perceptron
– Hmm... notice that input, output, network can each be binary
or continuous (8 possible combinations)
Open issues, things to do
Things to tackle:
• The world is analog, but spiking neurons propagate digital
signals.
• The world evolves over time, but artificial neural networks
tend to deal with static data points.
Is this going to work at all?
• Multi-layer perceptrons already can compute any function.
• What does spiking computation really offer? Surely not
computational capacity, but efficiency?