A Simple Feed Forward System
- To make a categoriser with biologically plausible components, all
you need is a way to encode the input, a way to encode the categories,
and STDP synapses to connect them.
- We encode the input features as a set of neurons that fire once
when the feature is present.
- We encode the output category as one (or more) neuron.
- We well connect the system in two layers of neurons, inputs
to outputs. These connections are STDP synapses.
- During training, we turn on the input neurons, then turn
on the output neurons.
- That's the key bit inputs just before outputs.
- There are a couple of versions of STDP. If you use multiplicative
weights, the weights will reflect the likelihood that the category
occurs when the feature occurs.
- The synapses are roughly co-occurence detectors.
- This is largely what we did last year.
- It's pretty easy to categorise using a simple two layer
topology.
- It's also reasonably effective, but generally not as good as
gradient descent.
