Filtering and feature extraction with neural networks

These notes are now complete for Winter, 1999.

Readings

Required reading: Chapter 4.

ways of representing information in state vectors (from Chapter 3)
- local codes ("grandmother cell codes")
- distributed representations
- coarse codes: semi-local, partially overlapping codes
Transforming physical stimuli to mechanical/electrical signals in sensors, e.g. tactile stimuli
the "power law" in sensory transduction: our percept is proportional to the log of stimulus intensity - and so is neural firing frequency e.g. in limulus eye

Lateral inhibition: A mechanism for signal enhancement and data compression.
An example in nature: Lateral inhibition in the limulus compound eye
Examples in signal processing: filtering image and speech data.
A unit with feedforward excitation and lateral inhibition approximates a second derivative filter. The signal enhancement and data compression go hand in hand: the unit filters out regions of constant and linearly varying signal intensities, while enhancing abrupt changes in the signal e.g. edges in images, or onset/offset of auditory or tactile stimuli.
Program 1: lateral inhibition in a feedforward network (Matlab program)
Program 2: lateral inhibition with feedback links (Matlab main program and activation function )
Computer demos in class of lateral inhibition (Mon. Jan. 18)
Lateral inhibition and filtering: the relationship between a centre-surround pattern of excitation/inhibition and a second derivative operator (using finite differencing)