Séminaire Simon Thorpe
CERCO Toulouse
Finding repeating patterns : the secret of biological intelligence?
The latest generation of deep learning trained AI systems can rival humans on many challenging tasks, including image categorisation, speech processing and text prediction. But it can be argued that such systems are not truly “intelligent” because of the way they learn, requiring millions of cycles of brute force learning with labelled data. I will argue that such systems lack one feature of biological learning that may be critical for true intelligence. Humans are incredibly good at spotting repeating patterns. For example, when presented with a stream of novel images at rates of up to 120 frames a second, an image repeated between 2 and 5 times can subsequently be picked out of a set of four images - where the other 3 have all been seen just once . Similarly impressive levels of performance have been seen with auditory snippets presented at rates of 20 per second - again, a few repeats are enough to make the sound easy to pick out. This suggests that this ability to detect patterns that repeat could be a general feature of the brain. Parallel modelling work has shown that a modified Spike-Time Dependent Plasticity Rule using binary synapses can allow neurons to learn to detect repeating patterns, again with only a few repeats. We would like to suggest that this sort of unsupervised learning could be a key to biological intelligence, and could provide a radical alternative to the supervised deep learning techniques that currently dominate the field.
titre et abtract à venir
