Supervised Learning and Self-Organisation

Stream 2a: Supervised Learning and Self-Organisation

Consistent with the main aim of developing a generic cognitive architecture that supports a wide range of tasks, recent work in our group has mainly focussed on modelling the human neocortex – a structure in the brain that plays a crucial role in memory, attention, analogy-making, reasoning, language and problem-solving. One of the interesting properties of the neocortex is that, although different regions process different kinds of information (such as sensory information or controlling voluntary muscle movements), the neocortex seems uniform in structure, suggesting to many researchers that there exists a general algorithm in the brain that supports intelligent behaviour irrespective of sensory modality or task. Our current research involves implementing a simplified neocortex that can be trained through a mixture of supervised and unsupervised learning, that can be used for recognition (equating high-level abstract representations with previous experience), as well as temporal prediction. Such temporal prediction is a central ability required for hypothetical thinking (i.e. problem solving, analogy-making and reasoning).

The adjacent video demonstrates the ability of our cortical model trained on visual input. The various windows show the hierarchical nature of processing, extracting more and more abstract features in the higher layers, leading to a high degree of invariant recognition.

The top left window shows the image seen by the camera, which is then processed with a DOG filter (top right image). The bottom left image shows edges of a specific orientation being detected (there are 8 such filters), whereas the last window shows the result of a higher-level feature detector (of which there are 50) detecting complex patterns .

The algorithm that we use is also generative, being capable of turning high-level abstract representations into specific predictions. For example, we have trained the system to extract the statistical regularities in words (by "reading" Alice in Wonderland), and using this information to solve anagrams. That is, given experience with words, the system can generate word-like candidates, using the same mechanisms that we believe are central to many instances of creativity (such as in the generation of music).

Home - Stream 1 - Stream 2a - Stream 2b - Stream 2c