Modification of Internal Representations as a Mechanism for Learning in Neural Systems

Description of Thesis Title: Modification of Internal Representations as a Mechanism for Learning in Neural Systems. 1. Incoming sensory signals are processed by hierarchically organised modules in the brain. In certain contexts, this may be modelled by a feedforward layered network of interconnected binary units. The activity patterns in the intermediate layers are internal representations. 2. A new learning algorithm uses projections from the desired output to modify internal representations. Biologically realistic 2-layer synaptic rules can then be applied to cause the associated input to evoke the modified representation(s) that are more readily trained to produce the target output. 3. Simulation is carried out on benchmark tasks for 3-layer feedforward networks. Comparisons with other popular algorithms are made. The results suggest that the new algorithm has better generalisation performance with faster or equal learning speed on the tasks simulated. 4. The learning algorithm is generalised to a multi-layer network setting, in which internal representations are dynamically constructed. 5. The above will be put into the context of efficient sensory coding that is based on Barlow’s ‘redundancy reduction’ proposal