Neural Network Model for Memory

We will now review the neural network model for memory and study how the knowledge of orthogonalization methods can help us in understanding certain cognitive functions. 

All types of perceptions - due to sight, touch, hearing etc. - lead to neural activities and are translated into patterns of firing and quiescent neurons. In a system of N neurons such a pattern corresponding to (say) an object can be 

The model represents learning in terms of the well known Hebb s hypothesis [10]. According to it learning manifests itself in the modification of synaptic connections between all i s and j s as a new object (a pattern of +1 and —1) comes to be recorded. The effectiveness or the strength of a synapse - whether it is excitatory (i.e. Jy 0) or inhibitory (i.e. Jy 0) - is believed to be altered by the activities of the neurons it connects. This flexibility of the synapses is modelled in the following choice of J, 

When one of the p memories of learnt patterns, say p = i/, is presented to the brain for association, it generates local fields hf on all the neurons according to 

The above condition of associative memory fails for p/N 0.14, i.e. the model gives a memory capacity of 0.14.1V. If the number of patterns that come to be recorded exceeds 0.14./V the retrieval process breaks down and a memory black-out results. The cause for this catastrophe is the correlation among the memories. When the vectors and have some overlap, they have a non-vanishing scalar product, i.e. 0 and they represent 

---