Associative memory: Lessons from engineering and neuroscience
Speaker(s): Vassilis Cutsuridis, Kings College LondonAssociative memory is one of the oldest artificial neural network (ANN) paradigms. The concept of the associative memory was first introduced by the formalism of a correlation matrix. In the correlation matrix, memory patterns are encoded as the activity patterns across a network of computing units. Patterns are stored in memory by
Hebbian modification of the connections between the computing units. A memory is recalled when an activity pattern that is a partial or noisy version of a stored pattern is instantiated in the network. Network activity then evolves to the complete stored pattern as appropriate units are recruited to the activity pattern, and noisy units are removed, by threshold-setting of unit activity. Memory capacity for accurate recall is strongly dependent on the form of patterns to be stored and the learning rule employed.
The hippocampus, one of the most studied brain regions, has been implicated in the storage and recall of memory patterns (e.g. declarative and spatial memories). Recent hippocampus research has yielded a wealth of data on network architecture, cell types, the anatomy and membrane properties of pyramidal cells and interneurons, and
synaptic plasticity. Understanding the functional roles of the different families of hippocampal neurons in encoding and retrieval of memory patterns, synaptic plasticity and network oscillations poses a great challenge, but also promises deep insights into how the brain stores and recalls memories. Computational models play an instrumental role in providing clues on how these processes may take place.
In my talk, I will present two computational models of hippocampal dynamics addressing the issues of memory capacity, recall performance and theta phase precession in the hippocampus.
Return to Seminars for 2011-2012