Memory Networks and STDP
Princeton University, Princeton NJ
During the spring semester I worked one-on-one with Princeton’s incomparable Professor John Hopfield to refine a computational model of associative memory. Using mathematics and a truckload of MATLAB, I explored the effects of this new addition to the model, from a single “integrate-and-fire” neuron to multiple, dynamic synapses.
The Memory Network Model Explained:
In Professor Hopfield’s simplified model of associative memory, one can think of each entity in a person’s memory (i.e. a name or a hair color) as a single neuron. When you are thinking about that entity, that neuron fires. A memory (i.e. a person you have met), then, is simply a set of learned connections between the neurons that describe something. This connection, or “synapse,” means that remembering one thing about a person (i.e. their name) will help you remember everything else about them (i.e. their hair color, gender, and height), because exciting one neuron also excites the neurons connected to it.
Refining The Model:
This simple model is incredibly accurate in simulating the activity of real memory. But the brain is not so simple, and computational neuroscience is all about using experimental data to refine a model. Professor Hopfield noted that in the presence of “synaptic drift,” or changes in synaptic strength caused by random cellular processes, the model does not retain its memories like a real brain does. The new model uses something called Spike Timing-Dependent Plasticity, or STDP, to stabilize these drifting connections. The most common form of STDP ensures that if one neuron fires just before a neuron it is connected to, it implies a causal relationship, and that connection will be strengthened. If the order is reversed, the connection is weakened.
The results of this study were not entirely conclusive, but at the end of my project it looked as if STDP could indeed account for the reliable maintenance of memories in the face of synaptic drift.
For More Information: