Memory in biology

Unfortunately I could not pay attention this morning fully.

Jean Pascal Pfister ("Bayesian inference is important for biology") was welcomed.

Emre Neftci started off with an introduction to the morning. Then Jean-Pascal talked about model types ranging from conceptual, biophysical, normative (i.e. putting a metric norm or rule, e.g. backprop) and "dream" models.

Next Bruno Averbeck took over with his usual excellent tutorial style. He started by sumarizing Q learning. He then reviewed how cortex, striatum and substantia niagra (SN) interact as the accepted model or the brain's reward system. Cortex tells animal what is available, striatum values it, and the reward prediction error RPE is expressed as activation of dopaminergic neurons in SN 

The basic equation from 1972 is shown in this sketch. It relates value, prediction of reward, and error in this prediction on their effect on the updated value function.



Barry Richmond emphasized that Bruno's recent work has shown very clearly the shortcomings of this model, e.g. that cortex is plastic, unlike in the model.

Yulia followed by showing models that express somehow these ideas.

  1. She started with a space code, where the location of the activations in an ordered array encodes a value and its distribution. 
  2. She then described how a reward could be presented at the end of an action to result in such a space code developing, 
  3. but how the simplest rules like Hebb or STDP don't work without anatomical constraints since there is too much correlation.
I had to leave after this to take care of hotel business.

Altogether, I had an interesting time trying to summarize some of the discussions. 

There was substantial pushback from those of us now doing synchronous digital logic on the non-utility of silicon SNNs and mixed signal approaches.  But there was also new discussion this year about interesting new observations in biology and from alternative models of computation, in particular, interacting relational maps.


Comments