Spring 2018

Brain and Computation Seminar

Feb 28, 2018 10:30 am – 11:30 am 

Add to Calendar

Parent Program: 

Calvin Lab Rm 116

Efficient Balanced Networks

The brain computes with spikes, and spikes are very costly to metabolism. Yet, the spiking responses of most cortical cells appear extremely noisy, to the extent that the only feature that repeats from trial to trial is the mean firing rate. If that was the case, the most sophisticated of central nervous system, the mammal brain, would be one of the less efficient and reliable in the animal kingdom.  Another prominent feature of cortical networks is that they maintain a tight balance between excitation and inhibition. This balance can result in chaotic dynamics, providing a mechanistic account for such variability. However this does not solve the conundrum of "why" circuits would be organized this way.

Here, we will strongly challenge this view and show that networks that learn to maintain a tight E/I balance converge in fact to state were they represent their stimuli most efficiently, i.e.  they achieve the highest possible accuracy and robustness given the number of spikes they spend.  Thought local, biologically plausible plasticity rules, biophysical entities such as membrane potentials and spikes acquire global functional meaning, constantly monitoring the network performance and correcting its errors.  Neural responses become highly variable, not because they are noisy, but because the population code is highly degenerate: Many different spiking patterns are equivalent in terms of coding. The mammal brain must combine high capacity, high robustness and low metabolic cost. According to our framework,  it results in unpredictable activity at the level of single cells, but extremely reliable population codes.

Finally, borrowing concepts from control theory, we will show that these networks can learn any dynamical systems. For example, they can learn to integrate sensory input, memorize, make decisions, control motor effectors, all based on a relatively small number of example trajectories. The infered learning rules are local and biophysically plausible, and guarrantied to converge to a highly efficient implementation of those tasks. Thus, the number of spikes and population size required are order of magnitude lower (and thus, the capacity order of magnitudes higher) than current rate-based models.