Omri Barak - Theoretical Neuroscience
The command “Take a left turn at the next junction”, while seemingly easy to interpret, entails integration of information across at least two different timescales. On the one hand the sensory information of the word “Take” has to be retained for about a second until the sentence is over. On the other, the meaning of the sentence has to be retained for about a minute until the junction is actually reached. This example presents two timescales out of a large range relevant for daily functioning.
When considering possible solutions utilized by the brain for these tasks, it is worthwhile to note that the bio-chemical processes in the brain have inherent time scales spanning many orders of magnitude such as milliseconds (action potentials), seconds (Calcium dynamics) and hours (protein synthesis).
On the functional level, we know that the nervous system can adapt to the statistics of the environment over a wide range of timescales, but the mechanisms for doing so are yet unknown.
As an example of two different ways to solve a task, consider the problem of holding the identity of a stimulus in memory for a delay of a couple of seconds. The prevailing dogma is that the brain does it by keeping the neurons which are related to the stimulus firing action potentials throughout the delay. Although every action potential is a process of a few milliseconds, if the neurons are recurrently connected to each other then positive feedback can maintain this activity over the two seconds. This is reminiscent of the way RAM works, or a car engine produces electricity.
Another option, which I explored during my PhD [Barak et al. 2007, Mongillo et al. 2008], is that the brain utilizes the slower dynamics of Calcium in the neurons in order to store the information there. According to this view the relevant neurons only fire transiently during the stimulus, but their associated synaptic Calcium levels remain elevated throughout the delay.
I am now exploring a more general formulation of this specific problem: given several mechanisms with different timescales, that could all solve the same task, can the brain effectively choose a mechanism such that its timescale matches that of the task. Potential advantages of such a solution are metabolic – slower processes are usually less energetically demanding than maintaining fast processes across a long time, and robustness – a slow process is less likely to be disrupted by fast noise. On the other hand, an obvious disadvantage is the loss of flexibility – a system tuned to a specific timescale (or a specific range of timescales) will perform poorly if forced to operate in a different regime.
What is the nature of neural representations? How does the activity of neurons in the brain reflect the external stimuli and internal states of the animal? Many studies addressing this question searched for single neurons with easily interpretable activity profiles – for example, neurons whose firing patterns tightly correspond to a particular stimulus or internal state. Striking examples of such neurons were found in several systems: orientation tuning in the early visual system, grid and place cells in the hippocampal and parahippocampal regions and even medial temporal lobe neurons responding to the face of specific actresses. Yet, it is likely that all behaviors are the result of a concerted action of large populations of neurons, and thus it would seem unlikely that such orderly single neurons can account for the entirety of network behavior. Indeed, in higher areas such as the prefrontal cortex, and even in more peripheral areas such as V1 there exist many neurons whose activity profiles defy simple explanations. The response of such neurons is typically explained by assuming that they do not directly participate in the task at hand.
I have recently begun testing an alternative hypothesis that explains both the orderly and the complex response profiles observed in experiments. Specifically, relying on results from the field of reservoir computing, I found that unstructured network models can be trained to perform tasks that were previously solved by carefully designed networks containing well-tuned single neurons.
I am now investigating the advantages and limitations of this approach, using measures such as performance, flexibility and robustness as indicators. Between the two extremes of engineered and unstructured networks lies a continuum of possible network models. I am exploring this continuum under the hypothesis that training a network for a specific task can gradually shift it from unstructured to structured. I will use experimental data from my collaborators to evaluate where along this continuum specific real networks reside. To this end, I am developing different analysis methods, evaluating their sensitivity to detect underlying network mechanisms, and later applying these methods to recorded data.