Adapted for adaptability-evidence suggests primate brains use reservoir computing

Photo: Christopher Walsh, Harvard Medical School

Primate brains are incredibly adaptive, responding to new situations and stimuli as needed. Think of all of the various skills we learn from computer programming to making pottery to neuroscience. Our brains learn new content and do so relatively efficiently, but how? How are humans and other primates so good at adapting to new information and scenarios?

Enel and colleagues (2016) wanted to better understand how our primate brains adapt to situations and stimuli that evolution could not have directly anticipated. They took advantage of developments from Rigotti and colleagues (2010) in reservoir computing, a branch of recurrent neural networks in which randomly connected neurons form a network of recurrent loops. Recurrent connections are fixed but connections from to output neurons can change. Rigotti and colleagues (2010) propose that recurrent neural networks with multiple, different neuronal responses are significant in the ability to complete complex cognitive tasks.

Enel and colleagues (2016) developed a recurrent neural network model that would perform a specific cognitive task and then compared predictions from the model to data from rhesus monkeys. Of four potential targets on a touch screen, only one would reward the monkeys with fruit juice. The monkeys needed to discover the fruit juice target by trial and error. The target corresponding to fruit juice changed after every trial.

Model Architecture, Figure from Enel and colleagues 2016 paper, Reservoir Computing Properties of Neural Dynamics in Prefrontal Cortex

Results show that the model was able to shift when it did not receive a reward and repeat when it did, a new extension of the functions of reservoir computing. The model was able to perform the task almost perfectly using a circular search and an ordered search and when trained on a schedule that was derived from the performance of one of the monkeys. Enel and colleagues (2016) compared the neural coding behavior of the model with the primate cortex. They found that variance in the presence of reservoir neurons in 1) target choice and 2) phase of the problem (either continue searching to find the reward or repeat because the reward has been discovered) show significant effects for both choice and phase. Choice and phase could not be directly derived from current inputs, but instead needed the history of previous inputs and their responses. This has also been observed for the primate brain (Barone and Joseph, 1989). See the published paper for the full results and more details about their findings.

This new research helps to explain how primates face and then handle a diverse and endless range of situations. It will be interesting to see if the same mechanism for behavioral adaptability is found in non-primate species or if it is unique to primates.

Works cited:

Barone, P., & Joseph, J. P. (1989). Prefrontal cortex and spatial sequencing in macaque monkey. Experimental brain research, 78(3), 447-464.

Enel, P., Procyk, E., Quilodran, R., & Domineer, P., F. (2016) Reservoir Computing Properties of Neural Dynamics in Prefrontal Cortex. PLoS Comput Biol, 12(6): e1004967. DOI: 10.1371/journal.pcbi.1004967 

Rigotti, M., Ben Dayan Rubin, D. D., Wang, X. J., & Fusi, S. (2010). Internal representation of task rules by recurrent dynamics: the importance of the diversity of neural responses. Frontiers in Computational Neuroscience, 4, 24.