Biologically realistic mechanisms for supervised learning

When we learn to do something new, we are often provided with an explicit example of how we should perform. For example, when we learn to play an instrument, we may have a teacher who corrects our mistakes by showing us the proper way to play. This form of learning is referred to as supervised learning, and although it is probably important for many of the things we learn in life, scientists know almost nothing about how the brain actually does it.

Interestingly, computer scientists have been able to develop algorithms for supervised learning that have been incredibly successful. These algorithms lie underneath many current artificial intelligence applications, such as computer vision. These algorithms also give neuroscientists some clues about how the brain might do supervised learning. However, the most successful algorithms for supervised learning, such as the back-propagation of error algorithm, are biologically unrealistic. We are exploring other, more biologically realistic mechanisms for supervised learning that theoretical neuroscientists have proposed. Using artificial neural network models and 2-photon imaging paired with targeted optical stimulation of individual cells, we are testing these proposed mechanisms to determine whether they are operating in the neocortex.