Thursday Colloquium

Neural networks in the brain and artificial neural networks (ANNs) in silico are famously able to learn complex computational functionalities, but they do so in very different ways.  Brains learn by Hebbian "fire together / wire together" rules, where neuron-neuron connection strengths change using only local information.  By contrast, ANNs are trained for machine learning tasks with algorithms like Backprogagation, where vast external CPU and memory use global information about network topology and connection weights to update each weight.  Here, I will discuss growing efforts to create physical networks in the lab that perform machine learning-like tasks simply by equilibrating when input data are supplied as boundary conditions. The emphasis will be on training methods that use local information, like in the brain, and experimental realizations in both electrical and mechanical networks.  We call these "Contrastive Local Learning Networks (CLLNs)".  Our CLLNs consist of networks of identically-constructed edges with tunable strengths, and can be considered as both metamaterials and novel forms of soft matter.  They have many brain-like advantages over ANNs, such as speed, energy efficiency, robustness to imperfections, and potential for scalability.  They also enable the study of learning as a bottom-up process wherein global functionality emerges from complex dynamics originating in simple local rules.