Bayesian networks are a computational model that is very efficient for computing in the presence of uncertainty. It excels in such tasks as predicting stock market behavior, disease progression, etc. It is ideal for taking an event that occurred and predicting the likelihood of different known causes to have been the contributing factor. For example, given the symptoms of a patient, it can compute the probabilities of various diseases that could be causing the symptoms. Unfortunately, implementing Bayesian networks usually requires complex hardware that is expensive, prone to failure, dissipates too much energy and consumes too much area on a computer chip. The goal of this research is to overcome these disadvantages by replacing traditional electronic hardware with magnetic devices that interact with each other in a special way to elicit Bayesian inference. This can reduce the hardware complexity and all associated costs dramatically, making Bayesian networks compact and efficient. This research will establish the viability of this approach through extensive simulations. Graduate students will be trained in this field to produce a pool of skilled scientists and engineers with cutting-edge knowledge.
Bayesian networks for computing in the presence of uncertainty leverage Bayesian inference engines implemented with complex hardware that often involves microcontrollers, shift registers, analog-to-digital converters, logic gates, etc. that dissipate exorbitant amounts of energy and have enormous footprints on a chip. It ha recently been shown by the project team that magnetic tunnel junctions (MTJs) that interact with each other by means of dipole coupling can implement Bayesian networks with vastly reduced energy cost and much smaller footprints. Two dipole coupled MTJs A and B can realize an extremely efficient 2-node Bayesian network, where the probabilities of high and low resistance states of MTJ A are set by current or voltage, while the (random) resistance state of MTJ B is determined by varying degrees of dipole coupling between the two MTJs. The degree of dipole coupling is tuned with local strain applied to the soft layer of MTJ B using electrical excitation. This allows one to generate any desired anti-correlation or correlation between the resistance states of the two MTJs that can be varied between 0% and 100% using electrical excitation. In turn, this allows the generation of programmable conditional probabilities that can be exploited for Bayesian networks. This research will build a simulation base for this approach, test the viability of MTJ-based inference engines under different scenarios and design optimal sub-systems.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
