EAGER: Neural Networks that Temporally Change (NOTCH)

University Of Massachusetts Amherst, Amherst MA

Investigators

Abstract

A common feature of many real applications of national significance and of great importance to people is the need to identify and characterize very early on, and with limited data, changes that occur, requiring a rapid response. This is the case with critically ill patients: Identifying early signs of sepsis, of cardiac or respiratory impairment, of stroke – and beginning rapid treatment can save lives; characterizing a natural disaster in advance, getting drones to take into account variances or finding them if failed, and identifying the earliest signs and type of a cyber-attack can prevent life-threatening or economically devastating events and in some scenarios provides a shield of privacy protecting national, business and individual interests. Like our brains, state-of-the-art AI utilizes pattern recognition for identification: Our brains utilize temporal pattern recognition as an integral part of modeling our environment and predicting events, which in turn influences and decides our behavior. Unlike our brains, AI lacks temporal awareness; this severely limits AI’s understanding of and ability to model its environment and ability to predict outcomes, which in turn limits its capabilities in real, out-of-the-lab scenarios. Our new technology gives AI systems temporal awareness and vastly improves overall pattern recognition and ability to operate safely, effectively and efficiently - with major implications for emerging autonomous and intelligent systems. The technology is based on a new type of neural networks, where much like the brain, the connections between neurons are no longer scalar numbers, but rather temporal functions. Additionally, learning the temporal connectivity among neurons does not require full reliance on training data, instead the network continues to adapt to changing conditions and improve its temporal understanding. The resulting network shows unparalleled capacity in temporal assisted prediction, surprisingly strong accuracy, and ability to remain effective even when most environmental measurements are minimal or are lost. Interestingly, our temporally changing network while more capable, is smaller in size and consumes significantly less power. Our project will develop the technology and demonstrate its capabilities in a spectrum of uses including autonomous and health applications, complex system prediction (e.g., weather and natural disaster prevention), in analysis of human behavior and intentions, and more. 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.