Real-Time Machine Learning (RTML) | NSF - National Science Foundation (nsf19566) - 0 views
-
MiamiOH OARS on 12 Mar 19The need to process large data sets arising in many practical problems require real-time learning from data streams makes high-performance hardware necessary, and yet the very nature of these problems, along with currently known algorithms for addressing them, imposes significant hardware challenges. Current versions of deep-learning algorithms operate by using millions of parameters whose optimal values need to be determined for good performance in real time on high-performance hardware. Conversely, the availability of fast hardware implementations can enable fuller use of Bayesian techniques, attractive for their ability to quantify prediction uncertainty and thus give estimates of reliability and prediction breakdown. The abilities of ML systems to self-assess for reliability and predict their own breakdowns (and also recover without significant ill effects) constitute critical areas for algorithm development as autonomous systems become widely deployed in both decision support and embodied AI agents. Only with attention to these challenges can we construct systems that are robust when they encounter novel situations or degradation and failure of sensors.