Surveillance systems [HC11] are equipped with a suite of sensors that can be switched and steered to focus attention on any target or location over a physical landscape (see Figure 1) to detect abrupt changes at any location. On the other hand, sensor suites are resource limited, and only a limited subset, among all the locations, can be probed at any time.

The recently proposed stochastic Polyak stepsize (SPS) and stochastic line-search (SLS) for SGD have shown remarkable effectiveness when training over-parameterized models. However, two issues remain unsolved in this line of work.

Numerous images of atarget human subject or ascene are required for training. Second, a learned model has limited generalization capability.

Furthermore,analysis of model interpretation reveals the capability of MAtt in capturing informative EEGfeatures andhandling thenon-stationarity ofbraindynamics.