Aside from a few forays into anomaly detection, the temporal data we have is underutilized. If we going to throw deep models at the problem can we make them care more about the context of the lengthy video than the distribution of the small sample of cells?
1) Anomaly detection: given a video, can we detect when significant changes occur in the cell? can we characterize or classify the nature of this change?
a) How do we design a model that is sensitive where we want it? i.e. the threshold problem outlined in Marcus et al.
2) Useful sequential representation: Can we summarize a video into a few points in time that tell a story or can be used to classify behavior? Can we translate this into something user friendly for our imaginary doctor friends?
Stub to start noting some ideas about thesis.
Aside from a few forays into anomaly detection, the temporal data we have is underutilized. If we going to throw deep models at the problem can we make them care more about the context of the lengthy video than the distribution of the small sample of cells?
1) Anomaly detection: given a video, can we detect when significant changes occur in the cell? can we characterize or classify the nature of this change? a) How do we design a model that is sensitive where we want it? i.e. the threshold problem outlined in Marcus et al.
2) Useful sequential representation: Can we summarize a video into a few points in time that tell a story or can be used to classify behavior? Can we translate this into something user friendly for our imaginary doctor friends?