Hexagonal Receptive Fields (RFs) for Macs at Multiple Levels during same Snippet
A 20-frame KTH snippet (jogger) being played while viewing RFs of macs at levels, L1 to L5. See related page for further information. One of our ultimate goals is to be able to show how the (spatiotemporal) features activating at a given level are linked (bound), and thus compose, higher-level features activating in the superjacent level. Ultimately, we'd like to automate generation of hierarchical networks (tree/DAG-structures) linking all features composing a given concept (class) or instance, across many levels. In conjunction with supervised training in which labels could be associated with concepts (features) at multiple levels, this would yield discrete, symbolic representations of organically discovered, natural, and intrinsically more continuously variable concepts, such as "diving header" (soccer), "double play" (baseball), etc. We're attempting to develop a clear and thorough explanation for learned (in contrast to imposed a priori) compositionality in representation. We emphasize that this problem, compositionality, is widely viewed throughout the machine intelligence community as a crucial and still largely uncracked problem. Moreover, learning compositionality qualifies as an instance of what the machine intelligence community terms, model selection, which is the search over all possible latent variable structures that could give rise to the observed data: since the number of possible structures is combinatorial, model selection is extremely difficult computationally.
You can use ctrl+<left/right-arrow> to step back/forward through frames. The videos are 128x128 (preprocessed from original KTH videos). Black pixels are those that are actively being represented by the mac in whose RF (cyan) the pixels are. Light gray pixels are those not being actively represented by any of the macs whose RFs are shown, but generally are being actively represented by other macs whose RFs are not shown (so as to avoid too much overlap between RFs).