Spreading activation: Seqsee has a web of concepts. Not a static, pre-defined web, but one that grows as Seqsee encounters new concepts as it solves sequences. When a node here is activated, it spreads activation to adjacent nodes. Many systems implement such “spreading activation”, but Seqsee differs in one crucial respect. As Seqsee works on a sequence, it is focusing on one object at a time and activation only spreads from this object rather than all activated objects.

In the debate, Bengio talked about how new information should result in only local changes, rather than changing the entire network. He called this the Consciousness Prior, and Seqsee’s spreading activation is thus analogous in a way other system’s spreading activation is not. In my dissertation, Section 7.1 lays out the benefits of this spreading strategy.

Thank you for pointing out the need to enlarge on that point: I was trying to keep the post short. I plan to write a series of posts on individual aspects of Seqsee, and I can be more detailed then.

I do unsupervised concept discovery at Pinterest (and previously at Google). Twitter: @amahabal