Jeff,I have watched the progress of your teem and effort as a little bit of an outsider, but with great interest. I have a question that may be a little outside of the current thread of discussion. How do you think the cortex, with all it’s input regions and layers performing different functions, operates at rest? In other words, without temporal input, what should the whole system be doing? For example, should it (layers) replay sequences and re-enforce memories, or improve on predictions based on new data? What do you think the hierarchy should do at rest?Joe Neff
This isn’t something we spend much time thinking about. “At rest” can mean a lot of things, sleeping, thinking quietly, and concentrating on something to the exclusion of other input. Sleeping is an entire field of research which suggests there are some important memory retention functions that occur during REM sleep. Thinking without input is something we all do every day. That would be part of the theory we are interested in. Basically, the cortex learns a predictive model and it can generate predictions and use them as input. Although we haven’t discussed this much we do think about it a lot.
Perhaps you are referring to what is termed the “Default Mode Network” (DMN).
This is the widely distributed pattern of sychrony involving many regions of cortex we see becoming active when a subject is daydreaming, or otherwise not focused on any particular task. This synchrony is perturbed when the subject attends to a particular task. There is a wide range of speculation as to why the brain behaves this way - ranging from it being a sort of base carrier signal (Olaf Sporns et al) to being involved in the formation of or reorganization of long term memories, to being involved in the attention system, and some even claim it does not exist at all (i.e. the fMRI imaging shows merely areas where larger blood vessels are anatomically arranged and for which there is some correlation).
However, the DMN remains speculative and has no corresponding mechanism in HTM that I’m aware of.
Thank you for your thoughtful feedback (and time). I think my question is more aligned with ‘thinking’ without input, or perhaps more accurately, the HTM system ‘wandering’ in-between clock cycles (or lulls in data stimulus). What is the best use of that ‘wandering’ time? The HTM could be replaying sequences and re-enforcing synapses to minimize anomaly detection (is this a good thing - I don’t know?). Or perhaps re-shuffling the encoder and the sparse representation to make more efficient use of the HTM and better prepare it for new un-experienced sequences? Food for thought! Many thanks and Best regards,
In my opinion, if I observe myself… experientially there is no “lull” time. It seems to me that we fill in objectively determined attention with attention to our own internal language or conversation. But I don’t think (at least not at this time), that it ever stops. Just my personal opinion for the moment.