Unfortunately the times haven’t worked for my schedule. I’ll try and make it to some of them when I can.
There are two, one in the US (bitking participates there)
There is a saying: “all models are wrong, but some are useful”. The HTM model is very minimal. It’s wrong because it leaves out a great many details, but its useful because it demonstrates a handful of key principles.
4 Likes
Thank you for the feedback regarding BAMI. Our understanding of predictive states has evolved over time and it looks like there are some mistakes in BAMI. You are correct - in the temporal memory, cells don’t become active due to lateral connections, they become depolarized (i.e. become predictive). This depolarization is an internal state of the neuron, and not visible externally. The figure is also a bit misleading. We will make the updates accordingly - I’ll post any updates to this thread.
Prediction is any change in state, in anticipation that something is about to happen based on other contextual input. In some cases, the predictive state can be purely internal to the neuron, which is what we call passive predictions. We also have active predictions, perhaps that’s what you mean by a “reactive framework”. You can ask me to explicitly make a prediction of something and in this case, my neurons would have to be active. On Intelligence didn’t really make a distinction between the two types of predictions.
3 Likes
I’m glad the feedback is useful.
By reactive I was thinking the output of a macrocolumn reflects the current input in the context of the history of the input. This constraint applies to each level in the hierachy so the system output is not predictive (i.e. reactive).
Which parts of HTM do you have in mind regarding the “active predictions” ?
I believe @clai is referring to the fact that the biological system clearly has capabilities that requires active predictions, and thus those will inevitably become part of the theory, since its goal is to model the biology. I know the concept of active predictions have been brought up numerous times in research meetings, but as far as I know nothing is currently formalized into official HTM theory yet (though I would be happy to learn otherwise, since it would give me some new material to dig into).
Did you/are you going to start a thread on your work on predictive coding?
1 Like
(if that was a question for me) Not yet (and BTW to make sure I am using established terms correctly, I’ll stay away from referring to “predictive coding”, and instead just use the term “active predictions”). That thread will be about Temporal Unfolding, but it requires establishing a couple of other concepts first. My first thread I mentioned that I am working on will be describing my implementation of the “output layer” (that is a necessary component, since before you can “unfold” something, you first have to “fold” it).
It is taking me longer than I anticipated, because illustrations alone are not enough to explain a couple of the more important dynamics of competing hex grids. I spent some time trying to simplify, but I’m not the best teacher, so I eventually had to give that up (I was just making the whole thing seem more confusing than it really is). Instead I’ve decided to for a couple of points in the thread, to link to demos that I’m building in HTM.js (rather than presenting the whole concept entirely in theory and illustration). Hopefully those in conjunction with discussion on the thread will convey the concept better.
1 Like