Hierarchical Temporal Memory (NTM) and its self-learning algorithms

What is it and why is it all? This is the latest development of a very unfamiliar Jeff Hawkins associates, simulating the work of individual layers of the cerebral cortex. This gizmo allows (if you don’t mess up everything with the right way) to extract from the input data stream similar events, their sequences, to carry out their recognition and prediction. All who are interested in the details, you are welcome under Habrakat.


About Jeff himself to talk here for the hundredth tenth time is meaningless, but if someone has not read his book "On Intelligence" ("On Intelligence", the Internet delivers), then it is desirable to read it, otherwise the "lyrical" context of the whole thing can be for you are foggy (if you need it at all).

The presented document fully describes the construction of the Hierarchical Temporal Memory (NTM) and its cortical (I mean, “brain-like”, forgive me, O Great and Powerful) algorithms of (self) learning. The authors (and the translator) sought to describe everything very clearly and in detail, in the spirit of “take and experiment,” which is recommended. It seems that no prior knowledge of contact with a Higher Mind is required for this. Moreover, according to the experience of running this translation on one of the AI ​​forums, the presence of fragmental knowledge on the topic generates excessive (co) suspiciousness and allows you to do nothing in practice, which leads to rotting of the ET, habrayuzer.

However, for all those interested in neurophysiology, “how is it arranged in the real brain?”, There are a couple of (most) interesting applications in the document that explain, in the authors' opinion (in fact, Hawkins himself), how their NTM correlates with the work real neurons in the natural thinking brain. In particular, a hypothesis has been put forward, and why in general there are 6 (or so, in some places) layers in the cerebral cortex and what they generally do there.

In general, a very interesting work on the topic of AI for theorists and practitioners. I would venture to suggest that it will eventually become a kind of “classic” in this area, such as Kohonen’s SOM. The company Numenta continues to develop the described systems and algorithms, advertises them where possible and hopefully will surprise us all with its achievements in this area (a startup by the way).

All Authors, as one translator who joined them, sincerely hope that the presented document will be useful to you in your developments in the field of AI and not only.