I am not sure if this question has been asked on the forum , but this bugs me a lot , Why isn’t Numenta using HTM theory and all the algorithms build an object recognition and Natural Language Understanding product ?
Our mission at Numenta is dual:
- understand how intelligence works in the mammalian neocortex
- create software based upon these principles
In following this mission, we find ourselves researching instead of working on products. We hope others will create products with our technology. We have simple licensing options.
But you guys have built Grok , Stock Predictor Software and among others , if you are gonna be invested in creating products based on HTM and its principles then why not create products that will disrupt the whole AI community and show them this approach is right ? like building a Object Recognizer or a NL Recogniser ?
Perhaps, but we are not making any assumptions about what capabilities the sensorimotor inference capabilities will give us yet. There are details that still need to be worked out.
We will eventually be testing this out on real world data, trying to apply it to the application space, but these will just be “example applications”. We want to show what you can do with HTM of course, but we don’t want to become a product company.
Matt’s right, but I can add a little bit more color. Our first priority is brain theory. We are making excellent and exciting progress on this right now. We don’t feel we can do both the basic research and product development simultaneously. Yes, we did more of the product development in the past, but we found it slowed down the research. We had to choose.
It would be fun apply HTM theory to build more products, but we take a long-term view. We have a real chance at completing a comprehensive theory of how the neocortex works and that will be our focus for the near future. I remain confident that our work will be at the core of AI, but we want to complete the theory first.
I hope that helps
Your question used to bug me a lot, too and I was also frustrated with the lack of companies that shared Numenta’s vision. However, I began to realize that building a product based on the applied principles of intelligence is extremely difficult these days because the concepts are still in their infancy. No product on this Earth can meet the mission. In my opinion this is why almost no companies devote the necessary capital and patience to get there. Therefore, as of now it is just research and a dream.
An effective analogy would be to compare the current state of HTM theory to the Perception in the 1950s, or perhaps a little bit after the idea began to solidify. Back then the concept was just research, couldn’t do anything impressive, and was outshined by other areas of technological interest. However, eventually its development and evolution gave rise to today’s deep learning, it’s various techniques, and the amazing technologies it has produced so far. As Jeff and Matt point out, HTM theory’s success is a long term investment.
As of now the theory is much like an unsupervised LSTM, it can learn sequences of patterns in various contexts albeit HTM has a few advantages discussed in one of their papers. If Numenta developed an awesome product just based on today’s ideas, I doubt the AI community would be impressed. The product would be just be more of the same. The real prize is building a comprehensive theory of the neocortex and applying it in software. Very few in today’s AI community have this vision in mind, but the implications and applications are staggering.