Hi Alexandre, HTMs cannot currently use inductive reasoning. That requires knowledge about what a "9" or a "10" is... HTMs just see streaming SDRs and don't really care what the semantic meaning of any particular input is.
This is expected an just like a single Region in the human neocortex. It's only after Regions of Layers are combined into multi-region hierarchies that you begin getting abstractions regarding lower level input, and eventually you would get "reasoning behavior". At this point, hierarchy is currently a research topic and in "development" as there are many different nuances to doing it properly and Numenta will not release anything that it isn't pretty darn sure is actually what is happening in the brain. Now, the reason this works when combining HTMs with Cortical.io's Semantic Folding, is because Cortical.io encodes both subject and context in their Fingerprints (SDRs) - so therefore a similarity can be extracted from a feature embedded SDR by going back into the Cortical.io API and finding the "match" for the abstracted SDR (an SDR or Fingerprint that has many of the same bits as the animal resulting from the query).This information however "lives" in the Cortical encoding and is one of the advantages to Cortical's offering. Eventually, HTMs will be able to do this work - but as Numenta is just getting started with their technology, there remains some time before more organic features are included in the software. Cheers, David On Tue, Apr 26, 2016 at 11:05 AM, Alexandre Vivmond <[email protected]> wrote: > Does HTM's have the ability to do inductive reasoning based on previously > seen similar patterns? > Here's are my thoughts: say that an input file will only ever include > values between 1 and 10. The values 1 and 2 should have a lot of semantic > overlap compared to 1 and 10. Then, let HTM learn the pattern 1, 2, 3, 4, > 5, 1, 2, 3, 4, 5, ... may hundred times, and at one point suddenly see the > pattern 6, 7, 8, 9, X, can HTM inductively reason that the next value X > should be 10 given that the pattern is similar to 1, 2, 3, 4, 5 and that > the semantic overlap between 4 and 5 is similar to 9 and 10? I tried this > exact problem in Nupic and it unfortunately did not manage to predict the > never before seen values 6, 7, 8, 9, 10. If HTM/nupic can inductively > reason could you provide some examples that should/might work? > > The reason I'm asking is because I'm curious since the Fox eats example > (from the nupic hackathon 2013) is similar to my mentioned problem (in my > understanding at least), it managed to correctly predict what the fox eats, > even though HTM had never previously learned what foxes eat nor been put in > any context other than the SDR's semantic understanding that foxes are > similar to other animals that eat rodents. > -- *With kind regards,* David Ray Java Solutions Architect *Cortical.io <http://cortical.io/>* Sponsor of: HTM.java <https://github.com/numenta/htm.java> [email protected] http://cortical.io
