Hierarchical Temporal Memory (HTM) is a biologically-constrained theory of intelligence originally described in the book On Intelligence. HTM is not a Deep Learning or Machine Learning technology. It is a software intelligence framework strictly based on neuroscience and the physiology and interaction of pyramidal neurons in the neocortex of mammalian brains.
If you’re interested in learning more about HTM, visit our educational series HTM School or browse through the topics on HTM Forum. You can also find a collection of research papers at numenta.com/papers/ and a living book authored by Numenta researchers and engineers that documents HTM called Biological and Machine Intelligence (BAMI).
Note: You may notice old references in our documents to the "Cortical Learning Algorithm", or "CLA". These are simply older terms used to describe our Hierarchical Temporal Memory (HTM) technology, which we no longer use.
Components of HTM
For an overview and introduction to HTM Theory, watch HTM School.
Works in Progress
HTM Sensorimotor Inference
Numenta is researching how Layer 4 of the cortex might represent spatial objects. Jeff talks about this in the research meeting below:
See further discussion on this on our forums: Preliminary details about new theory work on sensory-motor inference
Rahul Agarwal, a former-Numenta engineer, gave a great introduction to some of the basics behind the CLA and its implementation within NuPIC. This is a great primer for anyone interested in learning more about how the neocortex works, as well as how it is implemented in NuPIC.
Subutai Ahmad, Numenta VP of Engineering, detailed some aspects of the CLA at our 2013 Fall Hackathon. He discussed an interesting property of SDR's affecting temporal pooling and hierarchies. The interactive session included a lot of Q&A. Slides.
Numenta engineer Chetan Surpur goes into great detail about the implementations of the spatial pooler and temporal memory in NuPIC.