Hierarchical temporal memory (HTM) is a biologically constrained machine intelligence technology developed by Numenta. Initially described within the 2004 ebook On Intelligence by Jeff Hawkins with Sandra Blakeslee, HTM is primarily used at this time for anomaly detection in streaming data. The technology is based on neuroscience and the physiology and interplay of pyramidal neurons within the neocortex of the mammalian (particularly, human) brain. At the core of HTM are learning algorithms that can retailer, learn, infer, and Memory Wave recall excessive-order sequences. Unlike most different machine learning strategies, HTM consistently learns (in an unsupervised course of) time-primarily based patterns in unlabeled data. HTM is sturdy to noise, and has excessive capacity (it could actually be taught a number of patterns concurrently). A typical HTM network is a tree-formed hierarchy of ranges (not to be confused with the "layers" of the neocortex, as described under). These levels are composed of smaller parts called regions (or nodes). A single stage within the hierarchy probably accommodates several areas. Higher hierarchy ranges usually have fewer areas.

Higher hierarchy ranges can reuse patterns learned at the decrease levels by combining them to memorize extra complicated patterns. Each HTM region has the same fundamental function. In studying and inference modes, sensory knowledge (e.g. knowledge from the eyes) comes into bottom-level regions. In technology mode, the bottom level areas output the generated sample of a given class. When set in inference mode, a region (in each stage) interprets information arising from its "little one" areas as probabilities of the classes it has in Memory Wave brainwave tool. Every HTM region learns by figuring out and memorizing spatial patterns-mixtures of enter bits that often occur at the identical time. It then identifies temporal sequences of spatial patterns that are more likely to occur one after another. HTM is the algorithmic element to Jeff Hawkins’ Thousand Memory Wave brainwave tool Brains Concept of Intelligence. So new findings on the neocortex are progressively incorporated into the HTM model, which adjustments over time in response. The brand new findings do not essentially invalidate the earlier parts of the mannequin, so ideas from one technology usually are not essentially excluded in its successive one.

Throughout training, a node (or region) receives a temporal sequence of spatial patterns as its input. 1. The spatial pooling identifies (within the enter) continuously noticed patterns and memorise them as "coincidences". Patterns which can be considerably related to one another are treated as the identical coincidence. A large number of potential enter patterns are decreased to a manageable number of identified coincidences. 2. The temporal pooling partitions coincidences which might be more likely to comply with each other in the training sequence into temporal groups. Every group of patterns represents a "cause" of the input pattern (or "name" in On Intelligence). The ideas of spatial pooling and temporal pooling are nonetheless fairly necessary in the current HTM algorithms. Temporal pooling is just not yet well understood, and its that means has modified over time (because the HTM algorithms evolved). Throughout inference, the node calculates the set of probabilities that a sample belongs to each known coincidence. Then it calculates the probabilities that the input represents every temporal group.

The set of probabilities assigned to the groups is named a node's "perception" in regards to the input sample. This perception is the result of the inference that's passed to a number of "mother or father" nodes in the next larger level of the hierarchy. If sequences of patterns are much like the training sequences, then the assigned probabilities to the teams is not going to change as usually as patterns are received. In a extra normal scheme, the node's belief might be despatched to the enter of any node(s) at any level(s), however the connections between the nodes are nonetheless fixed. The upper-degree node combines this output with the output from other child nodes thus forming its own input sample. Since decision in area and time is misplaced in every node as described above, beliefs formed by increased-stage nodes characterize a fair larger range of house and time. This is supposed to reflect the organisation of the physical world as it is perceived by the human brain.