Glossary

B

Belief
Within the context of an HTM, a belief is the probability distribution on a cause or set of causes. Specifically, belief refers to the distribution over a set of potential causes once all top-down, bottom-up and lateral evidence has been considered.

Bindings
Exposure to a target language of APIs originally written and implemented in a (different) source programming language. For example, Numenta Tools has Python bindings to a C++ library. In this case, Python is considered the target language, and C++ is the source.

Bundle
Collection of files stored in a single directory hierarchy. On some operating systems, the bundle can be made to appear as a single file. In Numenta Tools, a session bundle holds all files associated with a single conceptual NRE session.

C

Category
The top-level, distinct class to which entities or concepts belong.

Category file
A category file classifies training data.

Cause
An object in the world. From the HTM perspective, what's important about the objects in the world is that they have persistence, that is, they exist over time. A cause is not necessarily a physical object.

Classification
Classification is first performed during training: the HTM system is presented with a category file, which maps training data to categories. After that, the HTM system is presented with new data and can decide on the closest category match for each pattern.

Client code
Code that is accessing an API. Client code is not part of the API or its implementation. Client code includes software developed internally by Numenta engineering or QA departments, or may be developed externally by customers.

Cluster
Set of servers networked together using Ethernet or other networking protocols.

Coincidence
A coincidence is the noteworthy alignment of two or more events or circumstances without obvious causal connection. In the context of an HTM Network, a specific combination of patterns that are likely to occur together at one point in time.

Coincidence Detection
The process of detecting frequently occurring coincidences among input patterns.

Coincidence Matrix
A matrix of the coincidences the HTM system found after performing learning at one level.

Confusion matrix
The confusion matrix allows you to see how many items were assigned to which category. You run the NetConfusion.py script to get a confusion matrix.

CPU
One-processor core. A server can have multiple CPUs per server or per chip. For example, a server with two dual-core chips has a total of four CPUs.

E

Effector
Effectors are nodes that receive the output of the top-level node as input. The effector might send the output to a file or hardware device.

F

Fan-in
Fan-in refers to the number of outputs leading to one input of a node.

Fan-out
Fan-out refers to the number of outputs going from a node to the inputs of other nodes.

G

Geometry
The network geometry specifies the number of levels and for each level the node parameters such as fan-in that determine how nodes are linked.

Group
A set of coincidence patterns that are likely to occur close together in time.

Grouping
Process of creating groups.

H

HTM
Hierarchical Temporal Memory. Theory describing the structural and computational properties of the neocortex.

HTM Network
A set of nodes, sensors, and effectors connected to perform a specific function. Serve as the HTM structure that is being computed by the NRE.

HTM System
A complete system for running HTM Networks consisting of software and hardware components.

I

Inference
Inference is the act or process of deriving a conclusion based solely on what one already knows. In the context of the Numenta platform, it can mean that during training, nodes can infer for example, the likelihood that a certain item is the next item in a sequence based on other sequences it has seen. After the HTM Network has been trained, you can feed it new data and the HTM can infer the corresponding category (as a statistical pattern).

Input
Any node can receive input from all nodes to which it is linked. A node can have multiple inputs.

Invariance
Occurs when a belief is unchanged by a wide range of real-world transformations, often those which cannot be specified in concrete mathematical terms.

L

Launcher
The Launcher process is part of the NRE. The process runs only briefly as it launches the NRE. As a rule, users don't interact with the launcher directly.

Learning
A node is in the learning state when it is receiving inputs, measuring the statistics of the inputs, and making modifications to its internal structures to represent the statistics of the inputs.

Learned State
Portion of a node's static state that is updated when learning occurs.

Link
Connection between nodes in an HTM Network.

M

maxDistance
The maxDistance parameter sets the maximum Euclidean distance at which two input vectors are considered the same during learning. When you set maxDistance to a higher number, you're more likely to get matches even if the noise-level is high. However, if maxDistance is too high, items that actually belong to different groups can end up in the same group.

N

NetExplorer tool
Numenta tool that allows you to test your HTM Network with different parameters and data and to see the results using gnuplot.

Network
The Python Network class implements an HTM Network.

Node
A node is the basic computational unit of an HTM Network. Node types include sensor, effector, and learning node. A learning node learns and represents the spatial and temporal statistics of the inputs to which it is exposed.

Node Input
See Input

Node Output
See Output

NP (Node Processor)
Software component that is responsible for running and scheduling a portion of an HTM Network.

NuPIC
Acronym for Numenta Platform for Intelligent Computing.

NRE
Numenta Runtime Engine. Software executables required for running HTM Networks. The NRE consists of the NP and the Supervisor.

NSAP (Numenta Supervisor Access Protocol)
A sockets-based protocol for communicating with the Numenta Supervisor. NSAP is a component of the Runtime API. Most developers don't use this protocol directly.

Numenta Network File Format
When you save a network after constructing it, or when you save a trained network, it is saved in Numenta Network File Format. Files in this format are in XML. If you modify an NFF file explicitly, it might no longer load. Use the tools for modification instead.

Numenta Platform for Intelligent Computing
Full name for the Numenta software platform. Includes the runtime engine and Numenta tools. Abbreviated NuPIC.

O

Output
The node output is the part of the node's state that's accessible by other nodes. Outputs can be arbitrary data types. Each node can have multiple named outputs. When a node is in inference node, it makes outputs available to other nodes.

P

Phase
A node's phase determines when the NRE executes it. You can specify the phase for each level during node creation.

Pipeline Scheduler
The pipeline scheduler is a high-performance node scheduler that can be used with feed-forward networks. The pipeline scheduler double-buffers node outputs and pipelines computation so that all nodes can be computed concurrently, making the pipeline scheduler ideal for multiprocessing.

Plugin
A plugin (or plug-in) is a computer program that interacts with a main application (a web browser or an email program, for example) to provide a certain, usually very specific, function. Numenta supports a node plug-in API that allows licensed users to create custom nodes.

Process
Single instance of a running program. Occupies system memory for program code, variables and objects.

R

Region
Regions are groups of nodes that all have the same configuration. Regions simplify common network topologies. Within a region, parameters cannot vary.

Runtime Engine
See NRE

S

Scheduler
The scheduler you choose determines the order in which nodes are executed. The basic scheduler uses phases. Advanced users can work with the pipeline scheduler in a multiprocessing environment.

Sensor
Input to HTM Networks. Sensors interface to external files, hardware devices, etc. and format data for input to other nodes.

Server
Physical computer containing one or more CPUs, hard drive, and power supply.

Session
A session includes all input data, output data, and interaction involved in a single use of the NRE. You can create and modify a session using the Python Session interface.

Session Bundle
See Bundle.

Static State
Portion of the node state that is independent of the runtime state of the system.

Supervised Learning
During supervised learning, the HTM Network is fed data and corresponding category information to learn the mapping between data and categories. After that, the HTM Network can perform inference on new data.

Supervisor
Portion of the NRE responsible for coordinating one HTM Network, and for communicating with external applications (e.g. the tools).

Supervisor command set
Portion of the NRE responsible for coordinating one HTM Network and for communicating with external applications (e.g. the tools).

Supervisor Command
Text commands for controlling the Supervisor. The command set is a subset of the API.

T

Time Adjacency Matrix
The system forms a time adjacency matrix by observing coincidences over time. The system uses that matrix to group the coincidences into temporal groups.

Tools (Numenta Tools)
Collection of software libraries, language bindings, and applications. Numenta tools provide access to the NRE, offer additional HTM related features, and are used by HTM applications.

Training
To train your HTM Network, you invoke the NRE with the network configuration and the training data. During training, the nodes in the HTM Network perform learning and inference.

U

Unsupervised Learning
During unsupervised learning, you feed data to the system without providing category information.

Visualizer Tool
The Numenta Visualizer tool allows you to examine a node to analyze its performance. It generates an HTML page for each node in the network, displaying groups and coincidences as well as general statistics.