Hierarchical Temporal Memory (HTM)

The fundamental data structure in HTM, representing information as a sparse binary vector where only a small percentage of bits are active (e.g., 2%). This mimics the sparse firing patterns of neurons in the neocortex. Key properties include:

• Noise Tolerance: The meaning is distributed, so the representation remains stable even if some bits are corrupted.
• Capacity: A single SDR can represent a vast number of unique patterns.
• Semantic Similarity: Similar concepts are represented by overlapping sets of active bits, enabling efficient similarity and union operations.

The process that converts raw, noisy input data into a stable Sparse Distributed Representation (SDR). It performs three key functions:

• Mapping: Creates a fixed representation for an input, regardless of minor variations.
• Sparsification: Ensures only a small percentage of columns/neurons become active.
• Overlap Preservation: Maintains semantic similarity; inputs that are similar produce SDRs with overlapping active bits. This stage provides invariance, allowing the system to recognize the same concept presented in slightly different forms.

The core learning mechanism that discovers and predicts temporal sequences in the SDR input stream. It models the contextual relationships between patterns over time.

• Neurons learn to recognize specific temporal contexts by forming dendritic segments.
• When a pattern is observed, the algorithm activates neurons that were predictive of that pattern based on the recent context.
• It learns transition probabilities between patterns, enabling it to make multi-step predictions about likely future inputs. This is the basis for HTM's anomaly detection capability.

HTM networks are organized into layers and regions that form a hierarchy, mirroring the neocortical columnar structure.

• Lower levels process fine-grained, concrete details and short-term patterns.
• Higher levels receive feed-forward inputs from lower levels, learning more abstract and stable representations over longer time scales.
• Feedback connections from higher to lower levels provide context, refining predictions at lower levels. This hierarchy enables the system to build complex, multi-scale models of the world from simple sensory inputs.

HTM is designed for continuous, unsupervised learning from non-stationary data streams. It does not require distinct training and inference phases.

• Learning happens incrementally with each new input.
• The model adapts continuously to changing statistics in the data without catastrophic forgetting of previously learned sequences.
• This makes it suitable for real-time applications like monitoring sensor data, network traffic, or financial tickers, where the underlying patterns may evolve.

A primary application of HTM is real-time anomaly detection. The system is constantly making predictions about the next expected input.

• A low prediction score indicates the current input was not well-predicted given the recent context, flagging it as a potential anomaly.
• This is fundamentally different from statistical outlier detection; it is based on the violation of learned temporal patterns.
• The hierarchical nature allows detection of anomalies at different levels of abstraction, from simple sensor faults to complex behavioral shifts.

A Sparse Distributed Representation (SDR) is the fundamental data structure in HTM theory, representing information as a binary vector where only a small percentage of bits are active (e.g., 2%). This mimics the sparse firing patterns of neurons in the neocortex. Key properties include:

• Capacity: An SDR of n bits with w active bits can represent a vast number of unique patterns (combinatorial capacity).
• Noise Tolerance: Two SDRs can be compared via overlap (intersection), making the representation robust to noise and partial information.
• Union Property: Multiple SDRs can be combined into a single SDR via a logical OR, preserving the original patterns without catastrophic interference. HTM uses SDRs for inputs, outputs, and the internal states of its columns and cells.

The Spatial Pooler is the first stage in an HTM region, responsible for converting raw input data into a stable Sparse Distributed Representation (SDR). It performs unsupervised learning of spatial patterns. Its core functions are:

• Mapping: Takes a potentially dense or overlapping input and maps it to a fixed-size SDR output.
• Learning: Uses a competitive Hebbian learning rule (boost and punish) to form connections between input bits and a set of HTM columns. Columns that frequently activate for a given input pattern strengthen their connections.
• Stability: Ensures the same input always produces a similar output SDR, while different inputs produce dissimilar SDRs.
• Boosting: Dynamically adjusts column activation rates to ensure all columns are used efficiently, preventing a few from dominating.

The Temporal Memory algorithm is the core of HTM's sequence learning capability. It models neurons with dendritic segments and predictive states. Operating on the SDRs from the Spatial Pooler, it:

• Learns Sequences: Models the activation of cells in columns over time. When a cell becomes active, it forms connections to cells that were active in the previous time step.
• Makes Predictions: Cells can enter a predictive state based on prior activity. If a column receives input that matches a prediction, only the predicted cell activates, forming a continuous representation of the sequence.
• Uses Context: The current activity is a function of both the present input and the recent past (context), enabling it to disambiguate identical inputs that occur in different sequences. This mechanism allows HTM to learn high-order sequences and make multi-step predictions in noisy data streams.

Cortical Learning Algorithms (CLA) is an earlier name for the suite of algorithms that constitute HTM. It emphasizes the biological inspiration drawn from the neocortex. The key principles include:

• Uniform Algorithmic Fabric: The hypothesis that a single, powerful algorithm (based on SDRs, spatial pooling, and temporal memory) is replicated throughout the cortical sheet, processing different modalities (vision, audio, touch).
• Hierarchical Processing: Real-world HTM systems are built as hierarchies of regions, where lower regions feed SDRs representing simpler features to higher regions that learn more complex, invariant representations.
• Online Learning: Learning occurs continuously from a never-ending stream of data, without separate training and inference phases. The term CLA underscores the goal of reverse-engineering cortical computation for machine intelligence.

Sequence Memory is the primary capability enabled by the Temporal Memory algorithm. It refers to a system's ability to:

• Encode: Store the temporal order of patterns in a way that captures transitions and context.
• Recall: Reproduce or continue a sequence given a starting cue or partial input.
• Predict: Anticipate the next element(s) in an ongoing sequence.
• Recognize: Identify which known sequence is currently being observed, even with noise or missing elements. In HTM, sequence memory is auto-associative and hierarchical. Lower-level sequences of sensory features become elements in higher-level sequences, allowing the system to build complex models of temporal structure, such as grammar in language or routines in sensor data.

Anomaly Detection is a major practical application of HTM systems. Due to their strong temporal prediction capability, HTMs excel at identifying deviations from learned patterns in real-time data streams. The process is intrinsic:

1. The Temporal Memory is constantly in a predictive state, generating an SDR of expected next inputs.
2. When new sensory input arrives, it is encoded into an SDR by the Spatial Pooler.
3. The anomaly score is calculated based on the overlap (or lack thereof) between the predicted SDR and the actual input SDR.
4. A low overlap indicates an unexpected event—an anomaly. This approach is unsupervised and online, requiring no pre-labeled anomalous data. It is used for monitoring IT infrastructure, financial transactions, and industrial sensor networks, where it can flag novel faults or fraud as they occur.