distributed-hierarchical-temporal-memory-dhtm

name	distributed-hierarchical-temporal-memory-dhtm
description	Distributed Hierarchical Temporal Memory (D-HTM) neuromorphic framework enabling cross-entity preemptive warning via Shared Associative Memory (SAM) — extends HTM beyond reactive detection to distributed predictive reasoning (arXiv: 2606.31789)
version	1.0.0
created	2026-07-01T00:00:00.000Z
author	Hermes Agent
category	neuroscience
tags	["neuromorphic","hierarchical-temporal-memory","anomaly-detection","sparse-distributed-representations","predictive-warning","time-series"]
activation	D-HTM, distributed HTM, shared associative memory, SAM, cross-entity warning, neuromorphic anomaly detection, SDR, temporal memory, preemptive warning
arxiv_id	2606.31789
arxiv_url	https://arxiv.org/abs/2606.31789

Distributed Hierarchical Temporal Memory (D-HTM) for Cross-Entity Preemptive Warning

Overview

D-HTM is a neuromorphic framework that extends Hierarchical Temporal Memory (HTM) by introducing a Shared Associative Memory (SAM) for cross-entity knowledge transfer. While traditional HTM operates independently on each data stream (reactive), D-HTM enables preemptive warning — issuing alerts before local anomaly onset by reusing precursor signatures learned from related entities.

Key Innovation: Demonstrates that transferable precursor structure can emerge within shared Sparse Distributed Representation (SDR) space and be reused for predictive reasoning across distributed systems.

Paper: arXiv:2606.31789 (2026-06-30) Authors: Pavia Bera, Jennifer Adorno, Sanjukta Bhanja

Core Architecture

D-HTM consists of three integrated components:

1. Spatial Pooler (SP) — Shared SDR Projection

• Projects observations from multiple entities into a common SDR space
• Uses HTM's canonical Spatial Pooling algorithm: overlap → inhibition → activation
• Creates a shared representation vocabulary enabling cross-entity comparison
• Key property: Similar precursors across entities map to overlapping SDRs

2. Temporal Memory (TM) — Entity-Specific Dynamics

• Each entity has its own TM module learning sequence statistics online
• TM maintains context via active dendrite segments with permanence values
• Learns transitions: given current SDR context, predict next SDR
• Detects anomalies when predictions fail (low overlap with actual input)
• Online learning: No offline training required; adapts continuously

3. Shared Associative Memory (SAM) — Cross-Entity Knowledge Transfer

• Novel component: Stores recurring pre-anomaly signatures shared across entities
• When TM detects precursor pattern in one entity, SAM can retrieve and check against other entities
• Enables preemptive warning: entity B gets alert before local anomaly based on entity A's precursor
• Maintains online learning capability of individual HTM modules

Methodology

Step-by-Step Pipeline

1. Input encoding: Convert multivariate time series observations to binary SDRs via SP
2. Temporal learning: Each entity's TM learns sequence transitions online