VSABrains

VSABrains

A discrete, CPU-first learning architecture inspired by A Thousand Brains (Hawkins et al.). Demonstrating that robust intelligence can emerge from parallel models operating in reference frames, achieving consensus through voting mechanisms.

Interactive Tutorial

A visual, step-by-step lesson that explains how the core pieces fit together: grids, displacement, multi-column consensus, replay, and verifiable reasoning.

Open the Tutorial
Runs entirely in the browser (canvas + optional narration). Designed to work on mobile too.

Plain-Language Guides

Start here if you want the big picture without ML jargon.

How VSABrains Works
Step-by-step walkthrough of ingestion, replay, and verification.
Demo Guide
A simple walkthrough for using the demo and reading its animations.
Frames & CNL
What frames are, how they are stored, and how to read the demo animation.
Glossary
High-school level definitions of core concepts and terms.
Why VSABrains?
Plain-language answers: what this architecture buys you and what it does not.

Design Specifications

The project is documented through a small set of design specifications that cover architecture, core algorithms, integration details, implementation, and evaluation. Each specification builds on the core insight: order as address, not superposition.

DS001 - Vision
Architecture overview and rationale. Why addressable memory avoids the "muddiness" problem of global superposition schemes.
DS004 - Core Algorithms
Core discrete runtime: GridMaps, displacement, localization, replay/checkpoints, reasoning primitives, diagnostics.
DS005 - Integrations
Text ingestion, LLM-backed extraction, fact validation, and verifiable answer contract for Exp3-style grounded RAG.
DS006 - Semantic Encoder (CNL)
CNL-driven frame definitions and semantic encoder contract for literature and dialogue analysis.
DS007 - Exp6 Literature & Dialogue
Evaluation design for semantic literary and conversational queries using virtual spaces.
DS002 - Implementation Plan (Index)
Reading order and index linking to DS002b (Implementation Guide) and DS002a (API Reference).
DS002b - Implementation Guide
Phases, entry points, end-to-end examples, testing strategy, and milestones.
DS002a - API Reference
File-level module APIs and data contracts to implement under src/.
DS003 - Evaluation Framework
Three experimental suites with metrics and success criteria for validating the architecture.
Glossary (in DS001)
Shared terminology: heavy-hitters, toroidal topology, coreference, work signatures, token/step concepts.
Why VSABrains?
Plain-language answers: VSA vs discrete maps, why final points are not meaning, how replay works, and how this differs from a simple list.
Frames & CNL
How semantic frames sit on top of events and enable higher-level queries.

Core Principles

The architecture is built on four foundational principles that distinguish it from traditional approaches like VSA (Vector Symbolic Architectures) or dense neural networks.

Order as Address
Temporal order is encoded as location in memory, not superposed content. Events are written along trajectories, making retrieval deterministic.
Multi-Column Consensus
Multiple parallel columns provide redundancy and multi-hypothesis tracking. Under ambiguity, consensus selects the most supported interpretation.
Addressable Memory
Episodic store with indexed retrieval enables long-context without keeping everything in RAM. Checkpoints enable replay-based verification.
Verifiable Reasoning
Every proposition traces to source facts. Outputs explicit verdicts: supported, conflicting, or unsupported.

Evaluation Experiments

The experiments validate core hypotheses. Each targets a specific architectural claim and provides quantitative success criteria.

Evaluation Status
Current pass/fail state mapped to DS003 success criteria.
Exp 1: Reference-Frame Alignment
Validates multi-column consensus under partial observability, noise, and ambiguous contexts. Target: consensus accuracy exceeds single column by 5%+.
Exp 2: Narrative Coherence
Tests state tracking with coreference, scene resets, and repetitive motifs. Target: graceful degradation, not catastrophic collapse.
Exp 3: Grounded RAG
Validates anti-hallucination: correct answers with evidence, refusal for unsupported queries, explicit conflict detection. Target: <5% hallucination.
Exp 6: Literature & Dialogue Semantics
Evaluates semantic queries about meaning, emotion, and discourse using virtual spaces against a naive list-scan baseline.

Deep Dives

Three high-signal pages that explain the “why” and show the current evidence.

Why VSABrains?
Plain-language answers: what this architecture buys you, what it does not, and how it compares to a simple symbolic list.
VSA vs VSABrains
A direct comparison: holographic superposition vs addressable trajectories, what we kept, what we changed, and why.
Evaluation Status
Experiment-by-experiment results, visualizations, and optimization sweeps.

Inspiration

The architecture draws from Jeff Hawkins' A Thousand Brains theory, operationalizing key concepts into discrete, auditable mechanisms:

Reference Frames
Internal location state on discrete GridMaps. Each column maintains its own coordinate system.
Movement Signals
Deterministic displacement updates location at each step, transforming temporal order into spatial trajectory.
Parallel Models
Multiple columns with different offsets provide redundancy and enable robust interpretation under ambiguity.
Voting Consensus
Weighted aggregation across column predictions. The ensemble outperforms any individual column.