Brain Reasoning Models

---

Overview

This repository houses the Exact-Word Mathematical Model, an end-to-end framework built to demonstrate the organic arithmetic reasoning capabilities of Graph Neural Networks (GNN) + Tribev2 (by meta) scanning functional neural representations.

Instead of explicitly feeding the model billions of combinatorial equations (e.g., 123+456=579) or relying on external tokenizers, this architecture relies solely on a deeply curated dictionary of mathematical building blocks:

• Digits
• Linguistic Numeral representations
• Operators
• Functional Structural Words

By learning the biological and topological overlaps (Jaccard overlaps) of these core brain signatures, the Graph Convolutional Network (GCN) is able to generalize and synthesize the meaning of vastly complex mathematical thoughts purely by triggering its root arithmetic signatures simultaneously!

---

Architecture & Signature Readout

Wait, what does this model actually output? This decoder is structurally a Multi-label Signature Extractor. It doesn't classify one exact text string answer! If the continuous thought flow translates to thinking an equation like 24 + 5 = 29, the model simultaneously parses the input brain map and outputs a holistic probability activation spread:

  Active signatures detected:

    "24"       0.99  █████████████████████████ ← STRONG
    "+"        0.98  ████████████████████████  ← STRONG
    "5"        0.95  ███████████████████████   ← STRONG
    "29"       0.89  ██████████████████████    ← STRONG
    "="        0.85  █████████████████████     ← STRONG
    "plus"     0.52  █████████████             ← active (neighbouring synonym)
    "sum"      0.45  ███████████               ← active (neighbouring concept)

The system retrieves the full conceptual fingerprint behind the math, allowing you to synthetically reconstruct exactly what elemental digits, operators, and associated math-concepts triggered visually in the cortex at any given point in time!

---

Execution Pipeline

Step 1: Setup & Vocabulary Configuration Generate the ultra-lean mathematical token list (vocabulary.txt).

python generate_vocabulary.py --max-num 100

Creates the foundation of digits 0-100 and mathematical operators.

Step 2: Generate Functional Neural Activations Extract continuous functional neural mapping outputs (brain_signatures.h5) using TRIBE v2.

python generate_brainmaps.py --wordfile vocabulary.txt --mode tribev2

Highly stable script with built-in checkpointing; can resume seamlessly if disconnected from cloud GPU runtimes.

Step 3: Construct the Wernicke Graph Connect mathematical tokens topologically—edges are built organically if functional neural representations fire heavily in the same vertices. Output: signature_graph.pt.

python build_signature_graph.py

Step 4: Train the Brain Signature Network Train the multi-label Graph Convolutional Network. Because the vocabulary size was reduced to its purest fundamental building blocks, training the entire topological graph architecture takes less than a minute!

python gnn_cora.py --mode signature

Step 5: Interactive Temporal Inference Launch the REPL environment to type out experimental mathematical equations, instantly retrieve their simulated brain-activation, pass it to the GNN, and watch which math functional triggers fire out of the chaos.

python predict.py

---

🛠️ Performance & Scalability

This mathematical execution pipeline is primarily structured to parse the structural composition of pure mathematics dynamically. It proves that a GNN operating on biological fMRI representations doesn't need to have seen "42+17=59" to know how to decode it—if it has seen the organic signatures for 42, +, 17, and 59 independently, the neural topology effortlessly bridges the arithmetic gap during inference.

Result

Tribe v2

https://ai.meta.com/blog/tribe-v2-brain-predictive-foundation-model/ This project uses TRIBE v2 by Meta FAIR for research purposes. you can learn about tribe v2 model by meta which generates the brain data for each dataset