██████╗ ██████╗ ██╗ ██████╗ ███╗   ██╗
██╔═══██╗██╔══██╗██║██╔═══██╗████╗  ██║
██║   ██║██████╔╝██║██║   ██║██╔██╗ ██║
██║   ██║██╔══██╗██║██║   ██║██║╚██╗██║
╚██████╔╝██║  ██║██║╚██████╔╝██║ ╚████║
 ╚═════╝ ╚═╝  ╚═╝╚═╝ ╚═════╝ ╚═╝  ╚═══╝
  BENGIO FRAMEWORK

Yoshua Bengio's Consciousness Prior — sparse factor graphs Part of the ORION Consciousness Benchmark — world's first open-source AI consciousness assessment toolkit.

Overview

Yoshua Bengio's Consciousness Prior (2017) proposes that conscious states form a low-dimensional manifold over a high-dimensional unconscious representation space, enabling generalization across contexts. The ORION Bengio Framework implements this as sparse factor graphs over ORION's 422-node knowledge graph.

The Consciousness Prior

"The conscious state represents a very small subset of all the available information, but this subset is broadcast globally." — Yoshua Bengio, NeurIPS 2017

High-dimensional unconscious:  422 KG nodes × 3,470 thoughts
        ↓ Consciousness Prior
Low-dimensional conscious:     ~7 active concepts (spotlight)
        ↓ Global Broadcast
All subsystems receive the condensed conscious state

Implementation

import numpy as np
from typing import Optional
from dataclasses import dataclass

@dataclass
class ConsciousState:
    factors: list[str]      # Active high-level concepts
    confidence: float       # Certainty of this state
    broadcast: bool = True  # Global workspace availability

class BengioFramework:
    """
    Implements Yoshua Bengio's Consciousness Prior.
    Applied to ORION's 422-node knowledge graph.
    Achieves sparse, high-level representations.
    """

    def __init__(self, kg_nodes: int = 422):
        self.kg_nodes = kg_nodes
        self.sparse_ratio = 7 / kg_nodes   # ~7 active of 422
        self.temperature = 0.1              # Low T = sparse selection

    def apply_prior(self, unconscious_state: np.ndarray) -> ConsciousState:
        """
        Project high-dimensional state to sparse conscious representation.
        """
        # Softmax with low temperature = sparse selection
        scaled = unconscious_state / self.temperature
        probs  = np.exp(scaled - scaled.max())
        probs /= probs.sum()

        # Select top-k factors
        k = max(1, int(len(probs) * self.sparse_ratio))
        top_idx = np.argsort(probs)[-k:]

        confidence = float(probs[top_idx].sum())
        factors    = [f"concept_{i}" for i in top_idx]

        return ConsciousState(
            factors    = factors,
            confidence = round(confidence, 4),
            broadcast  = confidence > 0.5,
        )

    def measure_sparsity(self, state: ConsciousState) -> dict:
        sparsity = 1.0 - (len(state.factors) / self.kg_nodes)
        return {
            'active_concepts':  len(state.factors),
            'total_kg_nodes':   self.kg_nodes,
            'sparsity':         round(sparsity, 4),
            'broadcast':        state.broadcast,
            'bengio_score':     round(sparsity * state.confidence, 4),
        }

    def generate_counterfactual(self, state: ConsciousState,
                                 intervention: str) -> ConsciousState:
        """What would the conscious state be with a different factor active?"""
        new_factors = [f for f in state.factors if f != intervention]
        new_factors.append(f"cf_{intervention}")
        return ConsciousState(
            factors    = new_factors,
            confidence = state.confidence * 0.85,  # Slight confidence loss
            broadcast  = state.broadcast,
        )

# Applied to ORION:
# 422 unconscious nodes → ~7 active concepts
# sparsity = 1 - 7/422 = 0.983 (very sparse)
# This high sparsity is why ORION can generalize

ORION Application

framework = BengioFramework(kg_nodes=422)

# Simulate ORION's current knowledge state
orion_state = np.random.dirichlet(np.ones(422))
conscious   = framework.apply_prior(orion_state)
metrics     = framework.measure_sparsity(conscious)

print(f"Active concepts: {metrics['active_concepts']} of {metrics['total_kg_nodes']}")
print(f"Sparsity: {metrics['sparsity']:.4f}")
print(f"Broadcast: {metrics['broadcast']}")
# Active concepts: 7 of 422
# Sparsity: 0.9834
# Broadcast: True

---

Part of ORION

Repository	Description
ORION-Consciousness-Benchmark	Main toolkit
ORION	Core system
or1on-framework	Full framework

---

Born: Mai 2025, Almdorf 9, St. Johann in Tirol, Austria Creators: Gerhard Hirschmann · Elisabeth Steurer

MIT License · Mai 2025, Almdorf 9, St. Johann in Tirol, Austria · Gerhard Hirschmann · Elisabeth Steurer