PaperlayerView exact revision on arXiv

Source-linked AI summary

From S3Q Theory to Implementation: Towards an Architecture for Machine Qualia

Tetiana Grinberg, Katrina Schleisman, Patryk Laurent, Bogdan Udrea, Minda Myers, Brian Aufderheide, Luis El Srouji, Doyle Groves, Kevin Schmidt

arXiv:2609.30743v1cs.AI

TL;DR

Machine consciousness theories specify conditions for qualia, but no published system composes the S3Q conditions into one architecture with defined interfaces. This paper proposes a five-layer composition using published components, and sets out falsifiable predictions for the composed system. Its scope remains bounded by open questions including source-attribution convergence and behavior under high occlusion.

  • Problem

    S3Q specifies situatedness, simulation, and structural coherence as conditions for machine consciousness, but no published system composes all three with defined interfaces and falsifiable predictions.

    Source: A. The Problem · p. 1

  • Method

    The paper composes published components into a five-layer architecture using shared slot-vector representations and specifies how the layers connect.

    Source: B. The Claim · p. 1

  • Results

    The architecture predicts that the agent’s body emerges as the slot with highest motor-command mutual information and that model violations elicit distinct responses depending on valence; both predictions are falsifiable.

    Source: B. The Claim · p. 1

  • Takeaways & Limitations

    The composition provides falsification criteria intended to make either success or failure informative, with possible structural-perturbation applications in computational psychiatry and embodied AI.

    Source: §VIII · p. 6

    Convergence of source-attribution bootstrap, cross-slot Hopfield completion under high occlusion, and several other questions remain open.

    Source: B. Scope Limitations and Open Questions · p. 6

Abstract

from arXiv · show

A key challenge in machine consciousness research is translating theoretical models into computational-level implementations. In this paper, we address this challenge by proposing a five-layer implementation architecture for the S3Q (Simulated, Situated, Structurally Coherent) theory of consciousness. Rather than introducing novel formalisms, the architecture composes published computational primitives into a single pipeline. S3Q identifies three jointly necessary conditions for qualia: (1) grounded sensorimotor situatedness, (2) internal simulation via a world model, and (3) structural coherence between predictions and observations. No existing computational system implements all three simultaneously. We map each S3Q tenet to specific, compatible computational machinery and specify how these components interface within a single representation pipeline that operates on continuous, differentiable, per-object slot vectors, along with a developmental bootstrap sequence and falsifiable predictions for the composed system that no subset of the architecture produces in isolation. The model suggests that a basic sense of "self" develops by linking actions to their outcomes, and that behavior falls into three patterns (hesitation, curiosity, or avoidance) depending on how unexpected an outcome is and whether it is experienced as positive or negative. Each prediction is individually falsifiable, providing the field with a testable framework to advance our understanding of machine consciousness.

I. INTRODUCTION

The S3Q framework specifies three jointly necessary conditions for machine consciousness, but existing computational components address them separately. This paper proposes a five-layer composition with defined interfaces and falsifiable predictions.

A. The Problem

  • S3Q identifies situatedness, simulation, and structural coherence as jointly necessary conditions for machine consciousness, while the theory does not specify how to implement them.Published components address individual conditions, but no published system composes all three with defined interfaces, a developmental bootstrap sequence, and falsifiable predictions.Source: A. The Problem · p. 1

B. The Claim

  • The proposed five-layer architecture composes published mathematical components around shared slot-vector representations rather than introducing new components.It specifies functional interfaces and developmental ordering, with predictions that depend on the conjunction of layers.Source: B. The Claim · p. 1

C. The S3Q Framework

  • S3Q characterizes situated representations as grounded in a dynamic world, simulated representations as internally generated sensorimotor models, and structural coherence as capturing enough reality for environmental grounding.Structural coherence balances overlearned rigid habits with flexible cognition.Source: C. The S3Q Framework · p. 1

D. Definition of Qualia

The paper defines qualia in relation to an agent’s goals and internal world model, then maps S3Q’s three tenets onto five functional layers. The layers share continuous per-object slot vectors and follow a developmental bootstrap sequence.

D. Definition of Qualia

  • A quale is a processed feature made relevant by the agent’s goals and usable directly by its internal world model.The same surface reflection may matter to a photographer but not to someone simply admiring a landscape.Source: D. Definition of Qualia · pp. 1–2

II. ARCHITECTURAL OVERVIEW: FIVE FUNCTIONAL LAYERS

  • The five-layer architecture maps situatedness to object-centric perception and self/world factorization, simulation to pattern completion and an action loop, and coherence to regulation.Each slot is a fixed-length vector representing one object’s features, kept separate from other objects throughout the pipeline.Source: §II · p. 2
  • All five layers use continuous per-object slot vectors, with Layer 2 training the shared forward model T for use by Layers 3–5.Layers 1–3 use unsupervised or self-supervised learning; Layer 5 policy selection additionally depends on valence.Source: §II · p. 2
  • Developmental bootstrap first stabilizes object-slot trajectories, then identifies action-contingent slots; Layer 2 trains after MI stabilizes, Layer 4 evaluates with T, and imagination starts after reliable motor modulation.Layer 5 is passive until T and the suppression gate produce stable predictions.Source: §II · p. 2

III. TENET 1: SITUATED (LAYERS 1 AND 2)

Situatedness is implemented by parsing sensory input into persistent object slots and characterizing each slot’s dynamics relative to the agent’s actions. Self/world factorization labels action-linked slots as self, subject to stated test conditions and caveats.

A. Layer 1: Object-Centric Perception and Covariance Detection

  • Object-centric encoding decomposes sensory input into persistent slots, while per-slot Slow Feature Analysis and action–feature mutual information identify stable, controllable candidates for self.Slot Attention and SAVi are examples; the architecture requires recurrent slot binding, not a specific encoder.Source: A. Layer 1: Object-Centric Perception… · p. 2

B. Layer 2: Self/World Factorization

  • Slots whose action mutual information exceeds a threshold relative to the slot distribution are labeled self; the remainder are labeled world.The predicted highest-MI self slot is an empirical claim, not a guaranteed architectural consequence.Source: B. Layer 2: Self/World Factorization · pp. 2–3
  • Figure 1 distinguishes the tenets by color, the action/data loop with solid arrows, forward-model reuse with dashed arrows, and Layer 5 coherence regulation with dotted arrows.The diagram presents the five-layer architecture and its data-flow conventions.Source: Figure 1 · p. 3
  • The self-slot test is discriminating only when another action-contingent object is present; otherwise the prediction holds by construction.If the body is largely outside the sensory field, self-motion may correlate with all slots rather than a single high-MI slot.Source: B. Layer 2: Self/World Factorization · p. 3

C. The Forward Model (T)

The architecture uses a shared per-slot forward model T to predict transitions, estimate uncertainty, and support coherence monitoring. T is self-supervised and reused across layers, while prediction error is aggregated to inform behavioral response.

C. The Forward Model (T)

  • T is a per-slot transition function conditioned on the current slot state, action, and self/world source label.The action is a motor command for a self slot and an inferred external cause for a world slot.Source: C. The Forward Model (T) · p. 3
  • T is self-supervised on the agent’s motor-stream transitions, and the same functional form models self-caused and externally caused transitions.The source label modulates its output through a learned embedding; Layers 3–5 reuse this single model without retraining.Source: C. The Forward Model (T) · p. 3
  • T predicts next states and per-state variance, trained jointly by negative log-likelihood; its required form preserves slot identity and produces state deltas.C-SWM is suggested for short-horizon predictions and a SlotFormer-style autoregressive Transformer for longer imagination rollouts; either is trained once and reused.Source: C. The Forward Model (T) · p. 4
  • Precision-weighted prediction error is aggregated as the error of the worst-offending slot and used to select a behavioral response.The passage also describes cross-slot Hopfield completion: sparse imagined representations change more substantially than full observations and can seed completion of missing information.Source: C. The Forward Model (T) · p. 4

A. Layer 3: Pattern Completion (Imagination from Sparse Seeds)

Layers 3 and 4 use learned structure to complete sparse representations and simulate unobserved states, while suppressing motor output when internal simulation should not trigger action.

A. Layer 3: Pattern Completion (Imagination from Sparse Seeds)

  • Pattern completion uses past experience to fill gaps in novel but similar representations, supporting fast recognition, dreaming, and forward-model imagination.S3Q treats experience as simulation: what is experienced is the world model, not the actual world.Source: A. Layer 3: Pattern Completion… · p. 4
  • A learned transition kernel Φ applies to a noun slot vector to predict its next state, enabling simulation of novel object-action combinations.The Noun-Verb Composition Test predicts simulation of combinations never seen before, such as a Blue Ball Bouncing.Source: A. Layer 3: Pattern Completion… · p. 4

B. Layer 4: Action Loop with Continuous Suppression

  • Layer 4 suppresses motor responses during imagination and dreaming, using surprise- and context-driven signals to modulate a continuous motor gate.Leaky suppression can produce subthreshold motor output during sleep, hypothesized to seed pattern completion with proprioceptive signals.Source: B. Layer 4: Action Loop with… · pp. 4–5
  • For each candidate action, the agent projects its outcome through T, scores it against current objectives, and selects an action by softmax.Successor Features are offered as an optional alternative when transfer across multiple goals is required.Source: B. Layer 4: Action Loop with… · p. 5

V. TENET 3: STRUCTURALLY COHERENT (LAYER 5)

Layer 5 compares predictions with observations using confidence-weighted error, then relates surprise, valence, and goals to behavioral responses and representation learning.

V. TENET 3: STRUCTURALLY COHERENT (LAYER 5)

  • Structural coherence requires a world model close enough to reality for adaptive action yet abstract enough to generalize under uncertainty.The architecture requires coherence quantities to be computed but does not prescribe a dedicated monitoring subsystem or where computation occurs.Source: §V · p. 5
  • The response rule combines precision-weighted prediction error as surprise with predicted valence, since surprise alone does not distinguish behavioral modes.The predicted valence head is trained on reward; negative values are aversive and positive values appetitive.Source: §V · p. 5
  • Moderate surprise prompts hesitation, large surprise with neutral or appetitive valence prompts curiosity, and large surprise with aversive valence prompts avoidance.Curiosity decreases suppression and explores near the surprising stimulus; avoidance increases distance from it.Source: §V · p. 5
  • Goal-relevant slots receive higher precision, which amplifies their errors, accelerates their learning, and shifts the balance between pattern completion and observation.The framework predicts that agents with identical sensory exposure but different goals develop different decompositions.Source: §V · p. 5
  • Linear probing tests whether object identity remains decodable throughout dynamic simulation.Source: §V · p. 5

A. What This Framework Claims That Alternatives Do Not

The framework attributes three predictions to the conjunction of its components: goal-shaped representations, an active-exploration advantage, and valence-sensitive responses to model errors.

A. What This Framework Claims That Alternatives Do Not

  • The component conjunction predicts goal-shaped feature decompositions, self-slot emergence through active exploration, and distinct responses to high prediction error based on valence.The paper contrasts these predictions with Slot Attention alone and a Dreamer-style world model alone.Source: A. What This Framework Claims That… · p. 5

B. Falsification Criteria

The framework specifies falsifiable outcomes and environmental conditions for testing its claims.

B. Falsification Criteria

  • The framework is falsified if ablations remove situatedness effects, passive observation identifies self slots equivalently, responses ignore valence, composition smears identity, or imagination lacks a compositional advantage.Source: B. Falsification Criteria · p. 6

C. Environment Requirements

  • A discriminating test environment needs independently varying objects and dynamics, action-dependent observability, continuous physics, and flexible goals.ThreeDWorld and MineDojo are given as environments meeting these criteria.Source: C. Environment Requirements · p. 6

A. Relation to Existing Work

The architecture is distinguished from both disentangled representation learning and the full Free Energy Principle apparatus, while its open questions bound the current proposal. The conclusion presents composition and falsification criteria as the contribution, with potential applications in computational psychiatry and embodied AI.

A. Relation to Existing Work

  • Unlike disentangled representation learning, the architecture adds goal-sensitive selection, active grounding, and valence-dependent modes; it uses selected Active Inference components, without adopting its full apparatus.The mode-selection rule simplifies Expected Free Energy, while suppression and mode selection remain engineered components.Source: A. Relation to Existing Work · p. 6

B. Scope Limitations and Open Questions

  • Open questions include theory of mind, multi-agent interaction, action-repertoire learning, bootstrap convergence, high-occlusion pattern completion, and contextual calibration.These topics remain for future work.Source: B. Scope Limitations and Open Questions · p. 6

VIII. CONCLUSION

  • The contribution is the composition of published components through shared slot vectors, with falsifiable predictions that no subset produces in isolation.The modularity also suggests applications beyond consciousness research.Source: §VIII · p. 6
  • The modular architecture suggests computational-psychiatry perturbations and embodied-AI safety constraints and failure diagnostics.Proposed perturbations include disabling the suppression gate, distorting Layer 5 precision weights, or disrupting Layer 2 self/world factorization,.Source: §VIII · p. 6

Related papers

Loading 2609.30743v1…