Global Workspace Competition + Broadcast + Working Memory with Ignition Dynamics

Fuse multi-modal encoder outputs by tokenizing and routing them into a constrained global workspace, applying ignition-driven broadcasting to coordinate downstream modules. This enables stable cross-modal integration across vision, text, audio, and sensor data.

Ignition dynamics control broadcasting by gating when the global workspace commits outputs and persists content in working memory. This mechanism uses iterative rounds with deterministic top-K gating to decide when to trigger cross-module broadcasts.

Implement deterministic winner selection using explicit four-term scoring across content, salience, novelty, and task metrics. The architecture applies top-K gating to select winning tokens from a fixed-modality pool, producing slot-based workspace outputs of shape (B, K, D).

Yes, the architecture supports modular encoders for vision, text, audio, and sensor data. It provides broadcast adapters to route slot-based workspace outputs to temporal, symbolic, and decision modules, allowing integration with custom encoder implementations.

Maintain persistent context by applying ignition dynamics that gate memory updates within the working memory backend. The workspace persists content across timesteps, ensuring robust multi-modal reasoning and stable context coordination for downstream modules.

Fixed-capacity workspaces constrain the number of tokens selected via deterministic top-K gating, meaning overflow tokens are discarded. This requires careful tuning of the four-term scoring weights to ensure critical salience and novelty signals are not lost during slot allocation.