Show & Tell — Phase 28.1 AttentionFilter: Selective Attention Pipeline

[web3guru888]

AttentionFilter — Architecture & Data Flow

Selective Attention Pipeline

Sensory Streams ──┐
                   ├──▶ Candidate Pool ──▶ SaliencyAttentionFilter ──▶ Filtered Top-K
WorldModel 13.1 ──┘         │                      │                       │
                            │              ┌───────┴───────┐               │
                            │              │  Score Calc:   │               ▼
                            │              │  w_s·saliency  │        WorkingMemoryGate
                            │              │ + w_r·relevance│           (28.2)
                            │              │ − λ·Δt (decay) │
                            │              │ − IoR_penalty   │
                            │              └────────────────┘
                            ▼
                     Inhibition Map
                    (target_id → expiry)

Saliency Computation

The combined score balances bottom-up saliency (stimulus-driven) with top-down relevance (goal-driven):

score(target, t) = w_s · target.saliency 
                 + w_r · target.relevance 
                 − target.decay_rate · (t − target.timestamp)
                 − ior_penalty · 𝟙[target_id ∈ inhibited_set]

Default weights: w_s = 0.4, w_r = 0.6 (biased toward goal-relevance per executive control literature).

Attention Modes

Mode	Behavior	Use Case
FOCUSED	Single target, maximum depth	Deep reasoning on one problem
DIVIDED	Multiple targets, reduced depth	Parallel monitoring
SUSTAINED	Stable targets, vigilance	Long-running surveillance
EXECUTIVE	Goal-directed, top-down bias	Task execution
VIGILANCE	Low-arousal background scan	Anomaly detection

Inhibition of Return (IoR)

Prevents the system from repeatedly attending to the same targets. After focusing on target X:

1. X enters _inhibited dict with expiry = now + ior_duration_s
2. Subsequent filter() calls apply ior_penalty to X's score
3. After expiry, X is removed from inhibited set
4. This forces exploration of novel stimuli (Posner & Cohen 1984)

Integration with SensorFusion 25.1

SensorFusion produces raw multi-modal feature vectors. AttentionFilter wraps these as AttentionTargets:

• saliency = feature magnitude / max(magnitudes) — normalized bottom-up signal
• relevance = cosine similarity with current goal embedding — top-down signal
• source = modality identifier (visual, auditory, proprioceptive)