Salience Map or Attention Heatmap¶
Interface — instantiates Cascaded Hierarchical Recognition
Shows where attention should move next based on current stage evidence.
Salience Map or Attention Heatmap overlays the input with a heat gradient showing where the current stage's evidence is strongest or most uncertain, telling the observer where to point attention next. Its defining idea is that it is a spotlight, not a shortlist: it answers "where should I look?" by highlighting a region, and it carries no set of candidate classes and walks no hierarchy of levels. That is exactly what separates it from a drilldown interface — the heatmap directs the eye to a location; it does not narrow a preserved list of what the thing might be. Its intensity is not just a location marker but an encoding of how confident, or how uncertain, the model is that the signal it cares about lives there.
Example¶
An automated optical-inspection station on an electronics line photographs each assembled circuit board. Instead of returning a bare pass/fail, it overlays a heatmap on the board image: regions the model finds most anomalous glow hot, ordinary regions stay cool. A technician's eye is pulled straight to a bright patch over one corner — a possibly cold solder joint — rather than scanning the whole densely-packed board component by component. The heat intensity carries the model's confidence: a faint warm smear says "slightly unusual, maybe nothing," a sharp hot spot says "strongly anomalous here." The technician then inspects that spot with a loupe. The heatmap never says what the defect is or which of several defect classes it belongs to; it says only look here first.
How it works¶
- Score each region at the current stage for evidence strength or uncertainty.
- Render as a heat overlay aligned to the input, hottest where the signal is strongest or most ambiguous.
- Route the eye. The hottest region is where attention goes next; the observer inspects there rather than everywhere.
- Update as evidence accrues. As a stage resolves, the map re-weights toward the next most informative region.
Tuning parameters¶
- What the heat encodes — raw evidence strength, model uncertainty, or class-activation; each sends attention to a different kind of region.
- Color scale and cutoff — where warm becomes hot; an aggressive cutoff makes one region dominate, a gentle one spreads attention.
- Spatial resolution — fine-grained versus smoothed heat; finer is precise but noisier and harder to read at a glance.
- Live vs static — whether the map updates as the observer works; live guidance is responsive but can feel jittery.
- Top-k vs continuous — highlight only the few hottest regions, or show a full gradient.
When it helps, and when it misleads¶
Its strength is directing scarce attention to the highest-value region fast, and making a model's focus legible so a human can sanity-check where it is attending.
Its failure mode is superficial-cue lock-in: a salient-but-irrelevant hotspot captures attention and biases everything that follows, and a heatmap can look explanatory without being faithful — saliency maps can highlight spurious correlations rather than the true cause.[n1] The classic misuse is treating the brightest region as the diagnosis rather than as a where-to-look hint, collapsing a routing aid into a decision. The guarding discipline is to treat the map strictly as a routing prompt and to verify the highlighted region with an actual test before acting.
How it implements the components¶
attentional_routing_rule— the heatmap is the routing rule made visible: the hottest region is where attention should move next, given current-stage evidence.confidence_and_uncertainty_state— the heat intensity encodes the model's per-region confidence or uncertainty, surfacing where the signal is strong versus merely ambiguous.
It never holds or narrows a set of live candidates and never walks a level hierarchy (candidate_set_carrier, feature_hierarchy — that's drilldown_recognition_interface); the heatmap points a spotlight, whereas the drilldown carries a shortlist through the levels.
Related¶
- Instantiates: Cascaded Hierarchical Recognition — the interface that routes attention to the region worth inspecting next.
- Consumes: Feature Pyramid or Hierarchical Model — the multi-scale evidence the heat is computed from.
- Sibling mechanisms: Cascade Error Audit · Coarse Screening Checklist · Drilldown Recognition Interface · Expert Review Checkpoint · Feature Pyramid or Hierarchical Model · Multi-Stage Classifier Pipeline · Stage Transition Log · Triage Queue with Escalation Rules
Editorial Notes¶
Form Classification¶
Form family: Interface, Display & Cue
Rationale: Salience Map or Attention Heatmap operates as a user-facing prompt, display, template, or perceptual cue that shapes attention and action at the point of use because it shows where attention should move next based on current stage evidence.
Independent corroboration: The frozen evidence defines Salience Map or Attention Heatmap as 'Shows where attention should move next based on current stage evidence', so its operative form is Interface, Display & Cue.
Nearest alternative: Monitoring, Sensing & Alerting — Salience Map or Attention Heatmap includes features of ongoing observation, sensing, or alerting that detects and surfaces state without itself executing the response, but its defining operation is a user-facing prompt, display, template, or perceptual cue that shapes attention and action at the point of use.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Human-Computer Interaction
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Attention heatmaps are established interface and usability evaluation artifacts.
Related originating lineages:
- Computer Science & Software Engineering — Computer science and software-engineering practice supplies a parallel or contributing lineage for the mechanism's defining operation: shows where attention should move next based on current stage evidence.
- Data Science & Analytics — Data science, analytics, and operational monitoring supplies a parallel or contributing lineage for the mechanism's defining operation: shows where attention should move next based on current stage evidence.
- Psychology — Experimental, clinical, and behavioral psychology supplies a parallel or contributing lineage for the mechanism's defining operation: shows where attention should move next based on current stage evidence.
Review resolution: Both blind reviewers agree that human_computer_interaction is the primary historical origin. Explicit reconciliation of alternate_origin_disagreement, encyclopedia_synthesis_disagreement starts from reviewer_a's mechanism-specific evidence: Attention heatmaps are established interface and usability evaluation artifacts. Reviewer A proposed alternates=data_science, psychology, origin_mode=cross_disciplinary_synthesis, domain_reach=multi_domain, and encyclopedia_synthesis=false; reviewer B proposed alternates=computer_science, data_science, psychology, origin_mode=cross_disciplinary_synthesis, domain_reach=multi_domain, and encyclopedia_synthesis=true. The final record retains every independently supported alternate from either review (data_science, psychology, computer_science) without an arbitrary cap, selects origin_mode=cross_disciplinary_synthesis to represent the combined lineage evidence, and records domain_reach=multi_domain and encyclopedia_synthesis=true. Present-day transfer is recorded as reach and is not treated as proof of historical origin.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] A saliency map highlights the regions of an input a model treats as most influential for its output. Such maps aid attention but are not guaranteed faithful — they can foreground spurious cues — which is why a highlighted region should be confirmed by an actual test rather than trusted as an explanation. ↩