Squeeze-and-Excitation block
Layers & Modules AIMech:SqueezeAndExcitationBlock DRAFT
Aliases: SE block
A neural-network feature recalibration component that globally averages each channel, transforms the resulting descriptor through a learned bottleneck, and applies sigmoid-derived channel weights to the input feature maps.
Identity Notes & Concept Scope
This is the SE component from the 2018 paper, not a complete SENet architecture or its trained checkpoints. The component receives the feature tensor produced by a preceding transformation. Its channel weights vary with the input but are shared over spatial positions within each channel. SE block is the scoped alias; the paper title denotes networks assembled using these units. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 279 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in worktree and main, returning 6 and 6 lines respectively. Every text hit was a prior candidate lead or a closely related variant lead. Seven exact AIO CLASS nodes with scoped synonyms/comments and outgoing edges and three native neighboring mechanisms were inspected. Absence is bounded to this snapshot and semantic neighborhood.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
For each channel of a spatial feature tensor, average all spatial values to obtain a channel descriptor. Pass that descriptor through a fully connected dimensionality reduction, ReLU, fully connected expansion to the original channel count and elementwise sigmoid. Broadcast each resulting weight across its corresponding spatial feature map and multiply elementwise. The reduction ratio controls the bottleneck capacity; the paper uses 16 in its main experiments rather than requiring that value for the concept. In its residual-network integration, this recalibration operates on the non-identity transformation before addition of the skip path. The module has learned parameters and remains part of the network during inference.
⚠️ Curated Limitations & Negative Results
- The sigmoid gates are independent weights, not a softmax distribution or guaranteed class probabilities. For finite real logits they lie strictly between zero and one, so the immediate scaling cannot increase absolute input-feature magnitudes; emphasis is relative to other channels and the surrounding network. Floating-point saturation can reach numerical endpoints.
- Global averaging discards spatial arrangement in the descriptor. Channel gating therefore does not itself supply a location-specific attention map, although the spatial feature tensor is retained in the output.
- Reduction ratio, placement and surrounding architecture affect cost and behavior. Additional learned weights and operations do not imply negligible cost on every device. The source’s interpretation of particular channel activations is not a causal explanation or a universal accuracy claim. No author implementation or trained model was executed.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Layer
Layer
|
broad_match | The native concept is a particular learned input-to-output feature-map component inside a neural network. The pinned generic Layer class is broader; the pinned AttentionLayer specifically denotes dot-product attention and is not used as a generic attention parent. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Jie Hu, Li Shen, Gang Sun
|
primary paper
full_text
|
sha256:2acc7faca912…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T12:43:35Z
Updated: 2026-10-11T12:43:35Z
Source File:
data/curated/SqueezeAndExcitationBlock.yaml