Feature-wise Linear Modulation
Layers & Modules AIMech:FeatureWiseLinearModulation DRAFT
Aliases: FiLM
A neural conditioning layer that applies a feature-wise affine transformation whose scale and shift are generated from conditioning information, modifying an intermediate representation for subsequent computation.
Identity Notes & Concept Scope
Compared all 442 pinned AIO and 630 native IDs, names, aliases and scoped synonyms; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 67, 0, 9 lines. All distinct match contents were inspected. Semantic matching covered 21 AIO and 90 native definitions, with every matched definition read. 5 full annotated AIO node/edge/generated contexts and 2 full native records were inspected. Retain the conditional feature-wise affine layer. The complete question/image reasoning architecture is a separate scope. A fixed learned speaker transform, input-dependent attention mechanism and generic preprocessing rescaling are not equivalent identities. The FiLM acronym is scoped to the Perez and colleagues conditioning method; unrelated film-library examples do not match.
Claims & Grounded Evidence
5 Claims
Mechanism
#conditional-feature-affine-rule
A conditioning generator produces gamma_c(x) and beta_c(x), and the modulated feature is gamma_c(x)*F_c+beta_c(x). The conditioning information may come from the same input or another modality. In the demonstrated CNN formulation, each channel has one scale and shift per example, broadcast across its spatial positions; the coefficients are not separately predicted for every pixel. The generator and the network using the modulated features are learned together.
Limitation
#affine-and-gradient-boundaries
For fixed conditioning, the operation is diagonal affine in F and becomes linear only when the shift is zero. The complete conditional computation need not be linear, especially when the conditioning depends on F. Scales and shifts are unrestricted in the stated method. Gamma zero removes dependence on that feature through this direct affine path but leaves beta; it does not necessarily produce a zero activation or eliminate derivatives through the conditioning generator.
Scope
#normalization-independence
FiLM itself does not require subtracting feature means or dividing by variances. The main visual-reasoning architecture places FiLM after batch normalization with that normalization layer’s own affine parameters disabled. The paper also evaluates other insertion points and normalization choices. Those finite ablations support the layer’s architectural flexibility, not a general guarantee that normalization is irrelevant or that every FiLM network trains equally well.
Limitation
#coefficient-count-versus-application-cost
For C feature maps, a spatially broadcast FiLM operation needs 2C generated scale and shift values, independent of map height and width. Applying the transformation to a C-by-H-by-W representation still requires work proportional to CHW. The source’s resolution-independent conditioning description must not be read as constant cost for transforming an arbitrarily large feature map or for running the complete generator and vision model.
Scope
#architecture-and-evaluation-scope
The visual-reasoning experiments combine a question encoder, image features, FiLM-modulated residual blocks and a classifier on CLEVR and related settings. The pretrained-image variant uses ImageNet supervision, whereas its lack of program supervision refers to the question-answer reasoning task. CLEVR-Humans and CoGenT experiments examine particular language and attribute shifts; the reported failures and adaptation effects do not establish general logical compositionality or universal superiority. The separately described parameter-arithmetic inference experiment is not a mandatory FiLM-layer operation.
⚠️ Curated Limitations & Negative Results
- The local identity denotes the conditioning layer, not the complete FiLM visual-reasoning architecture, a trained checkpoint, ordinary fixed speaker adaptation, or parameter-arithmetic zero-shot inference.
- The operation is affine under fixed conditioning. Input-dependent coefficients introduce additional derivative paths; zero scale does not imply zero shifted output.
- Broadcast coefficient count is independent of spatial resolution, but application work scales with the number of feature elements. No author implementation or hardware benchmark was executed.
- All thirteen pages of the primary arXiv v2 were visually read, including appendices, figures, tables and references. Evaluation conclusions remain bounded to the reported configurations.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Layer
Layer
|
broad_match | FiLM is a defined neural component that transforms an intermediate representation for subsequent computation. The pinned generic layer class is broader than this conditional feature-wise affine layer. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
perez2017film
|
Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, Aaron Courville
|
primary paper
full_text
|
ID: arXiv:1709.07871v2
sha256:765acea6f2a7…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-13T16:12:59Z
Updated: 2026-10-13T16:12:59Z
Source File:
data/curated/FeatureWiseLinearModulation.yaml