Home / Curated / Architectures / Attention-Based Configurable Convolutional Neural Network

Attention-Based Configurable Convolutional Neural Network

Architectures AIMech:AttentionBasedConfigurableConvolutionalNeuralNetwork DRAFT

Aliases: ABC-CNN

A visual-question-answering architecture that derives a convolution kernel from an encoded question to compute normalized spatial attention, then fuses weighted and unweighted image features with the question for answer-word classification.

Identity Notes & Concept Scope Compared all 442 pinned AIO and 610 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 23, 0, 0 lines. All distinct match contents were inspected. Semantic matching covered 166 AIO and 76 native definitions, with every matched definition read. 7 full annotated AIO node/edge/generated contexts and 3 full native records were inspected. Visual ABC-CNN derives configurable kernels from question semantics and predicts answer words from weighted/unweighted image features. Sentence-pair ABCNN instead compares sentence representations by distance affinities; CBAM is a feature-refinement component and DMN+ uses recurrent episodic fact aggregation. Shared attention language does not establish identity.

Claims & Grounded Evidence

6 Claims

Mechanism #question-configured-spatial-attention

An LSTM encodes the question, and the model averages its hidden states over time to obtain the question vector. A learned affine transform followed by sigmoid converts that vector into a configurable convolution kernel with channels matching the spatial image features. Convolving this kernel over the image feature map and applying one softmax over spatial positions produces a question-conditioned attention map. The convolution is padded to preserve the spatial extent; illustrative kernel drawings do not establish an uninspected implementation kernel size.

supports Source: ABC-CNN: An Attention Based Convolutional Neural Network for Visual Question Answering · Locator: Sections 3.1–3.3, PDF pp. 3–5, Equations 1–3 and Figures 2–3

Mechanism #weighted-and-unweighted-feature-fusion

The same spatial attention map multiplies each image-feature channel. A one-by-one convolution reduces the weighted feature channels, and a nonlinear projection fuses the reduced weighted map, the original unweighted image map and the question vector. The reported answer task is classification over a separate answer-word dictionary, selecting the most probable word. The original-image branch allows predictions to use information outside the weighted path; the architecture does not require every answer to depend on the attention map.

supports Source: ABC-CNN: An Attention Based Convolutional Neural Network for Visual Question Answering · Locator: Section 3.4, PDF p. 5, Equations 4–6; Section 4.5, p. 8, original-feature bypass discussion

Scope #notation-and-training-boundaries

The answer-generation prose specifies a softmax classifier, while Equation 6 reuses g after g was defined as scaled tanh in Equation 5. This record follows the explicit classifier description and does not treat the reused symbol as a verified probability formula. Although the architecture permits end-to-end training, the reported experiments keep the image feature extractor fixed for speed. The question encoder, attention and answer components are trained with answer supervision; the absence of attention-region labels is not an unsupervised-learning claim.

supports Source: ABC-CNN: An Attention Based Convolutional Neural Network for Visual Question Answering · Locator: Sections 3–3.5 and introductory training description, PDF pp. 2–5, Equations 5–6 and Training and Testing

Scope #feature-and-answer-vocabulary-variants

The experiments distinguish coarse VGG-plus-color features, denser fully convolutional segmentation features, and a combined feature configuration. These are separate evaluated settings of ABC-CNN rather than interchangeable synonyms. The VQA comparison also distinguishes a dictionary of one thousand frequent answers from a full answer dictionary, with different reported accuracies. Spatial resolution, pretrained features, answer vocabulary and fixed-encoder training therefore belong to the empirical protocol, not universal architecture guarantees.

supports Source: ABC-CNN: An Attention Based Convolutional Neural Network for Visual Question Answering · Locator: Section 4.1, PDF pp. 5–6; Section 4.5, pp. 7–8, Tables 2 and 5

Limitation #filtered-evaluation-and-table-arithmetic

Toronto COCO-QA, DAQUAR and VQA results use different datasets and answer restrictions; the reported method is evaluated on single-word answers. Aggregate gains do not extend to every question category or metric: the listed image-plus-bag-of-words comparator scores higher on COCO-QA color questions, and the DAQUAR-reduced ensemble scores higher on WUPS at threshold 0.9. The prose attention-ablation differences also disagree with Table 2: subtracting NO-ATT from ATT yields 3.89, 3.31 and 1.11 percentage points for accuracy, WUPS 0.9 and WUPS 0.0. These table-derived differences are not the prose values 1.34, 0.85 and 0.35.

supports Source: ABC-CNN: An Attention Based Convolutional Neural Network for Visual Question Answering · Locator: Sections 4.2–4.5, PDF pp. 6–8, Tables 2–5; constructed Table 2 arithmetic

Limitation #attention-interpretation-and-identity

The normalized map is a learned spatial weighting, not a demonstrated probability that a region is the correct object or a certified causal explanation. Selected visualizations illustrate question-dependent responses, and the paper explicitly allows correct answers when the attention map does not focus on the queried object because original image features are also supplied. Extending the classifier to generate full sentences is proposed rather than evaluated here.

supports Source: ABC-CNN: An Attention Based Convolutional Neural Network for Visual Question Answering · Locator: Introduction and Section 3, PDF pp. 2–3; Section 4.5 and Figure 5, pp. 8–9; Conclusion, p. 8

⚠️ Curated Limitations & Negative Results

This is Chen and colleagues’ visual ABC-CNN, not Yin and colleagues’ sentence-pair ABCNN or every attention-based CNN.
Printed classifier notation and ablation arithmetic contain inconsistencies; interpretation and table-derived calculations remain explicit, without assuming unseen code.
Reported image encoders are fixed, answer outputs are single words, and dataset/feature/vocabulary settings differ. No unrestricted generation, universal superiority or causal attention faithfulness is established.
All ten pages of arXiv1511.05960v2 were visually read. No author code, training, benchmark replication or other source version inspected. DRAFT remains.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	broad_match	Learned convolutional, recurrent and classification components form a particular multimodal neural architecture; ANN is broader.
`AIO:ConvolutionalLayer` Convolutional Layer	related_match	Image feature extraction, question-configured convolution and channel reduction are components of the whole architecture.
`AIO:LongShortTermMemory` Long Short Term Memory	related_match	An LSTM encodes the question and its hidden states are averaged; that component does not define the entire visual-question-answering architecture.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`chen2016`	ABC-CNN: An Attention Based Convolutional Neural Network for Visual Question Answering Kan Chen, Jiang Wang, Liang-Chieh Chen, Haoyuan Gao, Wei Xu, Ram Nevatia	primary paper full_text	ID: arXiv:1511.05960v2 sha256:43d0fc2f07dd…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-13T11:31:08Z

Updated: 2026-10-13T11:31:08Z

Source File: data/curated/AttentionBasedConfigurableConvolutionalNeuralNetwork.yaml