Home / Curated / Architectures / Dynamic Memory Network Plus

Dynamic Memory Network Plus

Architectures AIMech:DynamicMemoryNetworkPlus DRAFT

Aliases: DMN+

A recurrent question-answering architecture that fuses text or image facts, aggregates them with question- and memory-conditioned attention gates in a GRU, and updates episodic memory through pass-specific rectified affine transformations.

Identity Notes & Concept Scope Compared all 442 pinned AIO and 605 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 39, 0, 0 lines. All distinct match contents were inspected. Semantic matching covered 65 AIO and 89 native definitions, with every matched definition read. 8 full annotated AIO node/edge/generated contexts and 3 full native records were inspected. DMN+ is the named successor with bidirectional fact fusion, scalar attention replacing the GRU update gate, and untied ReLU memory updates. It is not an alias of the original DMN, the intermediate DMN2/DMN3 ablations, KV-MemNN or CommNet. Existing incidental DMN+ benchmark mentions are not a maintained identity.

Claims & Grounded Evidence

6 Claims

Mechanism #fused-text-and-image-facts

DMN+ replaces the original word-stream input with sentence representations formed by position-weighted word embeddings followed by bidirectional GRU input fusion; forward and backward states are added. Its visual variant projects VGG-19 regional features into the question space and applies recurrent fusion along a snake traversal of the image patches. These are contextual fact encoders, not independently addressed raw tokens or exact object segmentation.

supports Source: Dynamic Memory Networks for Visual and Textual Question Answering · Locator: Sections 3.1–3.2, PDF pp. 2–4, Figures 2–3

Mechanism #normalized-attention-gru

At each memory pass, an MLP scores elementwise products and absolute differences between each fact, the question and previous memory. Softmax across facts produces scalar gates. The attention-based GRU replaces its usual vector update gate with this scalar, mixing the candidate state with the previous state and returning the final scan state as context. This differs from the original DMN gate multiplying a complete GRU update. A softmax-weighted sum is an evaluated alternative, not the final recurrent aggregation; any permutation-invariance statement for that sum holds only for fixed fact representations.

supports Source: Dynamic Memory Networks for Visual and Textual Question Answering · Locator: Section 3.3, PDF pp. 4–5, Equations 8–11; Section 6.1, pp. 6–7

Scope #untied-memory-and-printed-notation

The final DMN+ memory update applies ReLU to an affine transformation of the concatenated previous memory, context and question, with different parameters across passes. The printed square matrix dimension next to Equation 13 conflicts with this three-vector concatenation; a dimensionally consistent transformation from three n_H vectors requires an n_H by 3*n_H matrix. Equation 10 also prints an inconsistent denominator bound rather than the declared fact count. These are explicit dimensional/notation interpretations of the paper, not a claim that the displayed typos define a working implementation.

supports Source: Dynamic Memory Networks for Visual and Textual Question Answering · Locator: Section 3.3, PDF pp. 4–5, Equations 10 and 13; Section 6.1, p. 6

Scope #answer-supervised-finite-protocol

The reported bAbI-10k experiments train from answers without supporting-fact labels. They use selected finite configurations, including three memory passes, truncated recent-sentence inputs, development-based early stopping and selection among multiple random restarts. Absence of gate labels does not mean absence of answer supervision. DMN2 and DMN3 are intermediate ablations; the complete DMN+ combines input fusion, attention-based recurrent aggregation and untied ReLU memory updates.

supports Source: Dynamic Memory Networks for Visual and Textual Question Answering · Locator: Sections 5 and 6.1–6.2, PDF pp. 6–7, Table 1

Limitation #task-and-metric-boundaries

Text QA, DAQUAR model analysis and VQA are separate protocols. bAbI reports per-task error and mean error; DAQUAR analysis restricts evaluation to single-word answers, and VQA scores agreement with human answers. DMN+ improves selected aggregate results but not every task or answer type: induction remains difficult, and its VQA number score of 36.8 is above SAN (36.6) but below D-NMN (37.4) in Table 3. Ablation and cross-paper comparisons do not establish universal superiority or isolate every architectural change as a causal effect.

supports Source: Dynamic Memory Networks for Visual and Textual Question Answering · Locator: Section 5, PDF p. 6; Sections 6.1–6.3, pp. 6–8, Tables 1–3

Limitation #order-and-visualization-limits

Sequential input fusion and the attention-based GRU can depend on fact or patch order. The snake scan does not establish arbitrary patch-permutation or geometric invariance. Selected visual attention heatmaps show model gates rather than a proof of faithful causal object grounding; the source notes that patch-based representations make counting objects crossing patch boundaries difficult. The mechanism is a specific DMN successor, not evidence that all memory networks share these properties.

supports Source: Dynamic Memory Networks for Visual and Textual Question Answering · Locator: Sections 3.1–3.3, PDF pp. 2–5; Section 6.3, p. 8; Figure 6, p. 9

⚠️ Curated Limitations & Negative Results

Original DMN, intermediate DMN2/DMN3 ablations and complete DMN+ are distinguished.
Source equation bounds and matrix dimensions contain inconsistencies; the record labels its dimensional interpretation and does not claim an inspected implementation.
Benchmark protocols, finite memory, selected restarts and selected visualizations bound the evidence. No author code, training, benchmark replication or hardware measurement was performed.
All ten pages of arXiv1603.01417v1 were visually read. Other versions were not inspected. DRAFT remains.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:RecurrentNeuralNetwork` Recurrent Neural Network	broad_match	Bidirectional input fusion and the attention-based recurrent fact scan define a specific recurrent architecture; the target is broader.
`AIO:GatedRecurrentUnit` Gated Recurrent Unit	related_match	GRUs are components of input fusion and the modified attention scan; the complete DMN+ is not equivalent to one GRU and its scalar attention gate replaces the usual update gate.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`xiong2016`	Dynamic Memory Networks for Visual and Textual Question Answering Caiming Xiong, Stephen Merity, Richard Socher	primary paper full_text	ID: arXiv:1603.01417v1 sha256:ce961743dc7b…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-13T10:44:20Z

Updated: 2026-10-13T10:52:59Z

Source File: data/curated/DynamicMemoryNetworkPlus.yaml