acceptodds
Withdrawn from ICLR 2027

Transmitter: Signal-Transduction-Inspired Reusable Response Pathways Through Decoupled Discrete State Selection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Biological signaling provides a role-level analogy for separating response selection from response expression. Inspired by this principle, we introduce Transmitter, a vision backbone whose token-mixing operator makes three roles explicit: query fragments select a route in a learned receptor book, the route retrieves a paired vector from a signal book, and an input-dependent gain modulates the retrieved response. The assembled response is added to each head input, exposing route and response traces for fixed-checkpoint analysis. During training, the model learns both which response to select and how to express it for classification. The classification loss trains the selected response path, while an auxiliary loss learns the query-to-receptor matching used for discrete selection. At one frozen intermediate checkpoint, replacing the release gain of all eight blocks by its within-image spatial mean lowers ImageNet-1K top-1 accuracy by 21.58 percentage points (pp), while moving whole per-token gain vectors to other tokens lowers it by 15.78–16.18 pp; forcing the original routes leaves drops of 20.46 and 15.32–15.60 pp, so gain placement matters even without route reselection. In exploratory COCO traces, the same-minus-different-class code-affinity gap is 0.0078 larger in the last two blocks than in the first two. These results position route-explicit token mixing as an inspectable decomposition of visual computation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Computer Vision, Vision Backbones, Representation Learning, Dynamic Routing, Conditional Computation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7xHtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces Transmitter, a vision backbone whose token-mixing operator (Transduction) is inspired by synaptic signalling. An input-dependent query hard-selects an entry in a learned receptor book, retrieves a paired response from a separate signal book, and scales it with an input- and usage-dependent gain; the selected routes and gains are logged for later analysis. The model is evaluated with an ImageNet-1K comparison against standard backbones, an ImageNet-100 ablation of the routing operator, and gain interventions and trace diagnostics on a single frozen checkpoint. The authors present route-explicit token mixing as an inspectable decomposition of visual computation.

Critical strengths
  1. Novel operator that decouples hard discrete route selection (receptor book) from a separately parameterized response (signal book) and its modulation, inside trainable token-mixing layers rather than on a frozen encoder — consequential because it exposes countable route/gain objects that can be logged and intervened upon, and the ImageNet-100 ablation (70.77% vs 67.19–69.53% for hard KV, soft KV and strict VQ) suggests the decoupled response path is useful.
  2. Careful, self-critical diagnostics: fixed-route replay, matched-size gain decomposition with image-level bootstrap CIs, random-initialization controls and a reported negative gain-cache result — consequential because it limits overclaiming and gives a reusable template for trace-based analysis.
Critical weaknesses
  1. The compactness claim is not supported. The ImageNet baselines score far below their usual accuracies (e.g., ResNet-50 at 63.3% instead of about 76%). Training recipes also differ across model families, even though the setting is called compute-matched. The CIFAR-100 result (88.4%, about equal to CIFAR-10 and far above all baselines) is implausible and suggests a protocol issue. — Consequential because the accuracy evidence cannot be trusted without properly trained baselines.

2: The inspectability claims rest on a single intermediate checkpoint. No seeds or other checkpoints are tested. Performance dropping after the gains are perturbed mostly shows that any network is sensitive to changes in its own activations, and no baseline gets the same test. The class-related structure found in the routes and gains is very small. — Consequential because the main claim, that routes and gains carry meaningful structure, is not established.

Questions for authors

Clarification: Q1. Can the checkpoint analyses be repeated over several seeds and over the final models in Table 1?

Clarification: Q2. Why are the baselines far below their standard ImageNet accuracies? Can you report them trained with standard recipes, or all models under one matched recipe?

Clarification: Q3. How do you explain the CIFAR-100 result being about equal to CIFAR-10? Was any pretraining (e.g., BYOL) or a non-standard split used?

Reviewer_6rNWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Transmitter selects learned response vectors through nearest-key lookup and modulates them with input-dependent gains and route-usage statistics. ImageNet classification, operator ablations, and frozen-checkpoint interventions evaluate this decomposition within a vision backbone.

Critical strengths

The explicit route/gain decomposition enables controlled interventions. Fixed-route replay (Table 3) shows that gain placement affects predictions without route reselection, while Table 9 distinguishes within-token allocation from common scaling.

Critical weaknesses
  1. The evaluation does not establish a competitive accuracy–compute trade-off. It relies mainly on traditional classification benchmarks without strong contemporary lightweight baselines. ResNet-18/50 achieve only 59.579%/63.319%, versus 69.758%/76.13% in the basic Torchvision references (ResNet-18, ResNet-50). These are different training recipes, but the large gaps need explanation. The claimed “compute-matched” protocol lacks per-model training and tuning budgets, weakening the compactness comparison.

  2. The contribution of the added mechanisms remains unclear. Discrete key/value retrieval is established prior work. Table 2 jointly changes gain, RBF masking, and response decoupling, so it cannot isolate their benefits. The relationship between the all_off registry models and the full operator is also unclear, and the auxiliary loss and RBF mask are incompletely specified. The single-checkpoint interventions establish gain sensitivity, but do not demonstrate an advantage over simpler matched operators.

Questions for authors
  1. What exactly is matched by “compute-matched,” and why are the ResNet baselines substantially below standard reference results? Does the claimed trade-off hold against a competitive lightweight backbone under the stated budget?

  2. What does all_off disable, and which configurations support each result? Please give the complete auxiliary loss and RBF-mask definition, and a control isolating gain/usage modulation while keeping the other components fixed.

Reviewer_b1SJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an alternative architecture for vision backbone inspired by synaptic transduction. At each token, the features select the nearest entry in a learned receptor book/table. That entry's paired vector is retrieved from a second table , scaled by a learned gain and added back to the token. A local convolution mixes neighboring tokens. The selected entries and gains are logged for analysis.

Evaluated on ImageNet-1K by training from scratch. Two model sizes Transmitter-512 and Transmitter-768 were trained and compared against other popular vision backbones such as ViT, ResNet, ConvNeXtV2. Other ablations include ImageNet-100, COCO trace diagnostics, CIFAR-10/100 study.

Critical strengths

The three steps are visible, namely, pick an entry, fetch its paired vector and scale it. The choices are recorded so that one can see which entry the model picked at each location and test what happens when the choices or gains are changed. Unlike attention maps, where one can observe where a token looked at, these records capture discrete choices that one can count and directly change to observe the effects. It could potentially help with interpretability.

Critical weaknesses

Why are the baseline numbers far lower than the official results on ImageNet? All 4 baselines are supposed to achieve over 70% of Top-1.

While interpretability is identified as a potential differentiator, it is not clearly shown what that benefit is for the recorded choices. In fact, on COCO dataset, the effects are very small. The recorded choices identify classes worse than ordinary features can do.

Questions for authors

Why are the baseline numbers far lower than the official results on ImageNet? All 4 baselines are supposed to achieve over 70% of Top-1.

While interpretability is identified as a potential differentiator, it is not clearly shown what that benefit is for the recorded choices. In fact, on COCO dataset, the effects are very small. The recorded choices identify classes worse than ordinary features can do.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.