- 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
- 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.
- 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
- 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?