Contribution-Aware Bandwidth Allocation for Multimodal Split Learning
Iason Ofeidis, Leandros Tassiulas
cs.LG · cs.DC · cs.NI
Quick verdict
ModalShare divides a fixed SL uplink budget by server-side Shapley contribution; sizable matched-payload gains (+15.4/+12.4 pp) but only M=2 tested.
Summary
The paper targets multimodal split learning (SL), where a client uploads compressed smashed activations for several modalities over one constrained uplink. Existing compressors (Top-K, RandTop-K, SplitFC) apply one keep-ratio to every modality, so transmitted payload divides in proportion to cut dimension rather than task contribution. The authors formalize this division as an iso-budget problem and propose ModalShare: the server masks already-received activations to form coalitions, computes exact Shapley scores, maps them through a temperature-controlled softmax with a keep-ratio floor into per-modality keep-ratios, and freezes once a confidence gap stabilizes. The probe adds no uplink traffic and no client compute. On CREMA-D, MVSA-Single, and UCI HAR with ResNet-18/DistilBERT/1D-CNN encoders and early cuts, ModalShare beats matched uniform allocation by 15.4 pp (CREMA-D) and 12.4 pp (MVSA) at 5x compression, +1.1 pp on UCI HAR, winning 29 of 36 configurations across three compressors and four budgets. A decomposition attributes most of MVSA's gain to contribution (7.23 of 12.40 pp) over dimension-skew correction (5.17 pp). Limitations: only two modalities evaluated, no multi-client experiments, high seed variance in some cells without significance testing, MVSA test accuracy taken from the validation partition, and no code release.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Equal keep-ratios in multimodal SL allocate transmitted payload in proportion to cut dimension, a quantity unrelated to a modality's role in the fused prediction (Sec. I, Table II).
- A server-side Shapley probe over coalitions of already-received smashed activations can set per-modality keep-ratios with zero uplink and zero client-side cost (Sec. III-D, Fig. 2).
- ModalShare improves accuracy over uniform allocation by 15.4 pp on CREMA-D and 12.4 pp on MVSA at matched payload under 5× compression with SplitFC (Table IIIa).
- The gain is contribution-driven, not dimension-skew correction: on MVSA 7.23 of the 12.40 pp exceed what any dimension-based rule supplies, and on balanced-dimension CREMA-D the skew component is absent (Table IV, Sec. V-C).
- Cut-local signals (activation norm, gradient norm, solo utility) cannot substitute for the fusion-level score and recover none of the gain (Table V).
Strengths
Novel problem formulation with clean isolation
Section III-C defines inter-modality allocation as an iso-budget problem (Eq. 6): all compared policies transmit the same expected payload TX, so accuracy differences are attributable to the split alone. Table II quantifies the status quo it corrects (MVSA image silently takes 80.3% of payload under equal keep-ratios).
Zero-overhead probe design
The Shapley probe (Sec. III-D) masks activations the server already holds: no uplink traffic, no client compute, no modality-identity prior. The freeze rule (γ=0.90, K=3, emin=10) eliminates residual server cost; Table VII shows freeze even outperforms continuous online allocation by +5.7 pp (CREMA-D) and +3.5 pp (MVSA).
Rigorous attribution of where the gain comes from
Honest negative results and failure-mode reporting
Above-average experimental protocol
Weaknesses
Only M=2 evaluated despite general-M formulation
No significance testing despite large variances
MVSA evaluation on the validation partition
Probe validity rests on an unvalidated masking convention
No code release; modest absolute baselines
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper addresses multimodal split learning (SL): one client hosts per-modality encoders truncated at an early cut and uploads compressed smashed activations for all modalities over a shared uplink; the server completes fusion and training. Its thesis (Sec. I, Sec. III-C) is that existing SL compressors (Top-K [7], RandTop-K [8], SplitFC [5]) make only an intra-tensor decision — the same keep-ratio for every modality — which silently divides transmitted payload in proportion to cut dimension (Table II: MVSA image takes 80.3% of payload because its cut is ~4× the text cut). The authors name the missing decision inter-modality allocation, formalize it as an iso-budget problem (Eq. 6), and propose ModalShare (Sec. III-D): a server-side Shapley probe over coalitions of already-received activations, mapped through a temperature softmax and keep-ratio floor to per-modality keep-ratios, with confidence-based freezing. Central claims: (i) dimension-proportional allocation is arbitrary; (ii) a fusion-level Shapley score is measurable at zero uplink/client cost; (iii) contribution-aware allocation beats uniform by 15.4/12.4 pp at 5× compression on CREMA-D/MVSA; (iv) the gain is contribution, not skew correction; (v) cut-local proxies fail.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ModalShare?
ModalShare is an allocator for multimodal split learning that divides a fixed uplink bandwidth budget across sensor modalities according to server-measured Shapley contribution scores, instead of giving every modality the same compression keep-ratio (which silently hands more bandwidth to whichever modality has the larger cut dimension).
How much does ModalShare improve accuracy?
At 5× compression with matched total payload, the paper reports +15.4 percentage points on CREMA-D (audio-video emotion recognition) and +12.4 pp on MVSA (image-text sentiment) over equal keep-ratios, using the SplitFC compressor; on UCI HAR, where the two sensor streams contribute near-equally, gains are only about +1 pp.
Does the Shapley contribution probe cost extra communication?
No. The server forms coalitions by masking activations it has already received, so the probe consumes no uplink traffic and no client computation. Its only cost is about 2^M extra server forward passes (four for two modalities) every 10 batches, which stops entirely once the allocation freezes.
Is code or data available for this paper?
The paper text contains no code or data release statement. The three datasets (CREMA-D, MVSA-Single, UCI HAR) are public, and hyperparameters and architectures are specified in enough detail for independent reimplementation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The split-depth ablation (Fig. 4) reports exact per-depth ∆ values only as figure data; the extraction captures the qualitative trend and the quoted uniform-share endpoints (MVSA image 80.3% at layer 1 → 33.8% at layer 4) but not the full numeric series, so those figure-only points were assessed qualitatively.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.