CoAtNet-DeepMoE: A Convolution-Attention Hybrid with DeepSeek Mixture-of-Experts for Parameter-Efficient Tomato Disease Classification
Md Nadim Mahamood, Md Arif Shahriar, Md Shafi Ud Doula, Kamrul Hasan
cs.CV
Quick verdict
Reports 99.8% tomato disease accuracy with 2.47M params (CoAtNet+DeepSeek-MoE); efficiency result notable but lacks variance, fair baselines, overlap controls.
Summary
This paper proposes CoAtNet-DeepMoE, a hybrid convolution-attention classifier that replaces the MLP sub-layers of a CoAtNet-style backbone with a DeepSeek-style Mixture-of-Experts (an always-on shared expert plus top-2 routed experts implemented as 1x1-conv pairs that shrink channels 4x), targeting parameter-efficient tomato leaf disease classification. Using ~29,160 images across ten classes from a Kaggle collection and PlantVillage, the authors report 99.80% accuracy/precision/recall/F1 on the balanced Kaggle set and 99.83%/99.85%/99.76%/99.80% on the imbalanced PlantVillage set, with 2.47M parameters - a claimed ~90.73% reduction versus 26.67M CoAtNet-Base (Table 1). Ablations include bidirectional cross-dataset evaluation with near-perfect class-wise metrics (Table 7) and an inference-cost analysis showing 1.34 GFLOPs, ~4 ms/img latency, and the best p95/p99 tail latencies among CoAtNet-Base, ResNet50, and ViT-Tiny (Table 8). Main limitations: single-run results with no statistical testing or variance; SOTA comparisons pool numbers from prior papers under differing splits and preprocessing; split sizes, expert count, load-balancing loss, and pretraining status are unreported; near-ceiling and near-perfect cross-dataset scores raise concerns about overlapping data sources and augmentation-before-split practices. Code is promised on GitHub but not yet verifiable.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Replacing CoAtNet's MLP sub-layers with a DeepSeek-style Mixture-of-Experts reduces total parameters from 26,674,252 to 2,473,316 (~90.73%) with no reported accuracy loss (Table 1).
- CoAtNet-DeepMoE achieves 99.80% accuracy/precision/recall/F1 on the balanced Kaggle dataset and 99.83%/99.85%/99.76%/99.80% on the imbalanced PlantVillage dataset, claimed as state of the art (Tables 5-6).
- The model generalizes across datasets, with macro accuracy of 0.9996 (PlantVillage-to-Kaggle) and 0.9997 (Kaggle-to-PlantVillage) in cross-dataset evaluation (Table 7).
- At 1.34 GFLOPs the model is 1.6-6.3x less compute-intensive and achieves the best tail latency (p95 4.19 ms, p99 4.24 ms) among CoAtNet-Base, ResNet50, and ViT-Tiny on identical hardware (Table 8).
Strengths
Large parameter reduction with retained accuracy, traceably accounted
Table 1 gives a block-wise breakdown: Transformer2D-2 falls from 12.15M to 0.44M and Transformer2D-3 from 12.85M to 1.72M (total 26,674,252 to 2,473,316); Table 8 corroborates the footprint with 1.34 GFLOPs and the best p95/p99 latency (4.19/4.24 ms) versus CoAtNet-Base, ResNet50, and ViT-Tiny.
Generalization checks beyond in-distribution accuracy
Section 3.2.1/Table 7 report bidirectional cross-dataset evaluation (PlantVillage-to-Kaggle macro accuracy 0.9996; Kaggle-to-PlantVillage 0.9997) with per-class precision/recall/F1 - a test rarely run in this subfield.
Explicit handling of class imbalance
Deployment-oriented cost analysis
Complete training configuration and shape-annotated architecture
Weaknesses
No statistical uncertainty anywhere
Heterogeneous, indirect SOTA comparisons
Missing experimental essentials
Dataset provenance and near-ceiling scores unexamined
MoE contribution not isolated
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 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 accuracy does CoAtNet-DeepMoE achieve on tomato disease classification?
The paper reports 99.80% accuracy (and 99.80% precision/recall/F1) on a ten-class Kaggle tomato leaf dataset, and 99.83% accuracy (99.85% precision, 99.76% recall, 99.80% F1) on PlantVillage, using only 2.47M parameters. These are single-run results on standardized lab-style leaf images, not field conditions, and no variance or statistical tests are reported.
How does the model cut parameters by about 90%?
It replaces the MLP sub-layers inside CoAtNet's transformer stages with a DeepSeek-style Mixture of Experts: one small always-active shared expert plus the top-2 of several tiny routed experts, each built from 1x1 convolutions that shrink the channel dimension fourfold. Per the paper's Table 1, the two transformer stages shrink from ~12.1M and ~12.9M parameters to ~0.44M and ~1.72M, taking the total from 26.67M to 2.47M.
Is the source code available?
The paper states that source code 'will be available' at github.com/nadimbrur/CoAt-MoE. As of this arXiv preprint's appearance the repository could not be verified as accessible, and key reproduction details (expert count, split sizes, pretraining status) are not in the text.
Does CoAtNet-DeepMoE work in real farm conditions?
Not demonstrated. Evaluation uses PlantVillage and Kaggle lab-style leaf images with plain backgrounds. The authors list field-level deployment, more diverse imagery, and early-warning integration as future work, and the near-perfect cross-dataset results may be inflated if the two benchmark datasets share source images.
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.
- Figures 1, 3, 4, 5, and 6 (architecture, scatter plots, confusion matrices, cost plot) are referenced but not visually analyzable from extracted text; quantitative judgments rely on the tables, which are fully present.
- The metadata abstract is truncated mid-sentence; the full embedded text was used and is complete through the references.
- The dataset-provenance concern (Kaggle tomatoleaf set derived from PlantVillage) rests on widely known dataset lineage external to the paper, not on anything verifiable within the extracted text.
- Whether the promised GitHub repository actually contains code could not be checked offline; codeAvailable reflects the paper's own statement.
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.