My TU Brain

Proposal for the Next Paper on Vision Transformer and Contrastive Learning for Aquaculture Diagnosis

·#vision-transformer#self-supervised-learning#xai#edge-deployment#aquaculture

Overview

The strongest proposal is to extend the current shrimp disease classification study into an edge-deployable and explainable aquaculture diagnosis framework built on self-supervised Vision Transformer representations. This direction is stronger than a simple dataset swap because it preserves the core contribution of the current paper — ViT plus SimCLR on limited labels — while adding two practical gaps that matter for real-world deployment: low-resource inference and model interpretability.

The proposed paper should position the current work as the teacher or baseline system: supervised ViT/EfficientNet reached about 98% accuracy and SimCLR plus ViT-Small reached about 85% on the 4-class shrimp dataset from 4,348 images. Building on that, the new work can aim to show that self-supervised features are not only useful for classification, but can also be compressed for edge devices and audited through explainable AI methods.

Proposed Paper Title

Edge-Deployable and Explainable Self-Supervised Vision Transformer for Aquaculture Disease Diagnosis

Alternative title:

From Contrastive Pretraining to Trustworthy Edge Inference: A Lightweight Explainable ViT for Shrimp Disease Detection

Core Research Question

Can a self-supervised ViT trained with SimCLR retain strong disease recognition performance after compression for edge deployment, while also producing trustworthy disease-focused explanations through transformer-specific XAI methods?

Main Hypothesis

SimCLR pretraining learns transferable visual representations from unlabeled aquaculture imagery, and these representations can be distilled into a lightweight student model with limited performance loss for on-device inference. In parallel, transformer-native explanation methods such as attention rollout and GMAR can verify whether the compressed model still focuses on anatomically meaningful disease regions rather than spurious background cues.

The paper should contribute in three layers:

  1. Self-supervised representation learning for aquaculture using SimCLR plus ViT, motivated by the low-data advantage of SSL in domain-specific vision tasks.
  2. Edge optimization through quantization, pruning, or knowledge distillation into MobileViT or another lightweight student model, following recent edge-ViT compression results.
  3. Explainability by comparing Grad-CAM, attention rollout, and GMAR to evaluate whether the model attends to disease symptoms such as white spots or dark gill areas.

This combination is the most publishable because many papers cover only one of these topics in isolation, while the joint formulation is still sparse in aquaculture and agricultural health monitoring.

Proposed Methodology

The four stages form a single pipeline from unlabeled imagery to explainable edge inference:

flowchart LR
  A["Unlabeled aquaculture images"] --> B["SimCLR pretraining"]
  B --> C["ViT-Small teacher"]
  C -->|"fine-tune on 4-class labels"| D["Teacher classifier"]
  D -->|"knowledge distillation"| E["MobileViT student"]
  E --> F["INT8 quantization"]
  F --> G["Edge inference (Raspberry Pi / Jetson)"]
  D --> H["XAI: Rollout / Grad-CAM / GMAR"]
  E --> H
  H --> I["Disease-region saliency check"]

Stage 1: Teacher Model

Use the current SimCLR plus ViT-Small pipeline as the teacher model. This preserves continuity with the current work and avoids rewriting the full training framework. The teacher should be pretrained on all available unlabeled shrimp or aquaculture images, then fine-tuned on the 4-class labeled shrimp dataset.

Stage 2: Student Model for Edge Deployment

Distill the teacher into a lightweight student such as MobileViT-XS, MobileViT-S, TinyViT, or EfficientViT. Recent agricultural IoT distillation work shows that attention plus logit distillation can preserve much of the teacher’s performance while reducing compute by about 95% and memory to about 13 MB.

Stage 3: Compression Pipeline

Evaluate three optimization options:

  • INT8 post-training quantization or quantization-aware training for deployment efficiency.
  • Structured pruning of attention heads or tokens for further latency reduction.
  • Distillation as the main route for performance-preserving compression.

The paper does not need to include all three as equal contributions. A cleaner design is to use distillation as the main method and add INT8 quantization as the deployment step.

Stage 4: XAI Module

Apply transformer-specific explainability methods to both teacher and student models:

  • Attention Rollout as the simplest ViT-native baseline.
  • Grad-CAM adapted for transformer blocks using pytorch-grad-cam.
  • GMAR as the main XAI contribution because it weights attention heads by gradient importance and has shown improved interpretability over standard rollout.

This allows the paper to answer a strong practical question: after compression, does the student still look at the same disease-relevant regions as the larger teacher?

Experimental Design

Primary Task

4-class shrimp disease classification: Healthy, BG, WSSV, and WSSV_BG, using the same dataset protocol as the current study for fair comparison.

Secondary Task

Optional cross-domain transfer experiment: pretrain the teacher using mixed unlabeled aquaculture imagery from shrimp and fish disease datasets, then fine-tune only on shrimp labels. This extension would strengthen the paper if additional unlabeled data are available.

Evaluation Metrics

Category Metrics
Classification Accuracy, macro F1, recall per class, confusion matrix
Edge deployment Model size, parameters, FLOPs, latency, FPS, peak memory
Explainability Qualitative saliency maps, insertion/deletion fidelity, expert inspection
Generalization Performance under different lighting, farm, or background conditions

The deployment evaluation should be performed on at least one realistic device such as Raspberry Pi 5 or Jetson Orin Nano, because edge claims are much stronger with actual hardware evidence than with FLOP-only estimates.

Paper Structure

A strong structure would be:

  1. Introduction: practical need for low-cost and trustworthy aquaculture diagnosis.
  2. Related work: shrimp/fish disease detection, SSL in low-data vision, ViT edge compression, and XAI for transformers.
  3. Method: SimCLR teacher, distilled student, compression pipeline, XAI module.
  4. Experiments: classification, deployment benchmarks, explanation comparisons.
  5. Discussion: trade-off between accuracy, latency, and interpretability.
  6. Conclusion: field-deployable and explainable aquaculture diagnosis.

Why This Proposal Is Strong

This proposal is stronger than a simple domain switch for three reasons. First, it directly extends the current paper instead of abandoning it, so your prior code, results, and narrative remain useful. Second, edge deployment addresses a real application barrier, because ViT models are often too expensive for small embedded devices without compression. Third, XAI improves scientific trust, especially for fine-grained disease cues where users need visual evidence that the model is detecting lesions rather than background artifacts.

Simpler Backup Proposal

If the full edge plus XAI proposal feels too large, the best reduced-scope paper is:

Explainable Self-Supervised Vision Transformer for Shrimp Disease Detection

This version keeps the current SimCLR plus ViT pipeline and adds a rigorous explainability study comparing Attention Rollout, Grad-CAM, and GMAR across the 4 classes. It is easier to finish and still gives a clear new contribution with minimal architectural change.

Final Recommendation

The best new paper proposal is to develop a distilled, edge-deployable, and explainable SimCLR-pretrained ViT system for shrimp disease detection. If time, hardware access, or implementation scope is limited, the best fallback is an XAI-focused extension of the current paper, because it is the fastest route to a publishable and defensible next contribution.