💻 computer science

BioAutoML-NAS: An End-to-End AutoML Framework for Multimodal Insect Classification via Neural Architecture Search on Large-Scale Biodiversity Data

BioAutoML-NAS is a novel end-to-end AutoML framework that leverages neural architecture search and multimodal fusion of images and metadata to achieve state-of-the-art accuracy in large-scale insect classification, significantly outperforming existing transfer learning, transformer, and AutoML methods on biodiversity datasets like BIOSCAN-5M.

Original authors: Arefin Ittesafun Abian, Debopom Sutradhar, Md Rafi Ur Rashid, Reem E. Mohamed, Md Rafiqul Islam, Asif Karim, Kheng Cher Yeo, Sami Azam

Published 2026-06-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Arefin Ittesafun Abian, Debopom Sutradhar, Md Rafi Ur Rashid, Reem E. Mohamed, Md Rafiqul Islam, Asif Karim, Kheng Cher Yeo, Sami Azam

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master chef trying to identify thousands of different types of insects just by looking at photos of them and reading their ID tags. This is a huge job for farmers and scientists because knowing exactly which bug is in a field helps protect crops and nature. But insects are tricky: there are too many of them, some look almost identical, and the data is messy (some bugs are super common, while others are rare).

The paper introduces a new "smart kitchen" called BioAutoML-NAS. Instead of a human chef trying to guess the best recipe, this system builds its own perfect recipe from scratch. Here is how it works, broken down into simple concepts:

1. The Two-Ingredient Soup (Multimodal Data)

Most old methods tried to identify insects using only a photo (like looking at a picture of a fruit). But this new system is smarter. It uses two ingredients:

The Visuals: The actual photo of the insect.
The Metadata: The "ID card" or biological notes attached to the photo (like its family name, DNA barcode, or scientific classification).

Think of it like identifying a person. Old methods just looked at a photo of their face. This new method looks at the face and reads their passport and family tree at the same time. By combining these two, the system gets a much clearer picture of who the insect is.

2. The Self-Designing Blueprint (Neural Architecture Search)

Usually, when scientists build a computer brain (a neural network) to recognize things, they have to manually design the blueprint. They decide how many layers to stack and what kind of filters to use. It's like an architect trying to design a skyscraper by hand, guessing which beams will hold the most weight.

BioAutoML-NAS does this automatically. It uses a technique called Neural Architecture Search (NAS).

Imagine a giant toolbox with 10 different types of tools (like different lenses, filters, or magnifying glasses).
The system builds a small "cell" (a mini-brain) and tries out every possible combination of these tools to see which one works best for the insect photos.
It stacks these cells together to build the full network.
The Magic Pruning: As it learns, the system realizes some tools are useless. It has a special "zero operation" (a "do nothing" button) that it can press to cut out the weak connections. It's like a sculptor chipping away the excess stone until only the perfect statue remains. This makes the final model fast, light, and efficient.

3. The Training Dance (Alternating Optimization)

Training this system is like a dance between two partners:

The Weight Partner: Adjusts the "strength" of the connections (learning what an insect looks like).
The Structure Partner: Adjusts the "blueprint" (deciding which tools to keep or cut).

They take turns. One step, the system learns the features; the next step, it tweaks the design. They keep switching back and forth until they find the perfect rhythm where the design and the learning work together perfectly.

4. The Results: A New Champion

The researchers tested this system on a massive dataset called BIOSCAN-5M, which contains 5 million insect images.

The Score: It got 96.81% accuracy.
The Comparison: It beat the previous best methods (which were like trying to solve a puzzle with a hammer) by a huge margin—about 16% better than standard deep learning, 10% better than Transformer models, and 8% better than other AutoML methods.
The Test: They also tested it on a different dataset (Insects-1M) with over a million images, and it still scored over 93% accuracy, proving it doesn't just memorize the first test but actually learns how to identify bugs.

Why This Matters

The paper claims this is the first system to use this specific "self-designing" approach on insect data that includes both photos and biological metadata.

For Farmers: It means more accurate pest detection, which helps in sustainable farming.
For Science: It helps scientists track biodiversity without needing to manually sort through millions of images.

In short, BioAutoML-NAS is a self-improving robot chef that reads both the picture and the recipe card to identify insects faster and more accurately than any human-designed system has ever done before. It builds its own brain, cuts out the useless parts, and gets the job done with incredible precision.

Technical Summary: BioAutoML-NAS

Problem Statement

Insect classification is critical for agricultural management and ecological research but faces significant challenges due to the complex morphological characteristics of insects, severe class imbalance in large-scale datasets, and the limitations of conventional deep learning approaches. Existing methods often rely on single-modal data (e.g., images only) or predefined architectures (such as standard CNNs, Transformers, or fixed transfer learning models) that fail to fully optimize performance on diverse, large-scale biodiversity data. Furthermore, many approaches lack the flexibility to co-optimize network structure and weights, leading to rigid, resource-intensive models that struggle with generalization and class imbalance without relying on synthetic data or extensive manual post-processing.

Methodology

The authors propose BioAutoML-NAS, an end-to-end AutoML framework designed for multimodal insect classification using Neural Architecture Search (NAS). The framework operates on the BIOSCAN-5M and Insects-1M datasets, integrating both image data and biological metadata (e.g., DNA barcodes, taxonomic ranks).

Core Components

Multimodal Architecture:
- Image Encoder: Utilizes a NAS-based approach to automatically discover the optimal architecture. The search space consists of a structured cell containing two intermediate nodes and ten heterogeneous primitive operations, including depthwise separable convolutions (3×3, 5×5), dilated convolutions, 1×1 convolutions, Squeeze-and-Excitation (SE) blocks, local pooling (average/max), skip connections, and zero operations.
- Metadata Encoder: A dedicated encoder transforms discrete biological descriptors (DNA barcodes, taxonomic hierarchy) into dense feature embeddings.
- Fusion Module: A concatenation-based fusion module integrates the image embeddings and metadata representations into a unified feature space for classification. The model employs modality-specific encoders that learn in a shared latent space, allowing interaction through aligned representations rather than complex cross-modal fusion operations.
Training Strategy:
- Alternating Bi-Level Optimization: The framework employs a gradient-based NAS strategy where network weights ( $w$ ) and architecture parameters ( $\theta$ ) are updated in an alternating fashion. In even mini-batches, weights are updated while architecture parameters are frozen; in odd mini-batches, the reverse occurs. This mitigates gradient interference and stabilizes the search process.
- Sparsity and Pruning: The inclusion of "zero operations" acts as a learnable pruning mechanism. During training, less important connections are suppressed (weights approach zero), resulting in sparse, efficient architectures without manual pruning.
- Imbalance Handling: The model addresses class imbalance through label smoothing (0.1) and dropout-regularized multimodal fusion, avoiding the need for re-sampling or class weighting.
Search Space Design:
- The search space is task-specific and biologically informed, designed to balance representational diversity and computational efficiency. It allows for the automatic selection of operations for each connection within the cell, enabling the discovery of compact, context-aware architectures.

Key Contributions

First Multimodal BioAutoML Model: To the authors' knowledge, this is the first BioAutoML model trained on multimodal data (images and metadata) specifically for insect classification, representing a novel approach to ecological inference.
Integrated Framework: The model uniquely combines multimodal fusion, alternating bi-level optimization, and zero-operation-based sparsity to deliver high-performing architectures with reduced complexity.
Robustness to Imbalance: The approach mitigates data imbalance through specific regularization techniques (label smoothing, dropout) and NAS-driven optimization, ensuring balanced learning without re-sampling.
Dynamic Architecture Evolution: Unlike static designs, the framework continuously refines the network structure during training, enabling data-driven, self-optimizing architecture learning.

Experimental Results

The framework was evaluated on the BIOSCAN-5M dataset (training/validation/testing) and the Insects-1M dataset (cross-dataset validation).

Performance on BIOSCAN-5M:
- Accuracy: 96.81%
- Precision: 97.46%
- Recall: 96.81%
- F1-Score: 97.05%
- Macro-F1: 80.29% (demonstrating robustness across minority classes).
Performance on Insects-1M:
- Accuracy: 93.25%
- Precision: 93.71%
- Recall: 92.74%
- F1-Score: 93.22%
Comparative Analysis:
- BioAutoML-NAS outperformed state-of-the-art (SOTA) transfer learning models (e.g., ResNet, DenseNet, EfficientNet) by approximately 16%.
- It surpassed SOTA Transformer-based models (e.g., ViT, Swin Transformer, BEIT) by approximately 10%.
- It exceeded existing AutoML and NAS-based methods (e.g., ENAS, NAO, AutoKeras) by approximately 8%.
- It significantly outperformed Vision-Language Models (VLMs) like LLaVA and BioCLIP in zero-shot settings.
Efficiency: The discovered architecture achieved these results with only 2.95 GFLOPs, 40.06 MB of memory, and an inference latency of 7.42 ms, demonstrating a superior balance between accuracy and computational cost compared to other search strategies (e.g., Reinforcement Learning or Evolutionary Search).

Significance and Claims

The paper claims that BioAutoML-NAS establishes a new state-of-the-art for large-scale insect classification in biodiversity research. Its significance lies in:

Automated Discovery: Moving beyond fixed or generic designs to discover architectures specifically tailored to the complexities of biological data.
Multimodal Synergy: Demonstrating that integrating metadata with images significantly improves classification confidence and accuracy compared to image-only models (a drop of ~10.87% was observed when metadata was removed).
Scalability: Proving that gradient-based NAS can effectively handle massive, imbalanced datasets (millions of images) without relying on synthetic data augmentation or hand-crafted post-processing.
Practical Application: The model provides accurate, confident insect classification that supports modern sustainable farming and ecological monitoring.

The authors acknowledge limitations, noting that extending the model to genus and species levels is challenging due to label sparsity in the current datasets. They suggest future work should focus on improving robustness under noisy or incomplete metadata conditions and exploring semi-supervised learning techniques.