Search-optimized quantization in biomedical ontology alignment.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Computational Science

Background:

• Large AI models present deployment challenges in resource-constrained environments due to size and computational demands.
• Efficient model optimization is crucial for edge devices, addressing energy, memory, and latency issues.
• Transformer-based models are increasingly complex, necessitating advanced optimization strategies.

Purpose of the Study:

• To introduce a systematic method for ontology alignment using transformer-based models.
• To optimize AI models for efficient deployment on edge devices.
• To achieve state-of-the-art performance in biomedical tasks with reduced resource consumption.

Main Methods:

• Employed supervised state-of-the-art transformer-based models for ontology alignment.
• Utilized cosine-based semantic similarity between a biomedical layman vocabulary and the Unified Medical Language System (UMLS) Metathesaurus.
• Leveraged Microsoft Olive, ONNX Runtime, Intel Neural Compressor, and Intel Extension for PyTorch (IPEX) for model optimization via dynamic quantization.

Main Results:

• Achieved a new state-of-the-art performance on two tasks from the DEFT 2020 Evaluation Campaign.
• Attained an average inference speed-up of 20x.
• Reduced memory usage by 70% while retaining performance metrics.

Conclusions:

• The proposed systematic method effectively optimizes large AI models for resource-constrained environments.
• Efficient AI model deployment in biomedical informatics is feasible with significant performance gains.
• Advanced optimization techniques like dynamic quantization are key to overcoming AI deployment challenges.