US20260260091
2026-09-03
Physics
G06N3/04
The patent application introduces a method to generate hardware-aware machine learning (ML) model architectures for various domains without the need for training. The solution involves an apparatus with memory and processing capabilities to execute machine-readable instructions. The processor generates multiple candidate architectures based on the target hardware and a predefined search space across different domains. It computes composite scores for these candidates, focusing on latency scores, and selects an optimal architecture that meets specific criteria.
Machine learning models, like neural networks, are essential for solving complex problems in fields such as pattern recognition and natural language processing. Traditionally, these models require a training phase to determine optimal parameters, which can be resource-intensive. Neural architecture search (NAS) programs automate model development but are often limited to single domains and require significant computing resources. Existing NAS programs also struggle with adapting to different hardware environments, which is a critical challenge for deploying models on resource-constrained devices.
Current NAS methods, including one-shot NAS and zero-cost proxy NAS, face challenges such as high computational demands and limited adaptability across domains. One-shot NAS reduces computational load by selecting sub-models from a general model but still requires significant resources. Zero-cost proxies attempt to alleviate this by using inherent model characteristics for evaluation, but they often consider limited factors, leading to suboptimal performance across diverse tasks.
The disclosed approach addresses these limitations by introducing a multi-model, hardware-aware, training-free NAS method. It constructs compact neural network architectures from a human-designed search space applicable to multiple domains. The method employs a hardware-aware search strategy and uses a holistic train-free score to evaluate candidate architectures. This score considers factors such as expressivity, complexity, saliency, diversity, and latency, allowing for the selection of optimal models without extensive training.
This innovation offers significant advantages, including reduced computational requirements and enhanced adaptability across various hardware and domains. By eliminating the need for training, it simplifies the deployment of ML models on constrained devices. The holistic evaluation approach ensures that selected architectures are not only efficient but also well-suited for their intended tasks, potentially transforming how ML models are developed and deployed across industries.