Dynamic neural architecture search : A pathway to efficiently optimized deep learning models
Chetan Gode, Bhushan Marutirao Nanche, Dharmesh Dhabliya, Rahul Dnyanoba Shelke, Rajendra V. Patil, Sushma Bhosle · Journal of Information and Optimization Sciences · 2025
Deep learning models have changed many areas by making it possible for them to do things like picture recognition, natural language processing, and reinforcement learning at the highest level. But optimizing these models is still very hard because they are very complicated and need a lot of computing power. This article suggests a way to make deep learning models work better by looking at important parts like model architecture design and training methods. Using methods like neural architecture search and model distillation, we present a methodical way to create systems that work well. This makes it possible to make models that are small but strong, which is good for places with limited resources. We also talk about how important it is to use the right start and regularization methods to avoid overfitting and make generalization better. We look into new ways to train that are meant to speed up convergence and lower the cost of computing. This includes methods like mixed-precision training, progressive shrinking, and transfer learning that use parallelism in both data and hardware to speed up training without lowering performance. We also look into how domain-specific information can be used to guide the optimization process and make models better able to meet the needs of particular tasks. To see how well our suggested route works, we run a lot of tests on a lot of different datasets and standards. The outcomes show big changes in model efficiency, with performance being on par with or even better than standard methods while using a lot less computing power.