A*-Inspired Batch Selection Accelerates CNN Training, Outperforms Deeper Models

Traditional Convolutional Neural Network (CNN) training often relies on randomly shuffled mini-batches, which can lead to slower convergence and a diminishing learning signal as many samples quickly become "easy." This research introduces A*-Inspired Batch Selection (A*-BS), a lightweight and model-agnostic strategy that reframes mini-batch scheduling as a heuristic search problem. Each batch is assigned an A*-like score, combining a loss-based difficulty measure with a reuse penalty, ensuring both informative gradient updates and batch diversity throughout the training process. A*-BS integrates seamlessly into existing training pipelines without requiring modifications to network architectures or optimization algorithms. The method was evaluated on the MedMNIST-v2 benchmark's twelve 2D classification tasks using a relatively simple CNN architecture. Remarkably, this lightweight model, when combined with A*-BS, achieved higher accuracy and AUC than both ResNet-18 and ResNet-50 baselines on half of the tasks, with relative gains up to 15%. Ablation studies confirmed A*-BS's superiority over random shuffling across all tasks. Furthermore, wall-clock measurements showed substantial speed improvements, indicating that intelligent batch ordering can effectively compensate for reduced architectural complexity, offering a computationally efficient alternative to deeper models.