Batch selection

[AstikaNehra]

PR Description:

• Addresses issues: [DEV] Training Loop- Repo Structure and Data Loader #10 and Batch Selection Efficiency Improvement #7
• A modular training loop for MNIST batch selection experiments
• Data loading MNIST from CSV (Kaggle: https://www.kaggle.com/datasets/oddrationale/mnist-in-csv)
• Two batch selection strategies: random and smart (highest-loss sample selection using only previous batch statistics)
• Plot accuracy and loss curves for results comparison

Key Results

• Both random and smart batch selection methods were implemented and benchmarked on MNIST.
• Smart selection, which prioritizes samples with higher recent loss, led to faster improvements and slightly higher test accuracy compared to purely random sampling, while avoiding overfitting. It can be further improved with some fine tuning.
• See plot screenshots below for side-by-side accuracy and loss curves:

Plots show that smart batch selection yields comparable or slightly better performance in both less loss and final test accuracy.

Usage Instructions

• Extract mnist_train.csv and mnist_test.csv from the zip files
• Run batch_selection_main.ipynb to reproduce experiments and plots.

Next Steps

• Extend to other batch selection strategies or datasets as needed