Implemented a transformer-based architecture for multi-label soccer action recognition on SoccerNet-v2. Integrated DINOv2 frame embeddings with a 6-layer temporal Transformer encoder and multi-head self-attention. Achieved ~60% micro-F1, demonstrating effective modeling of long-range temporal dependencies.

Trained a Deep Convolutional GAN on CIFAR-10 and explored the generator’s latent space using PCA, SLERP, and t-SNE. Identified disentangled axes of variation corresponding to semantic factors. These analyses provided insight into how the GAN encodes visual concepts internally and helps understand generative representations.

Built a counterfactual explanation framework combining XGBoost, PyTorch models, and a VAE for plausibility scoring. Benchmarked DiCE, gradient-based, surrogate, and genetic search CFs, evaluating validity, sparsity, and interpretability. Visualized results with boxplots, t-SNE maps, and patient records.

Developed a GRU-based encoder-decoder model in PyTorch that converts numeric sequences (e.g., 123) into English words (one two three). It uses teacher forcing during training, cross-entropy loss, and greedy decoding to handle variable-length sequences accurately, demonstrating sequence-to-sequence modeling for real-world applications.