Ali Ghasemi

This project introduces a pipeline for facial animation retargeting that combines unsupervised expression transfer with 3D blendshape prediction, enabling the geometric reconstruction of facial expressions from 2D images. By bridging 2D reenactment with 3D parametric models, this work offers a modular framework for animation workflows, eliminating the need for paired training data or manual blendshape annotation. The results highlight its potential for applications in virtual avatars, AR/VR, and automated animation pipelines.

DiffiT introduces a novel approach to image generation by combining Vision Transformers (ViTs) with diffusion-based models, leveraging a Time-dependent Multihead Self Attention (TMSA) mechanism to achieve high-fidelity image synthesis. The project explores the image space model, outperforming other Transformer-based diffusion models while maintaining parameter efficiency. The implementation includes a detailed diffusion process with a linear noise schedule, efficient noisification techniques, and a learned reverse process. The model is preconfigured for datasets like CIFAR-10 and Tiny ImageNet, with flexible hyperparameters for customization. Key features include the TMSA mechanism, parameter efficiency, and dual support for latent and pixel space implementations. The project provides training and evaluation code, along with pretrained models, making it a comprehensive resource for researchers and practitioners in generative AI.

This project explores the Natural Language Inference (NLI) task, focusing on enhancing dataset performance against adversarial test sets. It incorporates word sense disambiguation and semantic role labeling to improve robustness. The primary goal is to develop an effective data augmentation strategy that strengthens the model’s ability to handle adversarial examples.

This project explores curiosity-driven exploration as an intrinsic reward mechanism for agents in environments with sparse or no extrinsic rewards. By formulating curiosity as the error in predicting action consequences within a self-supervised visual feature space, the approach enables efficient exploration and generalization. It allows agents to reach goals with fewer interactions, explore more effectively in the absence of rewards, and adapt to new scenarios by leveraging prior experience.