Mandela Patrick, PhD

Mandela Patrick's PhD thesis, focusing on computer vision, self-supervised learning, and multi-modal learning.

Introduces trajectory attention, a self-attention block for video transformers that implicitly tracks space-time patches along motion paths, achieving state-of-the-art results on action recognition datasets.

Focuses on better leveraging latent time and space for video representation learning by computing efficient multi-crops and using a shallow transformer to model time.

Introduces a systematic framework for transformations in contrastive learning, leading to state-of-the-art video representation learning.

Explores pre-training techniques for vision-language models to enable zero-shot cross-lingual transfer.

Research contributing to the understanding and interpretability of deep neural networks.