Self-supervised learning is an approach in which AI systems learn from data without relying on human-labeled examples – by identifying patterns on their own. With the rise of models like ChatGPT and multimodal systems such as SigLIP, which are trained on vast amounts of labeled data, an important question arises: is self-supervised learning still the right path forward?

This talk argues that it is. Inspired by how humans – especially infants – learn from the world around them, self-supervised learning remains a powerful and scalable paradigm. It is particularly valuable in domains where labeled data is scarce, expensive, or difficult to obtain, such as image, video, and 3D data, as well as multimodal applications.

Despite the growing dominance of large multimodal models, self-supervised approaches continue to play a key role in advancing AI capabilities. In particular, we will explore why video is emerging as the next frontier – offering unprecedented scale and richness of data. Leveraging video data opens the door to a new generation of visual AI systems with more robust, human-like understanding.