The AI community is increasingly recognizing the importance of language in connecting vision and cognition, as seen with recent work on Transformers. The question remains whether language or vision is more fundamental to our understanding of the world.
Professor Yoav Artzi believes that training neural nets differently, such as teaching them to focus on causal explanations and encouraging joint learning between language and the world it references, can improve language and world modeling in artificial intelligence.
The integration of symbolic representation and statistical machine learning in identifying patterns in language and image processing. It serves as a tool to explore the world of computable knowledge.
The difficulty with self-supervised learning lies in representing uncertainty, multiple outcomes and unpredictability. Although vision appears to be less challenging than language for achieving full self-supervised learning capability, the challenges still remain vast.
The idea is that machine learning models, such as GANs or contrastive models, can all be explained by an energy function they're trying to minimize or maximize. By using a common language for these models, their similarities can be better understood and used to achieve higher levels of performance.
The discovery of good predicates is the essence of human-level intelligence that AI seeks. While deep learning is a popular method, it still relies on mediocre predicates, leaving much room for improvement.