"Understanding Large Vision Models" - Antonio Torralba, TTIC Distinguished Lecture Series скачать в хорошем качестве

"Understanding Large Vision Models" - Antonio Torralba, TTIC Distinguished Lecture Series

“Understanding Large Vision Models” Antonio Torralba, Massachusetts Institute of Technology (MIT) Originally recorded on April 25, 2025, at TTIC, 6045 S. Kenwood Avenue, Chicago, IL. In this talk, Antonio Torralba explores the internal mechanisms of large pretrained vision-language models. He introduces MAIA (Multimodal Automated Interpretability Agent), a system that emulates human-like experimentation to interpret AI model components. MAIA iteratively tests hypotheses using tools like synthetic input generation, dataset sampling, and behavior summarization. The talk also reflects on the role of data in training vision systems, questioning whether image datasets are even necessary in the future of AI. Timestamps: 00:00 Introduction 02:18 Talk begins 48:05 Q&A #AI #ComputerVision #DeepLearning #Interpretability #VisionLanguageModels #MAIA #MLResearch #MIT #TTIC #PhDThesis #DistinguishedLecture #LargeScaleModels #ModelUnderstanding #GenerativeModels