Explore a comprehensive lecture from ICCV'21 Tutorial on human-centric AI for computer vision and machine autonomy. Delve into the challenges and successes of AI, examining whether deep AI models can think like humans. Investigate explainable AI techniques such as Class Activation Mapping (CAM) for scene classification and medical imaging. Learn about network dissection, rapid progress in image generation, and identifying causality in latent space. Discover how to steer generative models and understand the role of individual units. Examine human-in-the-loop reinforcement learning and its impact on improving environment diversity and generalizability. Gain insights into the key impact areas of human-centric AI in this informative 39-minute presentation by Bolei Zhou.
Overview
Syllabus
Intro
Success and Challenge of Al
Can Deep Al Models Think like Humans?
Toward Human-centric Al
Al Model for Scene Classification
Explaining the Prediction Class Activation Mapping (CAM)
Explaining Model Prediction
Application to Medical Imaging
Explainable Al for Classification
Quantifying the Interpretability of Individual Units Network Dissection
Key units for classifying Living room
Key units for classifying Restaurant: Tables
Rapid Progress for Image Generation 2014
2021: Text2Image Model from OpenAI
Al Model for Image Generation
Identifying Causality in Latent Space
Pushing Latent Code to the Subspace Latent Space
Steering Generative Model
Steerable Al Model for Generation
Understanding the Role of Individual Units
Unsupervised Learning of Steerable Dimensions
Al Model for Machine Autonomy
Human-in-the-Loop Reinforcement Learning
Improving the diversity of the environment
Generalizability is improved by environment diversity
Impact Areas of Human-Centric Al
Taught by
Bolei Zhou
