Rebooting AI: Adding Knowledge to Deep Learning | Gary Marcus | Knowledge Connexions 2020

Gary Marcus discusses the limitations of current AI, particularly deep learning, in achieving true artificial general intelligence (AGI). He contrasts the AI we have, which excels at specific tasks like playing games and transcribing speech, with the AGI we desire, capable of complex reasoning, causal inference, and operating in novel environments. Marcus critiques deep learning's superficial understanding and susceptibility to errors, exemplified by misclassifications and nonsensical language generation. He advocates for a hybrid approach that combines the strengths of deep learning (perception) with classical AI (abstraction, reasoning, and knowledge representation). Marcus emphasizes the importance of incorporating innate knowledge and common sense into AI systems, drawing parallels to human cognition and the need for AI to move beyond ad tech and address real-world problems like healthcare and pandemic preparedness. The discussion extends to the technical aspects of integrating knowledge graphs with deep learning, the political biases within the AI community, and the need for better evaluation benchmarks that assess comprehension rather than just task-specific performance.