(Time differs from regular CLIP Colloquiums)

 

Despite considerable advances in deep learning, AI remains to be narrow and brittle. One fundamental limitation is its lack of common sense: intuitive reasoning about everyday situations and events, which in turn, requires a wide spectrum of commonsense knowledge about how the physical and social world works, ranging from naive physics to folk psychology to ethical norms. In this talk, I will share our recent adventures in modeling neuro-symbolic commonsense models by melding symbolic and declarative knowledge stored in large-scale commonsense graphs with neural and implicit knowledge stored in large-scale neural language models. I will conclude the talk by discussing major open research questions, including the importance of algorithmic solutions to reduce incidental biases in data that can lead to overestimation of true AI capabilities.