Machine Learning Street Talk (MLST) avatar

Dexa/Machine Learning Street Talk (MLST)

Learn more
Published Jan 28, 2023

#97 SREEJAN KUMAR - Human Inductive Biases in Machines from Language

Sreejan Kumar discusses the integration of human inductive biases in AI using program synthesis and meta-learning, aiming to boost AI's generalization and human-like efficiency. His research sheds light on aligning AI more closely with human cognitive processes through natural language and program abstractions, as highlighted in his NeurIPS presentation.
Episode Highlights
Machine Learning Street Talk (MLST) logo

Topics covered

Popular Clips

Episode Highlights

  • Meta Learning

    Meta learning serves as a powerful framework to embed human-like inductive biases into AI systems. explains that by training on a distribution of tasks, meta learning allows AI to acquire inductive priors that enhance its ability to generalize and learn new tasks more efficiently 1. This approach is crucial for aligning AI models with human cognitive priors, thereby improving their generalization capabilities.

    Meta learning is a really good computational framework to understand kind of rigorously inductive bias.

    ---

    By incorporating human-generated language and program abstractions, AI can achieve more human-like behavior, which is essential for effective collaboration and understanding between humans and machines 2.

       

    Generalization

    The research highlights the importance of human-like generalization in AI, achieved through meta learning and reinforcement learning. describes how AI agents trained on human-generated tasks can generalize better to machine-generated tasks, demonstrating the value of human inductive biases 3. This dual-task distribution approach reveals the distinct differences in how humans and machines process information.

    Humans have this ability to learn things really fast or pick up new things really quickly and strongly generalize to novel tasks.

    ---

    By co-training AI on natural language descriptions and program abstractions, the agents not only improve in tasks humans excel at but also align more closely with human values and behaviors 1.

Related Episodes