MIT EI Seminar - Phillip Isola - Emergent Intelligence- Getting More Out of Agents Than You Bake In

Explore emergent intelligence in AI systems through this MIT Embodied Intelligence Seminar talk. Delve into three projects showcasing surprising outcomes beyond explicit programming: unsupervised learning discovering semantics without labels, generative models producing physically plausible "videos" from static images, and self-assembling "creatures" emerging from primitive limbs without centralized control. Gain insights into the goal of maximizing AI output while minimizing input constraints, and examine topics such as representation learning, colorization, neural representations, self-supervised objectives, generative modeling, latent variable models, multi-agent interactions, and self-assembly morphologies. Discover how these concepts relate to evolution in nature and the emergence of complex systems from simple components.

Introduction
Recipe for intelligence
Representation learning
Colorization
Neural representation
Selfsupervised objectives
The Cave
Multiview representation
Contrastive multiview representation
Data augmentation
Semantics
Morris Different
generative modeling
dream of a model
continuum
latent variable models
camera transformations
latent space vector
word Tyvek
Vectors
Circles
Transformations
Bias
Biases
Multiagent interactions
Emergent arms race
Evolution in nature
Spiral of evolution
Emergence of multicellular life
Primitive agents
Selfassembly morphologies
Long creatures
Training
Modularity
Slime Mold Creatures
Obstacles
Multiagent
Selfassembling