Sat Dec 09 08:00 AM -- 06:30 PM (PST) @ 203
Learning Disentangled Features: from Perception to Control
Emily Denton · Siddharth Narayanaswamy · Tejas Kulkarni · Honglak Lee · Diane Bouchacourt · Josh Tenenbaum · David Pfau
An important facet of human experience is our ability to break down what we observe and interact with, along characteristic lines. Visual scenes consist of separate objects, which may have different poses and identities within their category. In natural language, the syntax and semantics of a sentence can often be separated from one another. In planning and cognition plans can be broken down into immediate and long term goals. Inspired by this much research in deep representation learning has gone into finding disentangled factors of variation. However, this research often lacks a clear definition of what disentangling is or much relation to work in other branches of machine learning, neuroscience or cognitive science. In this workshop we intend to bring a wide swathe of scientists studying disentangled representations under one roof to try to come to a unified view of the problem of disentangling.
The workshop will address these issues through 3 focuses:
What is disentangling: Are disentangled representations just the same as statistically independent representations, or is there something more? How does disentangling relate to interpretability? Can we define what it means to separate style and content, or is human judgement the final arbiter? Are disentangled representations the same as equivariant representations?
How can disentangled representations be discovered: What is the current state of the art in learning disentangled representations? What are the cognitive and neural underpinnings of disentangled representations in animals and humans? Most work in disentangling has focused on perception, but we will encourage dialogue with researchers in natural language processing and reinforcement learning as well as neuroscientists and cognitive scientists.
Why do we care about disentangling: What are the downstream tasks that can benefit from using disentangled representations? Does the downstream task define the relevance of the disentanglement to learn? What does disentangling get us in terms of improved prediction or behavior in intelligent agents?