AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
Learning With Latent Language
The latent descriptions project grew directly out of the policy sketches project in that. The more interesting set of experiments in that paper said if you train a policy with this extra scaffolding provided by language it is also useful for fast adaptation to new tasks even when you no longer have access to those structured annotations. So somehow the language gives me a representation of a policy that's just like better as a policy and generalizes better despite not being able to use them at training time. Yeah so it does seem like a good direction forward.