Danilo J. Rezende (@DaniloJRezende) | Robocrunch.Net

DaniloJRezende

Danilo J. Rezende @DaniloJRezende

Researcher, Lead of the Generative Models & Inference team @ #DeepMind Building models for decision making and hard science problems.

Tweets by Danilo J. Rezende

Being "causal" is a property of a model, not a variable. A variable can "cause" another given a conditional dependence structure of a model.

Shared by Danilo J. Rezende at 5/12/2022

Cool work from colleagues at @DeepMind and collaborators approximating turbulent dynamics with convnets.

Shared by Danilo J. Rezende at 4/28/2022

Amortized variational inference doesn't work, hum?

Shared by Danilo J. Rezende at 3/17/2022

🔥New paper on flows for Lattice-QED🔥 Previously we built (S)U(N) Gauge-equivariant flows for "pure Gauge" theories. We added fermions to our U(1)-equivariant flow to explore the Schwinger model at criticality (a particularly hard regime) From our MIT - @DeepMind collaboration

Shared by Danilo J. Rezende at 2/25/2022

Not amongst expressive classes (e.g AR, VAEs, GANs, flows); exact expectations are typically cheap for discrete mixtures of simple expfam dists (e.g GMMs, prob circuits). GPs have known (un)conditional moments (given hyper params).

Shared by Danilo J. Rezende at 2/16/2022

In a hierarchical world of the form p_env(sources)p(reward | task, source)p(task). You should still expect to learn something useful by focusing on p_env(sources) and p(task) under weak assumptions about p(reward | task, source). 2/2

Shared by Danilo J. Rezende at 10/4/2021

We're hiring interns on the Generative Models & Inference team at @DeepMind If you're interested in generative models, representation learning, RL and model-based RL please consider applying: https://t.co/69lZQYbaWL

Shared by Danilo J. Rezende at 9/28/2021

Also share this view: If we include the measurement process as part of the model, all forms of uncertainty are just uncertainties at different levels in the model hierarchy.

Shared by Danilo J. Rezende at 9/27/2021