I sent someone a list of authors and papers in causality and machine learning. I figured I should post them here so you can see what it’s all about.
- David Lopez-Paz. His thesis is a great overview of using machine learning to do causal inference. I especially like his paper marrying deep embeedings and causal inference.
- Murat Kocaoglu. His CausalGAN paper is an example of how causality can do things that would be impossible in normal machine learning.
- Uri Shalit, Fredrik Johansson, and David Sontag. Their papers on counterfactual inference ([1], [2]) – developing neural architectures and bounds on treatment effects – are great.
- Sister-brother pair Finnian and Tor Lattimore. They authored a paper on “Causal Bandits” which is a step towards automated scientific hypothesis testing.
- Some intriguing pseudo-causal work came out of Montreal where they showed how you might use VAEs to learn a latent representation of the environment such that each factor in the representation was independently controllable. This is useful for experimenting in the RL setting. (Thanks for the tip, Giorgio Patrini.)
- Chris Louizos co-authored an interesting paper on building variational approximations of a latent space to optimally represent confounders. It sounds insane to pull a confounder out of thin air.
- Jonas Peters, Dominik Janzing, and Bernhard Schölkopf have just published the phenomenal Elements of Causal Inference (PDF here). Schölkopf and Janzing are leading an exceptional group at MPI Tübingen pioneering methods for causal inference. The two weeks I spent there for MLSS 2017 were among the best two of the DPhil.
- Brenden Lake et al wrote Building Machines that Learn and Think like People, which argues causality (among other patterns) are inherent in human-like intelligence. DeepMind wrote a great response arguing that these patterns should emerge in an automated way.
- Other names I have not mentioned: Ricardo Silva, Joris Mooij, Jonas Peters, Mateo Rojas-Carulla, Paul Rubinstein, and many others. Most common names are left out because I’m more interested in how causality can improve machine learning, not how machine learning can do causal inference.
Hopefully these illustrate the subfield’s potential. We can achieve a lot if we relax our dogmatic restrictions about causality, and instead borrow from its toolkit useful tools like experiments, confounding, counterfactuals, and conditional independence.
If you want to follow literally every paper on causality and machine learning, checkout this GitHub repo I update semi-frequently that tracks them