Adversarial Physics, Transparent Robots, and a Band of Bayesians: NIPS Days 5/6
If I was yearning for a time-turner during the four-way-tracked Symposium on Thursday (and: I was), that was nothing to intellectual FOMO of Friday and Saturday, when on the order of 20 workshops were scheduled on top of each other each day.
The workshop on “Machine Learning for the Physical Sciences” highlighted a lot of ways that ML was being used to further scientific discovery, but two had particular interest for me, due to their having some element of creativity or cleverness that might be able to be more broadly leveraged.
First, Giles Louppe spoke from a group connected to the ATLAS experiment at the LHC, and discussed building GANs to try to replace highly expensive physics simulations necessary to do exploratory work there. “Simulations” here refers to simulations of the expected particle scatter behavior that you would expect from different potential underlying theories. Simulations are necessary because of the physical complexity of the processes the scientists are trying to capture; the workflow consists simulating the particle behavior we would expect under a given physical theory, and then comparing statistics of that behavior to statistics of the observed behavior.
These simulations, which today are mostly carried out via long Monte Carlo Markov Chains, are enormously computationally costly (side note: in fact-clarifying this, I became aware of the Worldwide LHC Computation Grid, where universities and data centers around the world contribute compute time to calculations like these). The paper being described using GANs, conditioned on the parameters of the experiment, to generate these simulated behaviors. They showed promising early work on one small subset of this behavior, but I’m not quite sure yet a) what physicists think about the plausibility of this approach, and b) how confident we’d have to be in this kind of non-parametric simulation in order for us to actually, practically implement. Nonetheless, it is an interesting space to watch
Secondly, researchers at the University of Illinois trained a network able to check for the presence of certain kinds of gravitational wave signal in LIGO data, again with the idea of cutting down on the computational time needed at classification time. What I find most compelling about this network is the way it was trained: entirely on realistic background noise, combined with artificial signals generated by applying Einstein’s equations. That is to say: it was trained entirely on simulated data, albeit simulations based on very well-trusted equations. And, despite not being trained on the real thing, it performed comparably to old fixed-template-matching methods at test time, with the use of many fewer computational resources. To me, this underlines the creative potential for using ML in cases where we have very high-quality simulations of the behavior we’re trying to learn.
Understanding T-Cell Behavior with ML
The next most memorable highlight for me came from machine learning being applied to a different kind of science: immunotherapy approaches to fight cancer. In immunotherapy, you try to leverage the body’s natural T-cell defenses to promote their being able to identify and destroy cancer cells. As part of this work, it’s useful to understand what kinds of antigens (potential foreign proteins) a person’s T-cells are capable of detecting. T-cells are unique to a given person, even when those people are genetically identical. Because of the huge number of different types of T-cells, and the varieties in which antigens a given T-cell can detect, it’s currently intractable to simple categorize those present in a patient’s body. However, Chayes and her team are working on methods to predict T-cell antigen binding using machine learning on top of the T-cell’s RNA sequence, with the hope of being able to find regularities in the mapping between that RNA expression and antigens that allow them to more easily test patients and hypothesize what kind of antigens their body is currently capable of detecting.
Designing Robots to Keep Us Informed
While watching humans interacting with robots, Anca Dragan made an observation: it can be pretty frustrating process for the humans. Humans have evolved to be quite good at reading the intentions and body language of other humans, but in the case of robots, it can be quite unclear what a robot is planning to do next, making it difficult to conduct cooperative operations. For example, if a human is trying to pilot a controlled robot to follow an autonomous robot, the robot may behave in ways the human didn’t predict, which makes it difficult for the human to plan their own actions properly.
Dragan’s work tries to remedy this problem by designing robots that have as part of their objective function a desire for the humans they interact with to have high and accurate confidence about the robot’s next action. They do this by modeling the human as a perfect Bayesian reasoner (obviously an incorrect model, but one that still gives interesting results). Some of the results of a system like this are: in cases where a robot is reaching for one of two cups very close to one another on a table, it might move its arm in a slightly over-exaggerated and not-fully-optimal rightward arc, to make it clear that it’s reaching for the right-hand glass.
“We’re the House Band of Bayesian Statistics”
After the last day of workshops on Sunday, the conference hosted a wrap-up party, complete with live band. “The Imposteriors” is a band made up entirely of stats professors, who played everything from polka to recent pop hits to their own cover of “All About That Bayes”
Because you know we’re
All about that Bayes, ‘bout that Bayes, no Neiman
All about that Bayes, ‘bout that Bayes, no Pearson
All about that Bayes, ‘bout that Bayes, no Fisher
At one point they even brought up David Blei on the accordion! All in all, it made for a delightfully dorky end to the conference.
Other Interesting Papers
- A machine learning approach to learning efficient database indexes
- A new kind of autoencoder that prioritizes learning independent representations
- “Progressive GAN” which tries to solve the problem of keeping the generator and discriminator at the same level by gradually scaling up the difficulty of the generation problem.
- At attempt to learn internal representations that map to human-understandable concepts by using a game rendering engine that expects such independent factors
Quote of the Day(s):
Presenter: “How much time do I have?”
Presenter: “I’ll take the 10”
Tweet of the Day(s):
Mood of the Day(s): “Does Ian Goodfellow secretly own a time turner?”
(he spoke at at least 5 workshops)