Bayesian optimization, active learning, and experimentation

One of the most challenging parts of solving real optimization tasks is translating what the human actually wants into a suitable objective function. We show that LLMs can solve this problem by getting feedback from the human on current results, and adjusting the objective accordingly.

We show how Gaussian process models can jointly fit data from different modalities, such as Likert-scale ratings and paired preferences, by mixing likelihoods. Specific applications include understanding haptic perception and preferences.

We combine short-term and long-term A/B tests to run fast optimizations of long-term outcomes, which we use for improving the efficiency of Meta ranking systems.

We develop active learning methods that target functions of derivatives of Gaussian processes, and use them for global sensitivity analysis and understanding parameter importance in tuning problems.

Standard Gaussian process regression models are ruined by outliers, common in hyperparameter optimization tasks. We developed a robust model that identifies and downweights outliers.

We develop models for preference learning over complex output spaces like images.

When asking humans their preferences or perception, we can measure reaction times alongside their response, which provides implicit information about how hard the decision is. We develop models that combine reaction times with preferences, and use them for psychophysics and preference optimization.

Interpretability is of vital importance in real systems. We developed methods for optimizing for sparse solutions, and understsanding the quality/interpretability trade-off, which we use for simplifying the sourcing component of a ranking system.

There is a large and old literature in psychology on measuring perceptual thresholds. We can combine parametric models from psychology with nonparametric Gaussian Process models to get the best of both worlds and better understand visual perception.

Active learning from preference and perception data is made challenging by the Bernoulli responses. We derive new posterior update formulae for the level-set estimation problem that allow better understanding human visual perception.

A thorough study of meta-learning and transfer learning for hyperparameter optimization.

There are subtle reasons why linear embeddings often fail for high-dimensional optimization, which we explore and correct for the purpose of learning robot locomotion policies.

BoTorch, our software and framework for using differentiable programming for Bayesian optimization.

Tuning contextual policies with A/B tests produces a challenging high-dimensional optimization problem, however the problem has structure that we can take advantage of to successfully tune video adaptive bitrate policies.

Ranking systems can be simulated with offline reply, but generally there will be bias. We show that these biased offline estimates can be combined with online A/B tests for multi-task Bayesian optimization. We used this approach to successfully improve the News Feed ranking model.

Methodological advances for using Bayesian optimization with noisy responses, particularly A/B tests. The paper also describes two systems optimizations that we've done with this approach: tuning a production ranking system, and optimizing web server compiler flags.

We meta-learning to accelerate hyperparameter optimization from similar past runs, and use it for hyperparameter optimization in Facebook's computer vision platform.

In biology we run wet-lab experiments to identify parameters of dynamical systems. How do we decide if we have run enough experiments, and if not, which additional experiments to run? We use minimax optimization to answer these questions, and use it to understand HIV dynamics.

Forecasting

Forecasting is a common data science task, yet also a specialized skill outside the expertise of many data scientists. The Prophet forecasting package is designed to be flexible enough to handle a range of business time series, while still being configurable by non-experts. We developed it for a collection of important forecasting tasks at Facebook, and have since open sourced it. It is built on Stan and has R and Python versions. It has received a tremendous response, filling a real need for forecasting these types of time series.

Modeling with sales transaction data

We developed a Bayesian hierarchical model for demand modeling in the presence of stockouts, which we use to infer cookie demand at a local bakery.

We use a copula model to model correlations in demand between various items, and optimally price bundle offers. This work was done during my summer internship at IBM Research.

Predictions from rules and information retrieval

We developed a Bayesian approach for fitting interpretable decision list models and used it to create a simple system for predicting stroke risk in atrial fibrillation patients.

We use techniques from supervised ranking to learn rule-based models that make predictions on data that are sequentially revealed.

The ECML PKDD Discovery Challenge 2013 was a competition to develop a recommender system for baby names. I developed a recommender system based on association rules and collaborative filtering. Despite its simplicity and a lack of feature engineering, I won 3rd place for the offline challenge. A major advantage of my approach was that it was fast enough that recommendations could be made entirely online, and I won 2nd place in an online challenge, where my recommender system was actually implemented into the baby naming website.

An information retrieval algorithm for finding and ranking a list of similar items given a small seed.

We developed a learning theory framework for using association rules for classification and sequential event prediction, with both VC bounds and results from algorithmic stability.

An earlier conference version of the paper above on theoretical bounds for association rule models in sequential prediction problems.

Psychophysics, biology and neuroscience

Understanding how long it takes for visual and auditory stimuli to produce responses in sensory cortices helps to elucidate neural processing pathways. This is follow-up work to the 2010 study below.

At Meta, psychophysics models are used for setting AR and VR hardware requirements for AR and VR. Nonparametric modeling and active learning are part of the framework for determining the requirements.

Our open-source toolbox AEPsych enables surrogate model-based active learning for threshold identification and other common psychophysics experimentation tasks.

I provided computational and statistical expertise for this study of the dynamics of HIV inhibition by interferon.

Measuring the timing of brain activations is made challenging by noise. We use robust statistics to better map cross-sensory interactions from MEG data.

While working at The Martinos Center for Biomedical Imaging at Mass. General Hospital I was involved in a study of cross-sensory interactions (auditory stimuli activate visual cortex and vice versa) using MEG and fMRI.

Machine learning-themed videos

	AISTATS paper reaction | Mega Compilation!! Paper reaction videos are the latest trend sweeping AI and ML. Efforts are underway to entirely replace the peer review process with the number of likes and subscribes on reaction videos. This video is a compilation of just a few of the reaction videos for a recent AISTATS paper. A lot of excitement! It's gonna be great!
	If AI papers were pharmaceuticals (a parody commercial) Do you suffer from high-dimensional Bayesian optimization? Alebo can help! Watch this commercial to understand how Alebo can make your life happier, your food tastier, and your friends funnier.
	Stockouts and substitutions at Lando's Bakery When KDD required me to make a video describing our paper on stockouts and substitutions at a bakery, I decided it was the perfect opportunity to try making a stop motion video, something I had always wanted to do. It was my first but not my last, and it was every bit as fun as I imagined it would be. Perhaps not quite what the conference organizers had in mind, but I think it got the ideas across pretty well!

Data science posts

	Debate bingo: 2020 presidential edition The second edition of Debate Bingo! I once again extracted commonly used phrases from each candidate to create a bingo card generator, for use during the debates.
	In Defense of Fahrenheit On a scale of 0 to 100, how warm is it? I combine Census data with daily temperatures from nearly 4000 weather stations to estimate the distribution of temperatures experienced by people in the U.S. The oft-maligned Fahrenheit scale nearly perfectly captures this distribution and is a great temperature scale for U.S. weather.
	Debate bingo: 2016 presidential edition Presidential candidates tend to repeat themselves, and our two for 2016 are no exception. I extracted some of their most commonly used phrases and used them to create a bingo card generator, for use during the debates.
	The best part of a lightsaber duel is the talking Why are the original Star Wars lightsaber duels so much better than those from the Prequels? It's not the simple fight choreography, it's the talking. In this post I quantify the balance between talking and fighting in Star Wars lightsaber duels and show that there is a clear difference between the Original Trilogy and the Prequels. Unfortunately, The Force Awakens is a typical Prequels duel.
	Was 2015 Boston's worst winter yet? When I finally had some time after spending the entire month of February shoveling snow, I did an analysis of snowfall data to determine exactly how unusual of a winter it was. Although 2015 beat the all-time total snowfall record by only a couple inches, a deeper look at the data revealed that this was by far the worst winter in recorded history.

Technical expositions

On NP-completeness

The theory of NP-completeness has its roots in a foundational result by Cook, who showed that Boolean satisfiability (SAT) is NP-complete and thus unlikely to admit an efficient solution. In this short paper I prove an analogous result using binary integer programming in the place of SAT. The proof is notationally cleaner and more straightforward to those for whom the language of integer programming is more natural than that of SAT.

On Bayesian Analysis

In this work I provide a somewhat rigorous yet simultaneously informal introduction to Bayesian analysis. It covers everything from the theory of posterior asymptotics to practical considerations of MCMC sampling. It assumes the reader is comfortable with basic probability and some mathematical rigor.

About me