Microsoft Research  recently had Warren Powell speak on Sequential Decision Analytics: A unified framework

Warren Powell is Professor Emeritus at Princeton University and Chief Analytics Officer of Optimal Dynamics. He’s also Founder and Director of Castle Labs at Princeton which manages over 70 grants and contracts with government agencies and leading companies working to develop models of algorithms, freight logistics, energy systems, and other industries.

He’s created a new field called sequential decision analytics which he covers in this talk and in his new book: Reinforcement Learning and Stochastic Optimization: A unified framework for sequential decisions.

Microsoft Research shares this amazing talk on  the optimization of many deep learning hyperparameters can be formulated as a bilevel optimization problem.

While most black-box and gradient-based approaches require many independent training runs, we aim to adapt hyperparameters online as the network trains. The main challenge is to approximate the response Jacobian, which captures how the minimum of the inner objective changes as the hyperparameters are perturbed. To do this, we introduce the self-tuning network (STN), which fits a hypernetwork to approximate the best response function in the vicinity of the current hyperparameters. Differentiating through the hypernetwork lets us efficiently approximate the gradient of the validation loss with respect to the hyperparameters. We train the hypernetwork and hyperparameters jointly. Empirically, we can find hyperparameter settings competitive with Bayesian Optimization in a single run of training, and in some cases find hyperparameter schedules that outperform any fixed hyperparameter value.

Sustaining growth in storage and computational needs is increasingly challenging thanks to those pesky laws of physics.

For over a decade, exponentially more information has been produced year after year while data storage solutions are pressed to keep up. Soon, current solutions will be unable to match new information in need of storage. Computing is on a similar trajectory, with new needs emerging in search and other domains that require more efficient systems. Innovative methods are necessary to ensure the ability to address future demands, and DNA provides an opportunity at the molecular level for ultra-dense, durable, and sustainable solutions in these areas.

In this webinar, join Microsoft researcher Karin Strauss in exploring the role of biotechnology and synthetic DNA in reaching this goal. Although we have yet to achieve scalable, general-purpose molecular computation, there are areas of IT in which a molecular approach shows growing promise. These areas include storage as well as computation.

Learn how molecules, specifically synthetic DNA, can store digital data and perform certain types of special-purpose computation.

Microsoft Research presents this talk on pretraining large neural language models, such as BERT, has led to impressive gains on many natural language processing (NLP) tasks.  However, most pretraining efforts focus on general-domain corpora, such as in newswire and web text. Biomedical text is very different from general-domain text, yet biomedical NLP has been relatively underexplored.

A prevailing assumption is that even domain-specific pretraining can benefit by starting from general-domain language models.

In this webinar, Microsoft researchers Hoifung Poon, Senior Director of Biomedical NLP, and Jianfeng Gao, Distinguished Scientist, will challenge this assumption by showing that for domains with abundant unlabeled text, such as biomedicine, pretraining language models from scratch results in substantial gains over continual pretraining of general-domain language models.

You will begin with understanding how biomedical text differs from general-domain text and how biomedical NLP poses substantial challenges that are not present in mainstream NLP. You will also learn about the two paradigms for domain-specific language model pretraining and see how pretraining from scratch significantly outperforms mixed-domain pretraining in a wide range of biomedical NLP tasks. Finally, find out about our comprehensive benchmark and leaderboard created specifically for biomedical NLP, called BLURB, and see how our biomedical language model, PubMedBERT, sets a new state of the art.

As people work from home, new opportunities and challenges arise around mobile office work.

On one hand, people may have flexible work hours and may not need to deal with traffic or long commutes. On the other hand, they may need to work at makeshift spaces, with less-than-optimal working conditions while physically separated from co-workers.

Virtual reality (VR) has the potential to change the way we work, whether from home or at the office, and help address some of these new challenges. We envision the future office worker to be able to work productively everywhere, solely using portable standard input devices and immersive head-mounted displays. VR has the potential to enable this by allowing users to create working environments of their choice and by relieving them of physical limitations, such as constrained space or noisy environments.

In this webinar, Microsoft Researcher Eyal Ofek presents a summary of research investigating opportunities and challenges for realizing a mobile VR office environment. In particular, you’ll learn how VR can be mixed with standard off-the-shelf equipment (such as tablets, laptops, or desktops) to enable effective, efficient, and ergonomic mobile knowledge work.

As AI is becoming part of user-facing applications and is directly impacting society, deploying AI reliably and responsibly has become a priority for Microsoft and several other industry leaders. In recent years, Microsoft has developed a set of AI principles and standards alongside a company-wide ecosystem to guide responsible AI development and deployment.

In this webinar, Microsoft researchers Dr. Besmira Nushi and Dr. Ece Kamar will share crucial learnings gained from founding and implementing such principles in practice in a large industry setting, where investments in AI span from automation to enhanced human productivity and augmentation.

Reinforcement learning is one of the most exciting collections of techniques for building self-learning systems.

Over the past five years, we’ve seen RL successfully meet such challenges as exceeding human performance on popular video games and board games.

Despite the excitement around the success of these RL agents, it has remained extraordinarily difficult for most people to get to build their own RL agents.

In this webinar, Microsoft researcher Kristian Holsheimer guides you through the landscape of RL agents and shows you how to build your own custom agent in just a few lines of code and without breaking a sweat.

Optimization is at the heart of machine learning, and gradient computation is central to many optimization techniques. Stochastic optimization, in particular, has taken center stage as the principal method of fitting many models, from deep neural networks to variational Bayesian posterior approximations.

Generally, one uses data subsampling to efficiently construct unbiased gradient estimators for stochastic optimization, but this is only one possibility. In this talk, I discuss two alternative approaches to constructing unbiased gradient estimates in machine learning problems. The first approach uses randomized truncation of objective functions defined as loops or limits. Such objectives arise in settings ranging from hyperparameter selection, to fitting parameters of differential equations, to variational inference using lower bounds on the log-marginal likelihood.

The second approach revisits the Jacobian accumulation problem at the heart of automatic differentiation, observing that it is possible to collapse the linearized computational graph of, e.g., deep neural networks, in a randomized way such that less memory is used but little performance is lost.

Microsoft Research hosts this talk on Automating ML Performance Metric Selection

From music recommendations to high-stakes medical treatment selection, complex decision-making tasks are increasingly automated as classification problems. Thus, there is a growing need for classifiers that accurately reflect complex decision-making goals.

One often formalizes these learning goals via a performance metric, which, in turn, can be used to evaluate and compare classifiers. Yet, choosing the appropriate metric remains a challenging problem. This talk will outline metric elicitation as a formal strategy to address the metric selection problem. Metric elicitation automates the discovery of implicit preferences from an expert or an expert panel using relatively efficient and straightforward interactive queries.

Beyond standard classification settings, I will also outline early work on metric selection for group-fair classification.