If machine learning is to discover knowledge from data, then machine teaching is an inverse problem to pass the knowledge on. More precisely, given a learning algorithm and a target model, the goal of machine teaching is to construct an optimal (e.g. the smallest) training set from which the algorithm will learn the target model. I will discuss several aspects of machine teaching that connect to interactive machine learning.

Wouldn't it be nice if machines could understand content in images and communicate this understanding as effectively as humans? Such technology would be immensely powerful, be it for aiding a visually-impaired user navigate a world built by the sighted, assisting an analyst in extracting relevant information from a surveillance feed, educating a child playing a game on a touch screen, providing information to a spectator at an art gallery, or interacting with a robot. As computer vision and natural language processing techniques are maturing, we are closer to achieving this dream than we have ever been.

In this talk, I will present two ongoing thrusts in my lab that push the boundaries of AI capabilities at the intersection of vision, language, and commonsense reasoning.

Visual Question Answering (VQA): I will describe the task, our dataset (the largest and most complex of its kind), our model, and ongoing work for free-form and open-ended Visual Question Answering (VQA). Given an image and a natural language question about the image (e.g., “What kind of store is this?”, “How many people are waiting in the queue?”, “Is it safe to cross the street?”), the machine’s task is to automatically produce an accurate natural language answer (“bakery”, “5”, “Yes”). We have collected and recently released a dataset containing >250,000 images, >760,000 questions, and ~10 million answers. Our dataset is enabling the next generation of AI systems, often based on deep learning techniques, for understanding images and language, and performing complex reasoning; in our lab and the community at large.

Learning Common Sense Through Visual Abstraction: Common sense is a key ingredient in building intelligent machines that make "human-like" decisions when performing tasks -- be it automatically answering natural language questions, or understanding images and videos. How can machines learn this common sense? While some of this knowledge is explicitly stated in human-generated text (books, articles, blogs, etc.), much of this knowledge is unwritten. While unwritten, it is not unseen! The visual world around us is full of structure bound by commonsense laws. But machines today cannot learn common sense directly by observing our visual world because they cannot accurately perform detailed visual recognition in images and videos. This leads to a chicken-and-egg problem: we would like to learn common sense to allow machines to understand images accurately, but in order to learn common sense, we need accurate image parsing. We argue that the solution is to give up on photorealism. We propose to leverage abstract scenes -- cartoon scenes made from clip art by crowd sourced humans -- to teach our machines common sense.

People understand many domains more deeply than today's machine learning systems. Having a good representation for a problem is crucial to the success of intelligent systems. In this talk, we discuss recent work and future opportunities for how humans can aid machine learning algorithms. Beyond simply labeling data, the crowd can help uncover the latent representation behind a problem. We discuss recent work on eliciting features using active learning as well as other aspects of crowdsourcing and machine learning, such as how crowdsourcing can help generate data, raise questions, and assist in more complex AI tasks.

The scale and complexity of biological systems makes biological research a fertile domain for active learning. This is because for complex biological systems, time and cost constraints make it infeasible to do all possible experiments. Previous applications of active learning in biology have been limited, and can be divided between retrospective studies, the goal of which is just to demonstrate the usefulness of active learning algorithms, and prospective studies, in which active learning is actually used to drive experimentation. The latter ones are rare. Furthermore, past studies mostly considered unidimensional active learning, where only a single variable is explored (i.e., which member of a set of drugs is most active on a single target). The goal was to reduce cost for a study that would have been feasible but expensive if carried out exhaustively. However, most biological systems have multiple, interacting components, and thus require multidimensional active learning (e.g., choose which pairs of drugs and targets to test in order to model possible effects of multiple drugs on multiple targets). This is far more challenging, but the goal is not just to reduce cost but tackle problems not otherwise addressable. In this talk, I will describe both retrospective and prospective applications of multidimensional active learning to biological systems. Considerations discussed will include the choice of the modeling method, incorporation of prior information using similarity matrices, and how to know when a model is good enough to stop doing experiments.

This talk considers a core question in reinforcement learning (RL): How can we tractably solve sequential decision making problems where the learning agent receives rich observations? We begin with a new model called Contextual Decision Processes (CDPs) for studying such problems, and show that it encompasses several prior setups to study RL such as MDPs and POMDPs. Several special cases of CDPs are, however, known to be provably intractable in their sample complexities. To overcome this challenge, we further propose a structural property of such processes, called the Bellman Rank. We find that the Bellman Rank of a CDP (and an associated class of functions) provides an intuitive measure of the hardness of a problem in terms of sample complexity and is small in several practical settings. In particular, we propose an algorithm, whose sample complexity scales with the Bellman Rank of the process, and is completely independent of the size of the observation space of the agent. We also show that our techniques are robust to our modeling assumptions, and make connections to several known results as well as highlight novel consequences of our results. This talk is based on joint work with Nan Jiang, Akshay Krishnamurthy, John Langford and Rob Schapire.

No abstract available.

Many clustering problems in computer vision and other contexts are also classification problems, where each cluster shares a meaningful label. Subspace clustering algorithms in particular are often applied to problems that fit this description, for example with face images or handwritten digits. While it is straightforward to request human input on these datasets, our goal is to reduce this input as much as possible. We present an algorithm for active query selection that allows us to leverage the union of subspace structure assumed in subspace clustering. The central step of the algorithm is in querying points of minimum margin between estimated subspaces; analogous to classifier margin, these lie near the decision boundary. This procedure can be used after any subspace clustering algorithm that outputs an affinity matrix and is capable of driving the clustering error down more quickly than other state-of-the-art active query algorithms on datasets with subspace structure. We demonstrate the effectiveness of our algorithm on several benchmark datasets, and with a modest number of queries we see significant gains in clustering performance.

No abstract available.

No abstract available.

No abstract available.

No abstract available.

We propose a pool-based non-parametric active learning algorithm for general metric spaces, which outputs a nearest-neighbor classifier. We give prediction error guarantees that depend on the noisy-margin properties of the input sample, and are competitive with those obtained by previously proposed passive learners. We prove that the label complexity of  the new algorithm is significantly lower than that of any passive learner with similar error guarantees. Our algorithm is based on a generalized sample compression scheme, and a new label-efficient active model-selection procedure.

Based on joint work with Aryeh Kontorovich and Ruth Urner

In this talk, I will present a new general active learning strategy, yielding significantly better label complexity guarantees than previously known to be possible. In particular, this work reveals that under Tsybakov's noise condition, the minimax optimal label complexity of active learning is significantly smaller than that of traditional (passive) supervised learning, for any hypothesis class of finite VC dimension. This new active learning strategy essentially uses sequential hypothesis tests, modified with an early cut-off stopping criterion, to enhance the noise-robustness of a simpler active learning subroutine. The method is surprisingly simple, and quite general, and also yields near-optimal label complexities in several other settings, such as learning with a smooth regression function or a smooth decision boundary.