This week's top Machine Learning stories, including robots to drive your car, diagnose your medical images, pick up your mess, and more! Machine Learning is one of the most exciting fields in the world. Every week we discover something new, something amazing, something revolutionary. It's incredible, but it can also be overwhelming. That's why we created This Week in Machine Learning!

A student in Ashok Goel's class last semester had a question: How long could the computer programs, or "agents," they were building take to solve problems? Since it was an online course, the student posted the question to the group discussion board. One teaching assistant replied, pointing to a portion of the assignment that set a 15 minute limit. The student clarified that their agent was running a little slow, and could take a bit longer. "It's fine if your agent takes a few minutes to run," she wrote.

One day in January, Eric Wilson dashed off a message to the teaching assistants for an online course at the Georgia Institute of Technology. "I really feel like I missed the mark in giving the correct amount of feedback," he wrote, pleading to revise an assignment. Thirteen minutes later, the TA responded. "Unfortunately, there is not a way to edit submitted feedback," wrote Jill Watson, one of nine assistants for the 300-plus students. Last week, Mr. Wilson found out he had been seeking guidance from a computer.

We study an online learning framework introduced by Mannor and Shamir (2011) in which the feedback is specified by a graph, in a setting where the graph may vary from round to round and is \emph{never fully revealed} to the learner. We show a large gap between the adversarial and the stochastic cases. In the adversarial case, we prove that even for dense feedback graphs, the learner cannot improve upon a trivial regret bound obtained by ignoring any additional feedback besides her own loss. In contrast, in the stochastic case we give an algorithm that achieves $\widetilde \Theta(\sqrt{\alpha T})$ regret over $T$ rounds, provided that the independence numbers of the hidden feedback graphs are at most $\alpha$. We also extend our results to a more general feedback model, in which the learner does not necessarily observe her own loss, and show that, even in simple cases, concealing the feedback graphs might render a learnable problem unlearnable.

Didi currently processes over 11 million trips, plans over 9 billion routes, and collects over 50TB of data per day. Machine learning strategies are vital to the company's success, and with growth comes the need to constantly improve on core algorithms, especially those that impact supply-demand forecasting. The competition is a challenge to machine learning and big data students around the world to improve how the company ensures riders always get a car when and where they need it, and drivers know where to be even before a ride is hailed. Didi has just published the competition data set, and registration closes on June 17 when the first round submission is due. The Top 10 teams will be invited to Didi in July to compete for the top prize.

Students at Georgia Institute of Technology found out that a teacher assistant giving them assistance they was actually Jill an artificially intelligent robot. Jill was created to provide faster answers and feedback to students and take some of the pressure of teaching large classes off the instructors. The class is a core requirement of Georgia Tech's online master's program in computer science, and it tends to draw a lot of questions from students. It's offered every semester, and each time, the 300 or so students enrolled post roughly 10,000 messages in the course's online forums, Goel estimates. That volume has often overwhelmed Goel and his eight teaching assistants, so this time, he added a ninth: Jill.

Excited to talk about Machine Learning, Data Science, and Artificial Intelligence? Ready to discover the key influencers you need to follow to stay current on all the latest happenings in these fields? We've got the resources you need. You see, at Udacity, we talk a great deal about Machine Learning, Data Science, and Artificial Intelligence. We talk to each other, we talk to our students, and we talk to the world at large.

For anyone who wants to learn "Data Science," I want to emphasize that it's such a fluid term. I would recommend starting out by analyzing a dataset in an area that is close to your heart, whether it's real estate pricing data for your neighborhood, or educational curriculum & outcome data. For me, it just happened to be genomic and patient health data, because that was a problem space I felt comfortable playing around in. If you explore the myriad technologies out there to help you find patterns and predict things from that data you're passionate about, before you know it you'll wake up one day and realize that you're actually a data scientist! It's all a matter of learning how to use different tools and when to apply them, and this just comes with practice.

One day in January, Eric Wilson dashed off a message to the teaching assistants for an online course at the Georgia Institute of Technology. "I really feel like I missed the mark in giving the correct amount of feedback," he wrote, pleading to revise an assignment. Thirteen minutes later, the TA responded. "Unfortunately, there is not a way to edit submitted feedback," wrote Jill Watson, one of nine assistants for the 300-plus students. Last week, Mr. Wilson found out he had been seeking guidance from a computer.

This work focuses on dynamic regret of online convex optimization that compares the performance of online learning to a clairvoyant who knows the sequence of loss functions in advance and hence selects the minimizer of the loss function at each step. By assuming that the clairvoyant moves slowly (i.e., the minimizers change slowly), we present several improved variation-based upper bounds of the dynamic regret under the true and noisy gradient feedback, which are {\it optimal} in light of the presented lower bounds. The key to our analysis is to explore a regularity metric that measures the temporal changes in the clairvoyant's minimizers, to which we refer as {\it path variation}. Firstly, we present a general lower bound in terms of the path variation, and then show that under full information or gradient feedback we are able to achieve an optimal dynamic regret. Secondly, we present a lower bound with noisy gradient feedback and then show that we can achieve optimal dynamic regrets under a stochastic gradient feedback and two-point bandit feedback. Moreover, for a sequence of smooth loss functions that admit a small variation in the gradients, our dynamic regret under the two-point bandit feedback matches what is achieved with full information.