This work addresses the classic machine learning problem of online prediction with expert advice. A potential-based framework for the fixed horizon version of this problem was previously developed using verification arguments from optimal control theory (Kobzar, Kohn and Wang, New Potential-Based Bounds for Prediction with Expert Advice (2019)). This paper extends this framework to the random (geometric) stopping version. Taking advantage of these ideas, we construct potentials for the geometric version of prediction with expert advice from potentials used for the fixed horizon version. This construction leads to new explicit lower and upper bounds associated with specific adversary and player strategies for the geometric problem. We identify regimes where these bounds are state of the art.

In this paper we propose to solve an important problem in recommendation -- user cold start, based on meta leaning method. Previous meta learning approaches finetune all parameters for each new user, which is both computing and storage expensive. In contrast, we divide model parameters into fixed and adaptive parts and develop a two-stage meta learning algorithm to learn them separately. The fixed part, capturing user invariant features, is shared by all users and is learned during offline meta learning stage. The adaptive part, capturing user specific features, is learned during online meta learning stage. By decoupling user invariant parameters from user dependent parameters, the proposed approach is more efficient and storage cheaper than previous methods. It also has potential to deal with catastrophic forgetting while continually adapting for streaming coming users. Experiments on production data demonstrates that the proposed method converges faster and to a better performance than baseline methods. Meta-training without online meta model finetuning increases the AUC from 72.24% to 74.72% (2.48% absolute improvement). Online meta training achieves a further gain of 2.46\% absolute improvement comparing with offline meta training.

We lower bound the complexity of finding $\epsilon$-stationary points (with gradient norm at most $\epsilon$) using stochastic first-order methods. In a well-studied model where algorithms access smooth, potentially non-convex functions through queries to an unbiased stochastic gradient oracle with bounded variance, we prove that (in the worst case) any algorithm requires at least $\epsilon^{-4}$ queries to find an $\epsilon$ stationary point. The lower bound is tight, and establishes that stochastic gradient descent is minimax optimal in this model. In a more restrictive model where the noisy gradient estimates satisfy a mean-squared smoothness property, we prove a lower bound of $\epsilon^{-3}$ queries, establishing the optimality of recently proposed variance reduction techniques.

December 5, 2019 Abstract The notion of drift refers to the phenomenon that the distribution, which is underlying the observed data, changes over time. Albeit many attempts were made to deal with drift, formal notions of drift are application-dependent and formulated in various degrees of abstraction and mathematical coherence. In this contribution, we provide a probability theoretical framework, that allows a formalization of drift in continuous time, which subsumes popular notions of drift. It gives rise to a new characterization of drift in terms of stochastic dependency between data and time. This particularly intuitive formalization enables us to design a new, efficient drift detection method. Further, it induces a technology, to decompose observed data into a drifting and a non-drifting part. Keywords: Online learning, learning theory, stochastic processes, learning with drift, continuous time models, drift decomposition 1 INTRODUCTION One fundamental assumption in classical machine learning is the fact that observed data are i.i.d. Yet, this assumption is often violated as soon as machine learning faces real world problems: models are subject to seasonal changes, changed demands of individual costumers, ageing of sensors, etc. In such settings, lifelong model adaptation rather than classical batch learning is required for optimum performance. Since drift, i.e. the fact that data is no longer identically distributed, is a major issue in many real-world applications of machine learning, many attempts were made to deal with this setting (Ditzler et al., 2015). Depending on the domain of data and application, the presence of drift is modelled in different ways. As an example, covariate shift refers to the situation of training and test set having different marginal distributions (Gretton et al., 2009). Learning for data streams extends this setting to an unlimited (but usually countable) stream of observed data, mostly in supervised learning scenarios (Gama et al., 2014). Learning technologies for such situations often rely on windowing techniques, and adapt the model based on the characteristics of the data in an observed time window. Active methods explicitly detect drift, usually referring to drift of the classification error, and trigger model adaptation this way, while passive methods continuously adjust the model (Ditzler et al., 2015).

Learning from Demonstrations (LfD) via Behavior Cloning (BC) works well on multiple complex tasks. However, a limitation of the typical LfD approach is that it requires expert demonstrations for all scenarios, including those in which the algorithm is already well-trained. The recently proposed Learning from Interventions (LfI) overcomes this limitation by using an expert overseer. The expert overseer only intervenes when it suspects that an unsafe action is about to be taken. Although LfI significantly improves over LfD, the state-of-the-art LfI fails to account for delay caused by the expert's reaction time and only learns short-term behavior. We address these limitations by 1) interpolating the expert's interventions back in time, and 2) by splitting the policy into two hierarchical levels, one that generates sub-goals for the future and another that generates actions to reach those desired sub-goals. This sub-goal prediction forces the algorithm to learn long-term behavior while also being robust to the expert's reaction time. Our experiments show that LfI using sub-goals in a hierarchical policy framework trains faster and achieves better asymptotic performance than typical LfD.

This is the supporting wiki for the upcoming book Machine Learning Engineering by Andriy Burkov. This book is distributed on the "read first, buy later" principle. I strongly believe that paying for the content before consuming it is buying a pig in a poke. You can see and try a car in a dealership before you buy it. You can try on a shirt or a dress in a department store.

New tools designed to help institutions meet accessibility requirements stand to personalize learning for all students. Artificial intelligence (AI) has seeped into almost every corner of higher education, popping up in the classroom, administrative offices, and even in dorm rooms and on campus grounds -- all with the promise to streamline tasks and create a more personalized college experience for students. As the technology steers colleges away from a one-size-fits-all approach, it is helping them make progress on one of their most long-running goals: making higher ed more accessible to all types of learners. It is doing that in several ways. Among them, by scanning class materials for accessibility issues, improving learning tools for students with disabilities and offering personalized resources for learners who may need additional support, such as those who speak English as a second language.

What do you think of when you think of schools and colleges? A classroom full of students furiously scribbling down notes while a teacher is droning on about a topic which is "very important for your midterms". Exams are a very important and indispensable part of education. They are important milestones in a student's educational journey, and students are understandably stressed about them. In an academic year, students have to give as many as 12 exams per semester, which means up to 24 exams in one year!

Where the training of machine learning models is concerned, there's always a risk of overfitting -- or corresponding to closely -- to a particular set of data. In point of fact, it's not infeasible that popular machine learning benchmarks like the Arcade Learning Environment encourage overfitting, in that they have a low emphasis on generalization. That's why OpenAI -- the San Francisco-based research firm cofounded by CTO Greg Brockman, chief scientist Ilya Sutskever, and others -- today released the Procgen Benchmark, a set of 16 procedurally-generated environments (CoinRun, StarPilot, CaveFlyer, Dodgeball, FruitBot, Chaser, Miner, Jumper, Leaper, Maze, BigFish, Heist, Climber, Plunder, Ninja, and BossFight) that measure how quickly a model learns generalizable skills. It builds atop the startup's CoinRun toolset, which used procedural generation to construct sets of training and test levels. "We want the best of both worlds: a benchmark comprised of many diverse environments, each of which fundamentally requires generalization," wrote OpenAI in a blog post.