When it comes to artificial intelligence (AI), few people can match Andrew Ng's breadth and depth of experience, combining research in a world-leading academic environment, accelerating AI education (one of his courses really helped me when I was breaking into AI myself) with founding AI teams at some of the most successful tech companies. Ng lead the Stanford AI Lab, was a founding leader at Google Brain where he worked with legendary engineer Jeff Dean, lead AI at Baidu, and currently serves as an adjunct professor in Computer Science at Stanford University. Among Andrew's other pursuits: being the founder of deeplearning.ai, In a recent episode of the Intel on AI podcast, Ng and host Intel's Abigail Hing Wen discuss why most of the important work yet to be done with AI is in industries outside of Silicon Valley, such as manufacturing, agriculture, and healthcare. Ng sees AI as driving huge growth for successful adopters, to the point where he sees a need for society to be prepared to offer additional support to workers in disrupted industries.

In a viral TikTok video, a college student showed an unseen way that academic artificial intelligence, produced by companies like ProctorU, can damage lives. However, she explains that her professor gave her a zero because the "Review " software built by ProctorU showed her mouthing one of the questions in its video recording. "I was just re-reading the question, so I could better understand it," she says. AI (artificial intelligence) proctor software, produced by companies like ProctorU, has been increasingly put under the microscope for data privacy and security issues. Proctoring software use has grown considerably, as the COVID-19 pandemic pushes schools and universities to transition more quickly to remote/online learning.

Rhiannon Chavez has always been an animated teacher, and her fourth-grade classroom is her theater. She sings, dances and gesticulates during her lessons, engaging her 23 students at 186th Street Elementary School in Gardena. But distance learning has elevated these tactics to a whole new level. Now, as she teaches into a computer screen from her home in San Pedro, she pours every last bit of energy into making sure that her enthusiasm is infectious enough to span the digital divide. From one of her three computer monitors screens she leads the children in a song about rounding numbers, her hand movements and facial expressions are bigger and more spirited than ever.

We study the problem of balancing effectiveness and efficiency in automated feature selection. After exploring many feature selection methods, we observe a computational dilemma: 1) traditional feature selection is mostly efficient, but difficult to identify the best subset; 2) the emerging reinforced feature selection automatically navigates to the best subset, but is usually inefficient. Can we bridge the gap between effectiveness and efficiency under automation? Motivated by this dilemma, we aim to develop a novel feature space navigation method. In our preliminary work, we leveraged interactive reinforcement learning to accelerate feature selection by external trainer-agent interaction. In this journal version, we propose a novel interactive and closed-loop architecture to simultaneously model interactive reinforcement learning (IRL) and decision tree feedback (DTF). Specifically, IRL is to create an interactive feature selection loop and DTF is to feed structured feature knowledge back to the loop. First, the tree-structured feature hierarchy from decision tree is leveraged to improve state representation. In particular, we represent the selected feature subset as an undirected graph of feature-feature correlations and a directed tree of decision features. We propose a new embedding method capable of empowering graph convolutional network to jointly learn state representation from both the graph and the tree. Second, the tree-structured feature hierarchy is exploited to develop a new reward scheme. In particular, we personalize reward assignment of agents based on decision tree feature importance. In addition, observing agents' actions can be feedback, we devise another reward scheme, to weigh and assign reward based on the feature selected frequency ratio in historical action records. Finally, we present extensive experiments on real-world datasets to show the improved performance.

Data augmentation is a widely used training trick in deep learning to improve the network generalization ability. Despite many encouraging results, several recent studies did point out limitations of the conventional data augmentation scheme in certain scenarios, calling for a better theoretical understanding of data augmentation. In this work, we develop a comprehensive analysis that reveals pros and cons of data augmentation. The main limitation of data augmentation arises from the data bias, i.e. the augmented data distribution can be quite different from the original one. This data bias leads to a suboptimal performance of existing data augmentation methods. To this end, we develop two novel algorithms, termed "AugDrop" and "MixLoss", to correct the data bias in the data augmentation. Our theoretical analysis shows that both algorithms are guaranteed to improve the effect of data augmentation through the bias correction, which is further validated by our empirical studies. Finally, we propose a generic algorithm "WeMix" by combining AugDrop and MixLoss, whose effectiveness is observed from extensive empirical evaluations.

A fundamental challenge in deep learning is that the optimal step sizes for update steps of stochastic gradient descent are unknown. In traditional optimization, line searches are used to determine good step sizes, however, in deep learning, it is too costly to search for good step sizes on the expected empirical loss due to noisy losses. This empirical work shows that it is possible to approximate the expected empirical loss on vertical cross sections for common deep learning tasks considerably cheaply. This is achieved by applying traditional one-dimensional function fitting to measured noisy losses of such cross sections. The step to a minimum of the resulting approximation is then used as step size for the optimization. This approach leads to a robust and straightforward optimization method which performs well across datasets and architectures without the need of hyperparameter tuning.

Stochastic optimization lies at the heart of machine learning, and its cornerstone is stochastic gradient descent (SGD), a method introduced over 60 years ago. The last 8 years have seen an exciting new development: variance reduction (VR) for stochastic optimization methods. These VR methods excel in settings where more than one pass through the training data is allowed, achieving a faster convergence than SGD in theory as well as practice. These speedups underline the surge of interest in VR methods and the fast-growing body of work on this topic. This review covers the key principles and main developments behind VR methods for optimization with finite data sets and is aimed at non-expert readers. We focus mainly on the convex setting, and leave pointers to readers interested in extensions for minimizing non-convex functions.

Federated learning (FL) is a framework for distributed learning of centralized models. In FL, a set of edge devices train a model using their local data, while repeatedly exchanging their trained updates with a central server. This procedure allows tuning a centralized model in a distributed fashion without having the users share their possibly private data. In this paper, we focus on over-the-air (OTA) FL, which has been suggested recently to reduce the communication overhead of FL due to the repeated transmissions of the model updates by a large number of users over the wireless channel. In OTA FL, all users simultaneously transmit their updates as analog signals over a multiple access channel, and the server receives a superposition of the analog transmitted signals. However, this approach results in the channel noise directly affecting the optimization procedure, which may degrade the accuracy of the trained model. We develop a Convergent OTA FL (COTAF) algorithm which enhances the common local stochastic gradient descent (SGD) FL algorithm, introducing precoding at the users and scaling at the server, which gradually mitigates the effect of the noise. We analyze the convergence of COTAF to the loss minimizing model and quantify the effect of a statistically heterogeneous setup, i.e. when the training data of each user obeys a different distribution. Our analysis reveals the ability of COTAF to achieve a convergence rate similar to that achievable over error-free channels. Our simulations demonstrate the improved convergence of COTAF over vanilla OTA local SGD for training using non-synthetic datasets. Furthermore, we numerically show that the precoding induced by COTAF notably improves the convergence rate and the accuracy of models trained via OTA FL.

Empirical Risk Minimization based methods are based on the consistency hypothesis that all data samples are generated i.i.d. However, this hypothesis cannot hold in many real-world applications. Consequently, simply minimizing training loss can lead the model into recklessly absorbing all statistical correlations in the training dataset. It is why a well-trained model may perform unstably in different testing environments. Hence, learning a stable predictor that can simultaneously performs well in all testing environments is important for machine learning tasks. In this work, we study this problem from the perspective of Invariant Risk Minimization. Specifically, we propose a novel Self-supervised Invariant Risk Minimization method based on the fact that the real causality connections between features are consistent no matter how the environment changes. First, we propose a self-supervised invariant representation learning objective function, which aims to learn a stable representation of the consistent causality. Based on that, we further propose a stable predictor training algorithm. This algorithm aims to improve the predictor's stability using the invariant representation learned by using our proposed objective function. We conduct extensive experiments on both synthetic and real-world datasets to show that our proposal outperforms previous state-of-the-art stable learning methods. The code will be released later.

Learning controllable and generalizable representation of multivariate data with desired structural properties remains a fundamental problem in machine learning. In this paper, we present a novel framework for learning generative models with various underlying structures in the latent space. We represent the inductive bias in the form of mask variables to model the dependency structure in the graphical model and extend the theory of multivariate information bottleneck to enforce it. Our model provides a principled approach to learn a set of semantically meaningful latent factors that reflect various types of desired structures like capturing correlation or encoding invariance, while also offering the flexibility to automatically estimate the dependency structure from data. We show that our framework unifies many existing generative models and can be applied to a variety of tasks including multi-modal data modeling, algorithmic fairness, and invariant risk minimization.