In June 2020, a new and powerful artificial intelligence (AI) began dazzling technologists in Silicon Valley. Called GPT-3 and created by the research firm OpenAI in San Francisco, California, it was the latest and most powerful in a series of'large language models': AIs that generate fluent streams of text after imbibing billions of words from books, articles and websites. GPT-3 had been trained on around 200 billion words, at an estimated cost of tens of millions of dollars. The developers who were invited to try out GPT-3 were astonished. "I have to say I'm blown away," wrote Arram Sabeti, founder of a technology start-up who is based in Silicon Valley. "It's far more coherent than any AI language system I've ever tried. All you have to do is write a prompt and it'll add text it thinks would plausibly follow. I've gotten it to write songs, stories, press releases, guitar tabs, interviews, essays, technical manuals. I feel like I've seen the future."

This article is part 2 of my Popular Machine Learning Interview questions. Here I feature more questions I usually see asked during interviews. I shall note that this isn't an interview prep guide nor a conclusive list of all questions. Rather, you should use this article as a refresher for your Machine Learning knowledge. I suggest reading the question then try to answer it yourself before reading the answer.

Current deep reinforcement learning (RL) algorithms are still highly task-specific and lack the ability to generalize to new environments. Lifelong learning (LLL), however, aims at solving multiple tasks sequentially by efficiently transferring and using knowledge between tasks. Despite a surge of interest in lifelong RL in recent years, the lack of a realistic testbed makes robust evaluation of LLL algorithms difficult. Multi-agent RL (MARL), on the other hand, can be seen as a natural scenario for lifelong RL due to its inherent non-stationarity, since the agents' policies change over time. In this work, we introduce a multi-agent lifelong learning testbed that supports both zero-shot and few-shot settings. Our setup is based on Hanabi -- a partially-observable, fully cooperative multi-agent game that has been shown to be challenging for zero-shot coordination. Its large strategy space makes it a desirable environment for lifelong RL tasks. We evaluate several recent MARL methods, and benchmark state-of-the-art LLL algorithms in limited memory and computation regimes to shed light on their strengths and weaknesses. This continual learning paradigm also provides us with a pragmatic way of going beyond centralized training which is the most commonly used training protocol in MARL. We empirically show that the agents trained in our setup are able to coordinate well with unseen agents, without any additional assumptions made by previous works.

Naively trained neural networks tend to experience catastrophic forgetting in sequential task settings, where data from previous tasks are unavailable. A number of methods, using various model expansion strategies, have been proposed recently as possible solutions. However, determining how much to expand the model is left to the practitioner, and often a constant schedule is chosen for simplicity, regardless of how complex the incoming task is. Instead, we propose a principled Bayesian nonparametric approach based on the Indian Buffet Process (IBP) prior, letting the data determine how much to expand the model complexity. We pair this with a factorization of the neural network's weight matrices. Such an approach allows the number of factors of each weight matrix to scale with the complexity of the task, while the IBP prior encourages sparse weight factor selection and factor reuse, promoting positive knowledge transfer between tasks. We demonstrate the effectiveness of our method on a number of continual learning benchmarks and analyze how weight factors are allocated and reused throughout the training.

Speech Affect Recognition is a problem of extracting emotional affects from audio data. Low resource languages corpora are rear and affect recognition is a difficult task in cross-corpus settings. We present an approach in which the model is trained on high resource language and fine-tune to recognize affects in low resource language. We train the model in same corpus setting on SAVEE, EMOVO, Urdu, and IEMOCAP by achieving baseline accuracy of 60.45, 68.05, 80.34, and 56.58 percent respectively. For capturing the diversity of affects in languages cross-corpus evaluations are discussed in detail. We find that accuracy improves by adding the domain target data into the training data. Finally, we show that performance is improved for low resource language speech affect recognition by achieving the UAR OF 69.32 and 68.2 for Urdu and Italian speech affects.

Unsupervised domain adaptation aims to learn a task classifier that performs well on the unlabeled target domain, by utilizing the labeled source domain. Inspiring results have been acquired by learning domain-invariant deep features via domain-adversarial training. However, its parallel design of task and domain classifiers limits the ability to achieve a finer category-level domain alignment. To promote categorical domain adaptation (CatDA), based on a joint category-domain classifier, we propose novel losses of adversarial training at both domain and category levels. Since the joint classifier can be regarded as a concatenation of individual task classifiers respectively for the two domains, our design principle is to enforce consistency of category predictions between the two task classifiers. Moreover, we propose a concept of vicinal domains whose instances are produced by a convex combination of pairs of instances respectively from the two domains. Intuitively, alignment of the possibly infinite number of vicinal domains enhances that of original domains. We propose novel adversarial losses for vicinal domain adaptation (VicDA) based on CatDA, leading to Vicinal and Categorical Domain Adaptation (ViCatDA). We also propose Target Discriminative Structure Recovery (TDSR) to recover the intrinsic target discrimination damaged by adversarial feature alignment. We also analyze the principles underlying the ability of our key designs to align the joint distributions. Extensive experiments on several benchmark datasets demonstrate that we achieve the new state of the art.

This paper presents a machine learning framework (GP-NODE) for Bayesian systems identification from partial, noisy and irregular observations of nonlinear dynamical systems. The proposed method takes advantage of recent developments in differentiable programming to propagate gradient information through ordinary differential equation solvers and perform Bayesian inference with respect to unknown model parameters using Hamiltonian Monte Carlo sampling and Gaussian Process priors over the observed system states. This allows us to exploit temporal correlations in the observed data, and efficiently infer posterior distributions over plausible models with quantified uncertainty. Moreover, the use of sparsity-promoting priors such as the Finnish Horseshoe for free model parameters enables the discovery of interpretable and parsimonious representations for the underlying latent dynamics. A series of numerical studies is presented to demonstrate the effectiveness of the proposed GP-NODE method including predator-prey systems, systems biology, and a 50-dimensional human motion dynamical system. Taken together, our findings put forth a novel, flexible and robust workflow for data-driven model discovery under uncertainty. All code and data accompanying this manuscript are available online at \url{https://github.com/PredictiveIntelligenceLab/GP-NODEs}.

Ensemble-based modifications of the well-known SHapley Additive exPlanations (SHAP) method for the local explanation of a black-box model are proposed. The modifications aim to simplify SHAP which is computationally expensive when there is a large number of features. The main idea behind the proposed modifications is to approximate SHAP by an ensemble of SHAPs with a smaller number of features. According to the first modification, called ER-SHAP, several features are randomly selected many times from the feature set, and Shapley values for the features are computed by means of "small" SHAPs. The explanation results are averaged to get the final Shapley values. According to the second modification, called ERW-SHAP, several points are generated around the explained instance for diversity purposes, and results of their explanation are combined with weights depending on distances between points and the explained instance. The third modification, called ER-SHAP-RF, uses the random forest for preliminary explanation of instances and determining a feature probability distribution which is applied to selection of features in the ensemble-based procedure of ER-SHAP. Many numerical experiments illustrating the proposed modifications demonstrate their efficiency and properties for local explanation.

We develop a new algorithm for non-convex stochastic optimization that finds an $\epsilon$-critical point in the optimal $O(\epsilon^{-3})$ stochastic gradient and hessian-vector product computations. Our algorithm uses Hessian-vector products to "correct" a bias term in the momentum of SGD with momentum. This leads to better gradient estimates in a manner analogous to variance reduction methods. In contrast to prior work, we do not require excessively large batch sizes (or indeed any restrictions at all on the batch size), and both our algorithm and its analysis are much simpler. We validate our results on a variety of large-scale deep learning benchmarks and architectures, where we see improvements over SGD and Adam.

Random feature methods have been successful in various machine learning tasks, are easy to compute, and come with theoretical accuracy bounds. They serve as an alternative approach to standard neural networks since they can represent similar function spaces without a costly training phase. However, for accuracy, random feature methods require more measurements than trainable parameters, limiting their use for data-scarce applications or problems in scientific machine learning. This paper introduces the sparse random feature method that learns parsimonious random feature models utilizing techniques from compressive sensing. We provide uniform bounds on the approximation error for functions in a reproducing kernel Hilbert space depending on the number of samples and the distribution of features. The error bounds improve with additional structural conditions, such as coordinate sparsity, compact clusters of the spectrum, or rapid spectral decay. We show that the sparse random feature method outperforms shallow networks for well-structured functions and applications to scientific machine learning tasks.