In September 2020, Regina Barzilay was announced as the winner of the inaugural AAAI Squirrel AI award. Regina was formally presented with the prize during an award ceremony at the AAAI2021 conference, following which she delivered an invited talk. She spoke about two particular areas of medicine that she has been researching: drug discovery and cancer diagnosis. It is well-known that the development of drugs is slow and expensive. Currently, drug discovery is primarily experimentally driven, with properties of molecules investigated empirically.

Despite advancements in deep reinforcement learning algorithms, developing an effective exploration strategy is still an open problem. Most existing exploration strategies either are based on simple heuristics, or require the model of the environment, or train additional deep neural networks to generate imagination-augmented paths. In this paper, a revolutionary algorithm, called Policy Augmentation, is introduced. Policy Augmentation is based on a newly developed inductive matrix completion method. The proposed algorithm augments the values of unexplored state-action pairs, helping the agent take actions that will result in high-value returns while the agent is in the early episodes. Training deep reinforcement learning algorithms with high-value rollouts leads to the faster convergence of deep reinforcement learning algorithms. Our experiments show the superior performance of Policy Augmentation. The code can be found at: https://github.com/arashmahyari/PolicyAugmentation.

The automation of probabilistic reasoning is one of the primary aims of machine learning. Recently, the confluence of variational inference and deep learning has led to powerful and flexible automatic inference methods that can be trained by stochastic gradient descent. In particular, normalizing flows are highly parameterized deep models that can fit arbitrarily complex posterior densities. However, normalizing flows struggle in highly structured probabilistic programs as they need to relearn the forward-pass of the program. Automatic structured variational inference (ASVI) remedies this problem by constructing variational programs that embed the forward-pass. Here, we combine the flexibility of normalizing flows and the prior-embedding property of ASVI in a new family of variational programs, which we named cascading flows. A cascading flows program interposes a newly designed highway flow architecture in between the conditional distributions of the prior program such as to steer it toward the observed data. These programs can be constructed automatically from an input probabilistic program and can also be amortized automatically. We evaluate the performance of the new variational programs in a series of structured inference problems. We find that cascading flows have much higher performance than both normalizing flows and ASVI in a large set of structured inference problems.

This graduate textbook on machine learning tells a story of how patterns in data support predictions and consequential actions. Starting with the foundations of decision making, we cover representation, optimization, and generalization as the constituents of supervised learning. A chapter on datasets as benchmarks examines their histories and scientific bases. Self-contained introductions to causality, the practice of causal inference, sequential decision making, and reinforcement learning equip the reader with concepts and tools to reason about actions and their consequences. Throughout, the text discusses historical context and societal impact. We invite readers from all backgrounds; some experience with probability, calculus, and linear algebra suffices.

Deep learning compiler frameworks are gaining ground as a more portable back-end for deep learning applications on increasingly diverse hardware. However, they face the daunting challenge of matching performance offered by hand-tuned target-specific libraries. While auto-tuning frameworks with statistical cost models can provide dynamic and efficient code optimization, they suffer from large space exploration and cost model training overheads. This paper proposes MetaTune, a meta-learning based cost model that more quickly and accurately predicts the performance of optimized codes with pre-trained model parameters. MetaTune encodes convolution kernel codes as structurally similar graphs to facilitate meta-learning, meta-trains a GNN model with a very small input data set, and then predicts optimization parameters for unseen convolution operations with varying sizes and structures during compilation. The resulting framework with MetaTune provides 8 to 13% better inference time on average for four CNN models with comparable or lower optimization time while outperforming transfer learning by 10% in cross-platform cases.

How much credit (or blame) should an action taken in a state get for a future reward? This is the fundamental temporal credit assignment problem in Reinforcement Learning (RL). One of the earliest and still most widely used heuristics is to assign this credit based on a scalar coefficient $\lambda$ (treated as a hyperparameter) raised to the power of the time interval between the state-action and the reward. In this empirical paper, we explore heuristics based on more general pairwise weightings that are functions of the state in which the action was taken, the state at the time of the reward, as well as the time interval between the two. Of course it isn't clear what these pairwise weight functions should be, and because they are too complex to be treated as hyperparameters we develop a metagradient procedure for learning these weight functions during the usual RL training of a policy. Our empirical work shows that it is often possible to learn these pairwise weight functions during learning of the policy to achieve better performance than competing approaches.

In this work, we study auxiliary prediction tasks defined by temporal-difference networks (TD networks); these networks are a language for expressing a rich space of general value function (GVF) prediction targets that may be learned efficiently with TD. Through analysis in an illustrative domain we show the benefits to learning state representations of exploiting the full richness of TD networks, including both action-conditional predictions and temporally deep predictions. Our main (and perhaps surprising) result is that deep action-conditional TD networks with random structures that create random prediction-questions about random features yield state representations that are competitive with state-of-the-art hand-crafted value prediction and pixel control auxiliary tasks in both Atari games and DeepMind Lab tasks. We also show through stop-gradient experiments that learning the state representations solely via these unsupervised random TD network prediction tasks yield agents that outperform the end-to-end-trained actor-critic baseline.

Data augmentation is often used to enlarge datasets with synthetic samples generated in accordance with the underlying data distribution. To enable a wider range of augmentations, we explore negative data augmentation strategies (NDA) that intentionally create out-of-distribution samples. We show that such negative out-of-distribution samples provide information on the support of the data distribution, and can be leveraged for generative modeling and representation learning. We introduce a new GAN training objective where we use NDA as an additional source of synthetic data for the discriminator. We prove that under suitable conditions, optimizing the resulting objective still recovers the true data distribution but can directly bias the generator towards avoiding samples that lack the desired structure. Empirically, models trained with our method achieve improved conditional/unconditional image generation along with improved anomaly detection capabilities. Further, we incorporate the same negative data augmentation strategy in a contrastive learning framework for self-supervised representation learning on images and videos, achieving improved performance on downstream image classification, object detection, and action recognition tasks. These results suggest that prior knowledge on what does not constitute valid data is an effective form of weak supervision across a range of unsupervised learning tasks. Data augmentation strategies for synthesizing new data in a way that is consistent with an underlying task are extremely effective in both supervised and unsupervised learning (Oord et al., 2018; Zhang et al., 2016; Noroozi & Favaro, 2016; Asano et al., 2019).

In recent years, reinforcement learning has seen interest because of deep Q-Learning, where the model is a convolutional neural network. Deep Q-Learning has shown promising results in games such as Atari and AlphaGo. Instead of learning the entire Q-table, it learns an estimate of the Q function that determines a state's policy action. We use Q-Learning and deep Q-learning, to learn control policies of four constraint satisfaction games (15-Puzzle, Minesweeper, 2048, and Sudoku). 15-Puzzle is a sliding permutation puzzle and provides a challenge in addressing its large state space. Minesweeper and Sudoku involve partially observable states and guessing. 2048 is also a sliding puzzle but allows for easier state representation (compared to 15-Puzzle) and uses interesting reward shaping to solve the game. These games offer unique insights into the potential and limits of reinforcement learning. The Q agent is trained with no rules of the game, with only the reward corresponding to each state's action. Our unique contribution is in choosing the reward structure, state representation, and formulation of the deep neural network. For low shuffle, 15-Puzzle, achieves a 100% win rate, the medium and high shuffle achieve about 43% and 22% win rates respectively. On a standard 16x16 Minesweeper board, both low and high-density boards achieve close to 45% win rate, whereas medium density boards have a low win rate of 15%. For 2048, the 1024 win rate was achieved with significant ease (100%) with high win rates for 2048, 4096, 8192 and 16384 as 40%, 0.05%, 0.01% and 0.004% , respectively. The easy Sudoku games had a win rate of 7%, while medium and hard games had 2.1% and 1.2% win rates, respectively. This paper explores the environment complexity and behavior of a subset of constraint games using reward structures which can get us closer to understanding how humans learn.

We first exhibit a multimodal image registration task, for which a neural network trained on a dataset with noisy labels reaches almost perfect accuracy, far beyond noise variance. This surprising auto-denoising phenomenon can be explained as a noise averaging effect over the labels of similar input examples. This effect theoretically grows with the number of similar examples; the question is then to define and estimate the similarity of examples. We express a proper definition of similarity, from the neural network perspective, i.e. we quantify how undissociable two inputs $A$ and $B$ are, taking a machine learning viewpoint: how much a parameter variation designed to change the output for $A$ would impact the output for $B$ as well? We study the mathematical properties of this similarity measure, and show how to use it on a trained network to estimate sample density, in low complexity, enabling new types of statistical analysis for neural networks. We analyze data by retrieving samples perceived as similar by the network, and are able to quantify the denoising effect without requiring true labels. We also propose, during training, to enforce that examples known to be similar should also be seen as similar by the network, and notice speed-up training effects for certain datasets.