Should sequences be modeled autoregressively---one symbol at a time? How much computation is needed to predict the next symbol? While local normalization is cheap, this also limits its power. We point out that some probability distributions over discrete sequences cannot be well-approximated by any autoregressive model whose runtime and parameter size grow polynomially in the sequence length---even though their unnormalized sequence probabilities are efficient to compute exactly. Intuitively, the probability of the next symbol can be expensive to compute or approximate (even via randomized algorithms) when it marginalizes over exponentially many possible futures, which is in general $\mathrm{NP}$-hard. Our result is conditional on the widely believed hypothesis that $\mathrm{NP} \nsubseteq \mathrm{P/poly}$ (without which the polynomial hierarchy would collapse at the second level). This theoretical observation serves as a caution to the viewpoint that pumping up parameter size is a straightforward way to improve autoregressive models (e.g., in language modeling). It also suggests that globally normalized (energy-based) models may sometimes outperform locally normalized (autoregressive) models, as we demonstrate experimentally for language modeling.

Large scale recommender models find most relevant items from huge catalogs, and they play a critical role in modern search and recommendation systems. To model the input space with large-vocab categorical features, a typical recommender model learns a joint embedding space through neural networks for both queries and items from user feedback data. However, with millions to billions of items, the power-law user feedback makes labels very sparse for a large amount of long-tail items. Inspired by the recent success in self-supervised representation learning research in both computer vision and natural language understanding, we propose a multi-task self-supervised learning (SSL) framework for large-scale item recommendations. The framework is designed to tackle the label sparsity problem by learning more robust item representations. Furthermore, we propose two self-supervised tasks applicable to models with categorical features within the proposed framework: (i) Feature Masking (FM) and (ii) Feature Dropout (FD). We evaluate our framework using two large-scale datasets with 500M and 1B training examples respectively. Our results demonstrate that the proposed framework outperforms traditional supervised learning only models and state-of-the-art regularization techniques in the context of item recommendations. The SSL framework shows larger improvement with less supervision compared to the counterparts. We also apply the proposed techniques to a web-scale commercial app-to-app recommendation system, and significantly improve top-tier business metrics via A/B experiments on live traffic. Our online results also verify our hypothesis that our framework indeed improves model performance on slices that lack supervision.

Prioritized Experience Replay (PER) is a deep reinforcement learning technique in which agents learn from transitions sampled with non-uniform probability proportionate to their temporal-difference error. We show that any loss function evaluated with non-uniformly sampled data can be transformed into another uniformly sampled loss function with the same expected gradient. Surprisingly, we find in some environments PER can be replaced entirely by this new loss function without impact to empirical performance. Furthermore, this relationship suggests a new branch of improvements to PER by correcting its uniformly sampled loss function equivalent. We demonstrate the effectiveness of our proposed modifications to PER and the equivalent loss function in several MuJoCo and Atari environments.

We provide new statistical guarantees for transfer learning via representation learning--when transfer is achieved by learning a feature representation shared across different tasks. This enables learning on new tasks using far less data than is required to learn them in isolation. Formally, we consider $t+1$ tasks parameterized by functions of the form $f_j \circ h$ in a general function class $\mathcal{F} \circ \mathcal{H}$, where each $f_j$ is a task-specific function in $\mathcal{F}$ and $h$ is the shared representation in $\mathcal{H}$. Letting $C(\cdot)$ denote the complexity measure of the function class, we show that for diverse training tasks (1) the sample complexity needed to learn the shared representation across the first $t$ training tasks scales as $C(\mathcal{H}) + t C(\mathcal{F})$, despite no explicit access to a signal from the feature representation and (2) with an accurate estimate of the representation, the sample complexity needed to learn a new task scales only with $C(\mathcal{F})$. Our results depend upon a new general notion of task diversity--applicable to models with general tasks, features, and losses--as well as a novel chain rule for Gaussian complexities. Finally, we exhibit the utility of our general framework in several models of importance in the literature.

Federated Learning (FL) is a machine learning setting where many devices collaboratively train a machine learning model while keeping the training data decentralized. In most of the current training schemes the central model is refined by averaging the parameters of the server model and the updated parameters from the client side. However, directly averaging model parameters is only possible if all models have the same structure and size, which could be a restrictive constraint in many scenarios. In this work we investigate more powerful and more flexible aggregation schemes for FL. Specifically, we propose ensemble distillation for model fusion, i.e. training the central classifier through unlabeled data on the outputs of the models from the clients. This knowledge distillation technique mitigates privacy risk and cost to the same extent as the baseline FL algorithms, but allows flexible aggregation over heterogeneous client models that can differ e.g. in size, numerical precision or structure. We show in extensive empirical experiments on various CV/NLP datasets (CIFAR-10/100, ImageNet, AG News, SST2) and settings (heterogeneous models/data) that the server model can be trained much faster, requiring fewer communication rounds than any existing FL technique so far.

Continual Learning has inspired a plethora of approaches and evaluation settings; however, the majority of them overlooks the properties of a practical scenario, where the data stream cannot be shaped as a sequence of tasks and offline training is not viable. We work towards General Continual Learning (GCL), where task boundaries blur and the domain and class distributions shift either gradually or suddenly. We address it through mixing rehearsal with knowledge distillation and regularization; our simple baseline, Dark Experience Replay, matches the network's logits sampled throughout the optimization trajectory, thus promoting consistency with its past. By conducting an extensive analysis on both standard benchmarks and a novel GCL evaluation setting (MNIST-360), we show that such a seemingly simple baseline outperforms consolidated approaches and leverages limited resources. We further explore the generalization capabilities of our objective, showing its regularization being beneficial beyond mere performance.

We propose a simple method to generate large amounts of multilingual question and answer pairs by a single generative model. These synthetic samples are then applied to augment the available gold multilingual ones to improve the performance of multilingual QA models on target languages. Our approach only requires existence of automatically translated samples from English to the target domain, thus removing the need for human annotations in the target languages. Experimental results show our proposed approach achieves significant gains in a number of multilingual datasets.

This paper shows how to construct knowledge graphs (KGs) from pre-trained language models (e.g., BERT, GPT-2/3), without human supervision. Popular KGs (e.g, Wikidata, NELL) are built in either a supervised or semi-supervised manner, requiring humans to create knowledge. Recent deep language models automatically acquire knowledge from large-scale corpora via pre-training. The stored knowledge has enabled the language models to improve downstream NLP tasks, e.g., answering questions, and writing code and articles. In this paper, we propose an unsupervised method to cast the knowledge contained within language models into KGs. We show that KGs are constructed with a single forward pass of the pre-trained language models (without fine-tuning) over the corpora. We demonstrate the quality of the constructed KGs by comparing to two KGs (Wikidata, TAC KBP) created by humans. Our KGs also provide open factual knowledge that is new in the existing KGs. Our code and KGs will be made publicly available.

Intelligent agents rely heavily on prior experience when learning a new task, yet most modern reinforcement learning (RL) approaches learn every task from scratch. One approach for leveraging prior knowledge is to transfer skills learned on prior tasks to the new task. However, as the amount of prior experience increases, the number of transferable skills grows too, making it challenging to explore the full set of available skills during downstream learning. Yet, intuitively, not all skills should be explored with equal probability; for example information about the current state can hint which skills are promising to explore. In this work, we propose to implement this intuition by learning a prior over skills. We propose a deep latent variable model that jointly learns an embedding space of skills and the skill prior from offline agent experience. We then extend common maximum-entropy RL approaches to use skill priors to guide downstream learning. We validate our approach, SPiRL (Skill-Prior RL), on complex navigation and robotic manipulation tasks and show that learned skill priors are essential for effective skill transfer from rich datasets. Videos and code are available at https://clvrai.com/spirl.

Optical aerial images change detection is an important task in earth observation and has been extensively investigated in the past few decades. Generally, the supervised change detection methods with superior performance require a large amount of labeled training data which is obtained by manual annotation with high cost. In this paper, we present a novel unsupervised self-training algorithm (USTA) for optical aerial images change detection. The traditional method such as change vector analysis is used to generate the pseudo labels. We use these pseudo labels to train a well designed convolutional neural network. The network is used as a teacher to classify the original multitemporal images to generate another set of pseudo labels. Then two set of pseudo labels are used to jointly train a student network with the same structure as the teacher. The final change detection result can be obtained by the trained student network. Besides, we design an image filter to control the usage of change information in the pseudo labels in the training process of the network. The whole process of the algorithm is an unsupervised process without manually marked labels. Experimental results on the real datasets demonstrate competitive performance of our proposed method.