NeurIPS 2023 Papers

We propose the first online quantum algorithm for zero-sum games with $\widetilde O(1)$ regret under the game setting. Moreover, our quantum algorithm computes an $\varepsilon$-approximate Nash equilibrium of an $m \times n$ matrix zero-sum game in quantum time $\widetilde O(\sqrt{m+n}/\varepsilon^{2.5})$. Our algorithm uses standard quantum inputs and generates classical outputs with succinct descriptions, facilitating end-to-end applications. Technically, our online quantum algorithm "quantizes" classical algorithms based on the optimistic multiplicative weight update method. At the heart of our algorithm is a fast quantum multi-sampling procedure for the Gibbs sampling problem, which may be of independent interest.

Modern neural language models that are widely used in various NLP tasks risk memorizing sensitive information from their training data. Understanding this memorization is important in real world applications and also from a learning-theoretical perspective. An open question in previous studies of language model memorization is how to filter out ``common'' memorization. In fact, most memorization criteria strongly correlate with the number of occurrences in the training set, capturing memorized familiar phrases, public knowledge, templated texts, or other repeated data. We formulate a notion of counterfactual memorization which characterizes how a model's predictions change if a particular document is omitted during training. We identify and study counterfactually-memorized training examples in standard text datasets. We estimate the influence of each memorized training example on the validation set and on generated texts, showing how this can provide direct evidence of the source of memorization at test time.

The tremendous success of large models trained on extensive datasets demonstrates that scale is a key ingredient in achieving superior results. Therefore, the reflection on the rationality of designing knowledge distillation (KD) approaches for limited-capacity architectures solely based on small-scale datasets is now deemed imperative. In this paper, we identify the small data pitfall that presents in previous KD methods, which results in the underestimation of the power of vanilla KD framework on large-scale datasets such as ImageNet-1K. Specifically, we show that employing stronger data augmentation techniques and using larger datasets can directly decrease the gap between vanilla KD and other meticulously designed KD variants. This highlights the necessity of designing and evaluating KD approaches in the context of practical scenarios, casting off the limitations of small-scale datasets. Our investigation of the vanilla KD and its variants in more complex schemes, including stronger training strategies and different model capacities, demonstrates that vanilla KD is elegantly simple but astonishingly effective in large-scale scenarios. Without bells and whistles, we obtain state-of-the-art ResNet-50, ViT-S, and ConvNeXtV2-T models for ImageNet, which achieve 83.1%, 84.3%, and 85.0% top-1 accuracy, respectively. PyTorch code and checkpoints can be found at https://github.com/Hao840/vanillaKD.

For some hypothesis classes and input distributions, \emph{active} agnostic learning needs exponentially fewer samples than passive learning; for other classes and distributions, it offers little to no improvement. The most popular algorithms for agnostic active learning express their performance in terms of a parameter called the disagreement coefficient, but it is known that these algorithms are inefficient on some inputs. We take a different approach to agnostic active learning, getting an algorithm that is \emph{competitive} with the optimal algorithm for any binary hypothesis class $H$ and distribution $\mathcal{D}_X$ over $X$. In particular, if any algorithm can use $m^*$ queries to get $O(\eta)$ error, then our algorithm uses $O(m^* \log H)$ queries to get $O(\eta)$ error. Our algorithm lies in the vein of the splitting-based approach of Dasgupta [2004], which gets a similar result for the realizable ($\eta = 0$) setting. We also show that it is NP-hard to do better than our algorithm's $O(\log H)$ overhead in general.

Understanding when neural networks can be learned efficiently is a fundamental question in learning theory. Existing hardness results suggest that assumptions on both the input distribution and the network's weights are necessary for obtaining efficient algorithms. Moreover, it was previously shown that depth-$2$ networks can be efficiently learned under the assumptions that the input distribution is Gaussian, and the weight matrix is non-degenerate. In this work, we study whether such assumptions may suffice for learning deeper networks and prove negative results. We show that learning depth-$3$ ReLU networks under the Gaussian input distribution is hard even in the smoothed-analysis framework, where a random noise is added to the network's parameters. It implies that learning depth-$3$ ReLU networks under the Gaussian distribution is hard even if the weight matrices are non-degenerate. Moreover, we consider depth-$2$ networks, and show hardness of learning in the smoothed-analysis framework, where both the network parameters and the input distribution are smoothed. Our hardness results are under a well-studied assumption on the existence of local pseudorandom generators.

For each of $T$ time steps, $m$ experts report probability distributions over $n$ outcomes; we wish to learn to aggregate these forecasts in a way that attains a no-regret guarantee. We focus on the fundamental and practical aggregation method known as *logarithmic pooling* -- a weighted average of log odds -- which is in a certain sense the optimal choice of pooling method if one is interested in minimizing log loss (as we take to be our loss function). We consider the problem of learning the best set of parameters (i.e. expert weights) in an online adversarial setting. We assume (by necessity) that the adversarial choices of outcomes and forecasts are consistent, in the sense that experts report calibrated forecasts. Imposing this constraint creates a (to our knowledge) novel semi-adversarial setting in which the adversary retains a large amount of flexibility. In this setting, we present an algorithm based on online mirror descent that learns expert weights in a way that attains $O(\sqrt{T} \log T)$ expected regret as compared with the best weights in hindsight.

Offline multi-agent reinforcement learning is challenging due to the coupling effect of both distribution shift issue common in offline setting and the high dimension issue common in multi-agent setting, making the action out-of-distribution (OOD) and value overestimation phenomenon excessively severe. To mitigate this problem, we propose a novel multi-agent offline RL algorithm, named CounterFactual Conservative Q-Learning (CFCQL) to conduct conservative value estimation. Rather than regarding all the agents as a high dimensional single one and directly applying single agent conservative methods to it, CFCQL calculates conservative regularization for each agent separately in a counterfactual way and then linearly combines them to realize an overall conservative value estimation. We prove that it still enjoys the underestimation property and the performance guarantee as those single agent conservative methods do, but the induced regularization and safe policy improvement bound are independent of the agent number, which is therefore theoretically superior to the direct treatment referred to above, especially when the agent number is large. We further conduct experiments on four environments including both discrete and continuous action settings on both existing and our man-made datasets, demonstrating that CFCQL outperforms existing methods on most datasets and even with a remarkable margin on some of them.

Although BERT-style encoder models are heavily used in NLP research, many researchers do not pretrain their own BERTs from scratch due to the high cost of training. In the past half-decade since BERT first rose to prominence, many advances have been made with other transformer architectures and training configurations that have yet to be systematically incorporated into BERT. Here, we introduce MosaicBERT, a BERT-style encoder architecture and training recipe that is empirically optimized for fast pretraining. This efficient architecture incorporates FlashAttention, Attention with Linear Biases (ALiBi), Gated Linear Units (GLU), a module to dynamically remove padded tokens, and low precision LayerNorm into the classic transformer encoder block. The training recipe includes a 30% masking ratio for the Masked Language Modeling (MLM) objective, bfloat16 precision, and vocabulary size optimized for GPU throughput, in addition to best-practices from RoBERTa and other encoder models. When pretrained from scratch on the C4 dataset, this base model achieves a downstream average GLUE (dev) score of 79.6 in 1.13 hours on 8 A100 80 GB GPUs at a cost of roughly $20. We plot extensive accuracy vs. pretraining speed Pareto curves and show that MosaicBERT base and large are consistently Pareto optimal when compared to a competitive BERT base and large. This empirical speed up in pretraining enables researchers and engineers to pretrain custom BERT-style models at low cost instead of finetune on existing generic models. We open source our model weights and code.

In this paper we tackle the problem of Generalized Category Discovery (GCD). Specifically, given a dataset with labelled and unlabelled images, the task is to cluster all images in the unlabelled subset, whether or not they belong to the labelled categories. Our first contribution is to recognise that most existing GCD benchmarks only contain labels for a single clustering of the data, making it difficult to ascertain whether models are leveraging the available labels to solve the GCD task, or simply solving an unsupervised clustering problem. As such, we present a synthetic dataset, named 'Clevr-4', for category discovery. Clevr-4 contains four equally valid partitions of the data, i.e based on object 'shape', 'texture' or 'color' or 'count'. To solve the task, models are required to extrapolate the taxonomy specified by labelled set, rather than simply latch onto a single natural grouping of the data. We use this dataset to demonstrate the limitations of unsupervised clustering in the GCD setting, showing that even very strong unsupervised models fail on Clevr-4. We further use Clevr-4 to examine the weaknesses of existing GCD algorithms, and propose a new method which addresses these shortcomings, leveraging consistent findings from the representation learning literature to do so. Our simple solution, which is based on `Mean Teachers' and termed $\mu$GCD, substantially outperforms implemented baselines on Clevr-4. Finally, when we transfer these findings to real data on the challenging Semantic Shift Benchmark suite, we find that $\mu$GCD outperforms all prior work, setting a new state-of-the-art.

We present an alternative approach to decompose non-negative tensors, called many-body approximation. Traditional decomposition methods assume low-rankness in the representation, resulting in difficulties in global optimization and target rank selection. We avoid these problems by energy-based modeling of tensors, where a tensor and its mode correspond to a probability distribution and a random variable, respectively. Our model can be globally optimized in terms of the KL divergence minimization by taking the interaction between variables (that is, modes), into account that can be tuned more intuitively than ranks. Furthermore, we visualize interactions between modes as tensor networks and reveal a nontrivial relationship between many-body approximation and low-rank approximation. We demonstrate the effectiveness of our approach in tensor completion and approximation.

We consider the problem of inferring latent stochastic differential equations (SDEs) with a time and memory cost that scales independently with the amount of data, the total length of the time series, and the stiffness of the approximate differential equations. This is in stark contrast to typical methods for inferring latent differential equations which, despite their constant memory cost, have a time complexity that is heavily dependent on the stiffness of the approximate differential equation. We achieve this computational advancement by removing the need to solve differential equations when approximating gradients using a novel amortization strategy coupled with a recently derived reparametrization of expectations under linear SDEs. We show that, in practice, this allows us to achieve similar performance to methods based on adjoint sensitivities with more than an order of magnitude fewer evaluations of the model in training.

With the emerging Tiny Machine Learning (TinyML) inference applications, there is a growing interest when deploying TinyML models on the low-power Microcontroller Unit (MCU). However, deploying TinyML models on MCUs reveals several challenges due to the MCU’s resource constraints, such as small flash memory, tight SRAM memory budget, and slow CPU performance. Unlike typical layer-wise inference, patch-based inference reduces the peak usage of SRAM memory on MCUs by saving small patches rather than the entire tensor in the SRAM memory. However, the processing of patch-based inference tremendously increases the amount of MACs against the layer-wise method. Thus, this notoriously computational overhead makes patch-based inference undesirable on MCUs. This work designs StreamNet that employs the stream buffer to eliminate the redundant computation of patch-based inference. StreamNet uses 1D and 2D streaming processing and provides an parameter selection algorithm that automatically improve the performance of patch-based inference with minimal requirements on the MCU’s SRAM memory space. In 10 TinyML models, StreamNet-2D achieves a geometric mean of 7.3X speedup and saves 81\% of MACs over the state-of-the-art patch-based inference.

Greedy layer-wise or module-wise training of neural networks is compelling in constrained and on-device settings where memory is limited, as it circumvents a number of problems of end-to-end back-propagation. However, it suffers from a stagnation problem, whereby early layers overfit and deeper layers stop increasing the test accuracy after a certain depth. We propose to solve this issue by introducing a simple module-wise regularization inspired by the minimizing movement scheme for gradient flows in distribution space. We call the method TRGL for Transport Regularized Greedy Learning and study it theoretically, proving that it leads to greedy modules that are regular and that progressively solve the task. Experimentally, we show improved accuracy of module-wise training of various architectures such as ResNets, Transformers and VGG, when our regularization is added, superior to that of other module-wise training methods and often to end-to-end training, with as much as 60% less memory usage.

As an efficient alternative to conventional full fine-tuning, parameter-efficient fine-tuning (PEFT) is becoming the prevailing method to adapt pretrained language models. In PEFT, a lightweight module is learned on each dataset while the underlying pretrained language model remains unchanged, resulting in multiple compact modules representing diverse skills when applied to various domains and tasks. In this paper, we propose to compose these parameter-efficient modules through linear arithmetic operations in the weight space, thereby integrating different module capabilities. Specifically, we first define an addition and negation operator for the module, and then further compose these two basic operators to perform flexible arithmetic. Our approach requires no additional training and enables highly flexible module composition. We apply different arithmetic operations to compose the parameter-efficient modules for (1) distribution generalization, (2) multi-tasking, (3) detoxifying, and (4) domain transfer. Additionally, we extend our approach to detoxify Alpaca-LoRA, the latest instruction-tuned large language model based on LLaMA. Empirical results demonstrate that our approach produces new and effective parameter-efficient modules that significantly outperform existing ones across all settings.

Although pairwise causal relations have been extensively studied in observational longitudinal analyses across many disciplines, incorporating knowledge of causal pairs into deep learning models for temporal event sequences remains largely unexplored. In this paper, we propose a novel approach for enhancing the performance of transformer-based models in multivariate event sequences by injecting pairwise qualitative causal knowledge such as `event Z amplifies future occurrences of event Y'. We establish a new framework for causal inference in temporal event sequences using a transformer architecture, providing a theoretical justification for our approach, and show how to obtain unbiased estimates of the proposed measure. Experimental results demonstrate that our approach outperforms several state-of-the-art models in terms of prediction accuracy by effectively leveraging knowledge about causal pairs. We also consider a unique application where we extract knowledge around sequences of societal events by generating them from a large language model, and demonstrate how a causal knowledge graph can help with event prediction in such sequences. Overall, our framework offers a practical means of improving the performance of transformer-based models in multivariate event sequences by explicitly exploiting pairwise causal information.

Many common types of data can be represented as functions that map coordinates to signal values, such as pixel locations to RGB values in the case of an image. Based on this view, data can be compressed by overfitting a compact neural network to its functional representation and then encoding the network weights. However, most current solutions for this are inefficient, as quantization to low-bit precision substantially degrades the reconstruction quality. To address this issue, we propose overfitting variational Bayesian neural networks to the data and compressing an approximate posterior weight sample using relative entropy coding instead of quantizing and entropy coding it. This strategy enables direct optimization of the rate-distortion performance by minimizing the $\beta$-ELBO, and target different rate-distortion trade-offs for a given network architecture by adjusting $\beta$. Moreover, we introduce an iterative algorithm for learning prior weight distributions and employ a progressive refinement process for the variational posterior that significantly enhances performance. Experiments show that our method achieves strong performance on image and audio compression while retaining simplicity.

Remote sensing images present unique challenges to image analysis due to the extensive geographic coverage, hardware limitations, and misaligned multi-scale images. This paper revisits the classical multi-scale representation learning prob- lem but under the general framework of self-supervised learning for remote sensing image understanding. We present Cross-Scale MAE, a self-supervised model built upon the Masked Auto-Encoder (MAE). During pre-training, Cross-Scale MAE employs scale augmentation techniques and enforces cross-scale consistency constraints through both contrastive and generative losses to ensure consistent and meaningful representations well-suited for a wide range of downstream tasks. Further, our implementation leverages the xFormers library to accelerate network pre-training on a single GPU while maintaining the quality of learned represen- tations. Experimental evaluations demonstrate that Cross-Scale MAE exhibits superior performance compared to standard MAE and other state-of-the-art remote sensing MAE methods.

The growth of Graph Convolution Network (GCN) model sizes has revolutionized numerous applications, surpassing human performance in areas such as personal healthcare and financial systems. The deployment of GCNs in the cloud raises privacy concerns due to potential adversarial attacks on client data. To address security concerns, Privacy-Preserving Machine Learning (PPML) using Homomorphic Encryption (HE) secures sensitive client data. However, it introduces substantial computational overhead in practical applications. To tackle those challenges, we present LinGCN, a framework designed to reduce multiplication depth and optimize the performance of HE based GCN inference. LinGCN is structured around three key elements: (1) A differentiable structural linearization algorithm, complemented by a parameterized discrete indicator function, co-trained with model weights to meet the optimization goal. This strategy promotes fine-grained node-level non-linear location selection, resulting in a model with minimized multiplication depth. (2) A compact node-wise polynomial replacement policy with a second-order trainable activation function, steered towards superior convergence by a two-level distillation approach from an all-ReLU based teacher model. (3) an enhanced HE solution that enables finer-grained operator fusion for node-wise activation functions, further reducing multiplication level consumption in HE-based inference. Our experiments on the NTU-XVIEW skeleton joint dataset reveal that LinGCN excels in latency, accuracy, and scalability for homomorphically encrypted inference, outperforming solutions such as CryptoGCN. Remarkably, LinGCN achieves a 14.2× latency speedup relative to CryptoGCN, while preserving an inference accuracy of ~75\% and notably reducing multiplication depth. Additionally, LinGCN proves scalable for larger models, delivering a substantial 85.78\% accuracy with 6371s latency, a 10.47\% accuracy improvement over CryptoGCN.

The capability of generalizing to out-of-distribution data is crucial for the deployment of machine learning models in the real world. Existing domain generalization (DG) mainly embarks on offline and discrete scenarios, where multiple source domains are simultaneously accessible and the distribution shift among domains is abrupt and violent. Nevertheless, such setting may not be universally applicable to all real-world applications, as there are cases where the data distribution gradually changes over time due to various factors, e.g., the process of aging. Additionally, as the domain constantly evolves, new domains will continually emerge. Re-training and updating models with both new and previous domains using existing DG methods can be resource-intensive and inefficient. Therefore, in this paper, we present a problem formulation for Continual Domain Generalization over Temporal Drift (CDGTD). CDGTD addresses the challenge of gradually shifting data distributions over time, where domains arrive sequentially and models can only access the data of the current domain. The goal is to generalize to unseen domains that are not too far into the future. To this end, we propose an Evolving Standardization (EvoS) method, which characterizes the evolving pattern of feature distribution and mitigates the distribution shift by standardizing features with generated statistics of corresponding domain. Specifically, inspired by the powerful ability of transformers to model sequence relations, we design a multi-scale attention module (MSAM) to learn the evolving pattern under sliding time windows of different lengths. MSAM can generate statistics of current domain based on the statistics of previous domains and the learned evolving pattern. Experiments on multiple real-world datasets including images and texts validate the efficacy of our EvoS.