NeurIPS 2023 论文列表

While most machine learning models can provide confidence in their predictions, confidence is insufficient to understand a prediction's reliability. For instance, the model may have a low confidence prediction if the input is not well-represented in the training dataset or if the input is inherently ambiguous. In this work, we investigate the relationship between how atypical~(rare) a sample or a class is and the reliability of a model's predictions. We first demonstrate that atypicality is strongly related to miscalibration and accuracy. In particular, we empirically show that predictions for atypical inputs or atypical classes are more overconfident and have lower accuracy. Using these insights, we show incorporating atypicality improves uncertainty quantification and model performance for discriminative neural networks and large language models. In a case study, we show that using atypicality improves the performance of a skin lesion classifier across different skin tone groups without having access to the group attributes. Overall, we propose that models should use not only confidence but also atypicality to improve uncertainty quantification and performance. Our results demonstrate that simple post-hoc atypicality estimators can provide significant value.

The quest for human imitative AI has been an enduring topic in AI research since inception. The technical evolution and emerging capabilities of the latest cohort of large language models (LLMs) have reinvigorated the subject beyond academia to cultural zeitgeist. While recent NLP evaluation benchmark tasks test some aspects of human-imitative behaviour (e.g., BIG-bench's `human-like behavior' tasks), few, if not none, examine *creative problem solving* abilities. Creative problem solving in humans is a well-studied topic in cognitive neuroscience with standardized tests that predominantly use ability to associate (heterogeneous) connections among clue words as a metric for creativity. Exposure to misleading stimuli --- distractors dubbed *red herrings* --- impede human performance in such tasks via the *fixation effect* and Einstellung paradigm. In cognitive neuroscience studies, such fixations are experimentally induced by pre-exposing participants to orthographically similar incorrect words to subsequent word-fragments or clues. The popular British quiz show Only Connect's *Connecting Wall* segment essentially mimics Mednick's Remote Associates Test (RAT) formulation with built-in, deliberate red herrings, that makes it an ideal proxy dataset to explore and study fixation effect and Einstellung paradigm from cognitive neuroscience in LLMs. In addition to presenting the novel Only Connect Wall (OCW) dataset, we also report results from our evaluation of selected pre-trained language models and LLMs (including OpenAI's GPT series) on creative problem solving tasks like grouping clue words by heterogeneous connections, and identifying correct open knowledge domain connections in respective groups. We synthetically generate two additional datasets: OCW-Randomized, OCW-WordNet to further analyze our red-herrings hypothesis in language models. The code and link to the dataset is available at [url](https://github.com/TaatiTeam/OCW).

We investigate area convexity [Sherman17], a mysterious tool introduced to tackle optimization problems under the challenging $\ell_\infty$ geometry. We develop a deeper understanding of its relationship with conventional analyses of extragradient methods [Nemirovski04, Nesterov07]. We also give improved solvers for the subproblems required by variants of the [Sherman17] algorithm, designed through the lens of relative smoothness [BBT17, LFN18}. Leveraging these new tools, we give a state-of-the-art first-order algorithm for solving box-simplex games (a primal-dual formulation of $\ell_\infty$ regression) in a $d \times n$ matrix with bounded rows, using $O(\log d \cdot \epsilon^{-1})$ matrix-vector queries. As a consequence, we obtain improved complexities for approximate maximum flow, optimal transport, min-mean-cycle, and other basic combinatorial optimization problems. We also develop a near-linear time algorithm for a matrix generalization of box-simplex games, capturing a family of problems closely related to semidefinite programs recently used as subroutines in robust statistics and numerical linear algebra.

We initiate the study of $\textit{behavioral information design}$ through deep learning. In information design, a $\textit{sender}$ aims to persuade a $\textit{receiver}$ to take certain actions by strategically revealing information. We address scenarios in which the receiver might exhibit different behavior patterns other than the standard Bayesian rational assumption. We propose HAIDNet, a neural-network-based optimization framework for information design that can adapt to multiple representations of human behavior. Through extensive simulation, we show that HAIDNet can not only recover information policies that are near-optimal compared with known analytical solutions, but also can extend to designing information policies for settings that are computationally challenging (e.g., when there are multiple receivers) or for settings where there are no known solutions in general (e.g., when the receiver behavior does not follow the Bayesian rational assumption). We also conduct real-world human-subject experiments and demonstrate that our framework can capture human behavior from data and lead to more effective information policy for real-world human receivers.

The success of deep learning relies heavily on large and diverse datasets, but the storage, preprocessing, and training of such data present significant challenges. To address these challenges, dataset distillation techniques have been proposed to obtain smaller synthetic datasets that capture the essential information of the originals. In this paper, we introduce a Sparse Parameterization for Epitomic datasEt Distillation (SPEED) framework, which leverages the concept of dictionary learning and sparse coding to distill epitomes that represent pivotal information of the dataset. SPEED prioritizes proper parameterization of the synthetic dataset and introduces techniques to capture spatial redundancy within and between synthetic images. We propose Spatial-Agnostic Epitomic Tokens (SAETs) and Sparse Coding Matrices (SCMs) to efficiently represent and select significant features. Additionally, we build a Feature-Recurrent Network (FReeNet) to generate hierarchical features with high compression and storage efficiency. Experimental results demonstrate the superiority of SPEED in handling high-resolution datasets, achieving state-of-the-art performance on multiple benchmarks and downstream applications. Our framework is compatible with a variety of dataset matching approaches, generally enhancing their performance. This work highlights the importance of proper parameterization in epitomic dataset distillation and opens avenues for efficient representation learning. Source code is available at https://github.com/MIV-XJTU/SPEED.

Large language models have been shown to struggle with multi-step reasoning, and do not retain previous reasoning steps for future use. We propose a simple method for solving both of these problems by allowing the model to take Self-Notes. Unlike recent chain-of-thought or scratchpad approaches, the model can deviate from the input context at any time to explicitly think and write down its thoughts. This allows the model to perform reasoning on the fly as it reads the context and even integrate previous reasoning steps, thus enhancing its memory with useful information and enabling multi-step reasoning. Experiments across a wide variety of tasks demonstrate that our method can outperform chain-of-thought and scratchpad methods by taking Self-Notes that interleave the input text.

Neuroprostheses show potential in restoring lost sensory function and enhancing human capabilities, but the sensations produced by current devices often seem unnatural or distorted. Exact placement of implants and differences in individual perception lead to significant variations in stimulus response, making personalized stimulus optimization a key challenge. Bayesian optimization could be used to optimize patient-specific stimulation parameters with limited noisy observations, but is not feasible for high-dimensional stimuli. Alternatively, deep learning models can optimize stimulus encoding strategies, but typically assume perfect knowledge of patient-specific variations. Here we propose a novel, practically feasible approach that overcomes both of these fundamental limitations. First, a deep encoder network is trained to produce optimal stimuli for any individual patient by inverting a forward model mapping electrical stimuli to visual percepts. Second, a preferential Bayesian optimization strategy utilizes this encoder to learn the optimal patient-specific parameters for a new patient, using a minimal number of pairwise comparisons between candidate stimuli. We demonstrate the viability of this approach on a novel, state-of-the-art visual prosthesis model. Our approach quickly learns a personalized stimulus encoder and leads to dramatic improvements in the quality of restored vision, outperforming existing encoding strategies. Further, this approach is robust to noisy patient feedback and misspecifications in the underlying forward model. Overall, our results suggest that combining the strengths of deep learning and Bayesian optimization could significantly improve the perceptual experience of patients fitted with visual prostheses and may prove a viable solution for a range of neuroprosthetic technologies

We consider a weakly supervised learning scenario where the supervision signal is generated by a transition function $\sigma$ of labels associated with multiple input instances. We formulate this problem as *multi-instance Partial Label Learning (multi-instance PLL)*, which is an extension to the standard PLL problem. Our problem is met in different fields, including latent structural learning and neuro-symbolic integration. Despite the existence of many learning techniques, limited theoretical analysis has been dedicated to this problem. In this paper, we provide the first theoretical study of multi-instance PLL with possibly an unknown transition $\sigma$. Our main contributions are as follows: First, we proposed a necessary and sufficient condition for the learnability of the problem. This condition nontrivially generalizes and relaxes the existing *small ambiguity degree* in PLL literature since we allow the transition to be deterministic. Second, we derived Rademacher-style error bounds based on the top-$k$ surrogate loss that is widely used in the neuro-symbolic literature. Furthermore, we conclude with empirical experiments for learning with an unknown transition. The empirical results align with our theoretical findings; however, they also expose the issue of scalability in the weak supervision literature.

Watermarking the outputs of generative models is a crucial technique for tracing copyright and preventing potential harm from AI-generated content. In this paper, we introduce a novel technique called Tree-Ring Watermarking that robustly fingerprints diffusion model outputs. Unlike existing methods that perform post-hoc modifications to images after sampling, Tree-Ring Watermarking subtly influences the entire sampling process, resulting in a model fingerprint that is invisible to humans. The watermark embeds a pattern into the initial noise vector used for sampling. These patterns are structured in Fourier space so that they are invariant to convolutions, crops, dilations, flips, and rotations. After image generation, the watermark signal is detected by inverting the diffusion process to retrieve the noise vector, which is then checked for the embedded signal. We demonstrate that this technique can be easily applied to arbitrary diffusion models, including text-conditioned Stable Diffusion, as a plug-in with negligible loss in FID. Our watermark is semantically hidden in the image space and is far more robust than watermarking alternatives that are currently deployed.

Neural networks with wide layers have attracted significant attention due to their equivalence to Gaussian processes, enabling perfect fitting of training data while maintaining generalization performance, known as benign overfitting. However, existing results mainly focus on shallow or finite-depth networks, necessitating a comprehensive analysis of wide neural networks with infinite-depth layers, such as neural ordinary differential equations (ODEs) and deep equilibrium models (DEQs). In this paper, we specifically investigate the deep equilibrium model (DEQ), an infinite-depth neural network with shared weight matrices across layers. Our analysis reveals that as the width of DEQ layers approaches infinity, it converges to a Gaussian process, establishing what is known as the Neural Network and Gaussian Process (NNGP) correspondence. Remarkably, this convergence holds even when the limits of depth and width are interchanged, which is not observed in typical infinite-depth Multilayer Perceptron (MLP) networks. Furthermore, we demonstrate that the associated Gaussian vector remains non-degenerate for any pairwise distinct input data, ensuring a strictly positive smallest eigenvalue of the corresponding kernel matrix using the NNGP kernel. These findings serve as fundamental elements for studying the training and generalization of DEQs, laying the groundwork for future research in this area.

Meta-learning has been proven an effective learning paradigm for training machine learning models with good generalization ability. Apart from the common practice of uniformly sampling the meta-training tasks, existing methods working on task scheduling strategies are mainly based on pre-defined sampling protocols or the assumed task-model correlations, and greedily make scheduling decisions, which can lead to sub-optimal performance bottlenecks of the meta-model. In this paper, we propose a novel task scheduling framework under Contextual Bandits settings, named BASS, which directly optimizes the task scheduling strategy based on the status of the meta-model. By balancing the exploitation and exploration in meta-learning task scheduling, BASS can help tackle the challenge of limited knowledge about the task distribution during the early stage of meta-training, while simultaneously exploring potential benefits for forthcoming meta-training iterations through an adaptive exploration strategy. Theoretical analysis and extensive experiments are presented to show the effectiveness of our proposed framework.

Understanding the fundamental principles behind the success of deep neural networks is one of the most important open questions in the current literature. To this end, we study the training problem of deep neural networks and introduce an analytic approach to unveil hidden convexity in the optimization landscape. We consider a deep parallel ReLU network architecture, which also includes standard deep networks and ResNets as its special cases. We then show that pathwise regularized training problems can be represented as an exact convex optimization problem. We further prove that the equivalent convex problem is regularized via a group sparsity inducing norm. Thus, a path regularized parallel ReLU network can be viewed as a parsimonious convex model in high dimensions. More importantly, since the original training problem may not be trainable in polynomial-time, we propose an approximate algorithm with a fully polynomial-time complexity in all data dimensions. Then, we prove strong global optimality guarantees for this algorithm. We also provide experiments corroborating our theory.

In this study, we propose a novel deep spatio-temporal point process model, Deep Kernel Mixture Point Processes (DKMPP), that incorporates multimodal covariate information. DKMPP is an enhanced version of Deep Mixture Point Processes (DMPP), which uses a more flexible deep kernel to model complex relationships between events and covariate data, improving the model's expressiveness. To address the intractable training procedure of DKMPP due to the non-integrable deep kernel, we utilize an integration-free method based on score matching, and further improve efficiency by adopting a scalable denoising score matching method. Our experiments demonstrate that DKMPP and its corresponding score-based estimators outperform baseline models, showcasing the advantages of incorporating covariate information, utilizing a deep kernel, and employing score-based estimators.

Existing deep learning-based computer vision methods usually operate in the spatial and frequency domains, which are two orthogonal \textbf{individual} perspectives for image processing. In this paper, we introduce a new spatial-frequency analysis tool, Fractional Fourier Transform (FRFT), to provide comprehensive \textbf{unified} spatial-frequency perspectives. The FRFT is a unified continuous spatial-frequency transform that simultaneously reflects an image's spatial and frequency representations, making it optimal for processing non-stationary image signals. We explore the properties of the FRFT for image processing and present a fast implementation of the 2D FRFT, which facilitates its widespread use. Based on these explorations, we introduce a simple yet effective operator, Multi-order FRactional Fourier Convolution (MFRFC), which exhibits the remarkable merits of processing images from more perspectives in the spatial-frequency plane. Our proposed MFRFC is a general and basic operator that can be easily integrated into various tasks for performance improvement. We experimentally evaluate the MFRFC on various computer vision tasks, including object detection, image classification, guided super-resolution, denoising, dehazing, deraining, and low-light enhancement. Our proposed MFRFC consistently outperforms baseline methods by significant margins across all tasks.

We study the problem of Federated Learning (FL) under client subsampling and data heterogeneity with an objective function that has potentially unbounded smoothness. This problem is motivated by empirical evidence that the class of relaxed smooth functions, where the Lipschitz constant of the gradient scales linearly with the gradient norm, closely resembles the loss functions of certain neural networks such as recurrent neural networks (RNNs) with possibly exploding gradient. We introduce EPISODE++, the first algorithm to solve this problem. It maintains historical statistics for each client to construct control variates and decide clipping behavior for sampled clients in the current round. We prove that EPISODE++ achieves linear speedup in the number of participating clients, reduced communication rounds, and resilience to data heterogeneity. Our upper bound proof relies on novel techniques of recursively bounding the client updates under unbounded smoothness and client subsampling, together with a refined high probability analysis. In addition, we prove a lower bound showing that the convergence rate of a special case of clipped minibatch SGD (without randomness in the stochastic gradient and with randomness in client subsampling) suffers from an explicit dependence on the maximum gradient norm of the objective in a sublevel set, which may be large. This effectively demonstrates that applying gradient clipping to minibatch SGD in our setting does not eliminate the problem of exploding gradients. Our lower bound is based on new constructions of hard instances tailored to client subsampling and a novel analysis of the trajectory of the algorithm in the presence of clipping. Lastly, we provide an experimental evaluation of EPISODE++ when training RNNs on federated text classification tasks, demonstrating that EPISODE++ outperforms strong baselines in FL. The code is available at https://github.com/MingruiLiu-ML-Lab/episode_plusplus.

In modern deep learning, algorithmic choices (such as width, depth, and learning rate) are known to modulate nuanced resource tradeoffs. This work investigates how these complexities necessarily arise for feature learning in the presence of computational-statistical gaps. We begin by considering offline sparse parity learning, a supervised classification problem which admits a statistical query lower bound for gradient-based training of a multilayer perceptron. This lower bound can be interpreted as a *multi-resource tradeoff frontier*: successful learning can only occur if one is sufficiently rich (large model), knowledgeable (large dataset), patient (many training iterations), or lucky (many random guesses). We show, theoretically and experimentally, that sparse initialization and increasing network width yield significant improvements in sample efficiency in this setting. Here, width plays the role of parallel search: it amplifies the probability of finding "lottery ticket" neurons, which learn sparse features more sample-efficiently. Finally, we show that the synthetic sparse parity task can be useful as a proxy for real problems requiring axis-aligned feature learning. We demonstrate improved sample efficiency on tabular classification benchmarks by using wide, sparsely-initialized MLP models; these networks sometimes outperform tuned random forests.

Constructing AI models that respond to text instructions is challenging, especially for sequential decision-making tasks. This work introduces a methodology, inspired by unCLIP, for instruction-tuning generative models of behavior without relying on a large dataset of instruction-labeled trajectories. Using this methodology, we create an instruction-tuned Video Pretraining (VPT) model called STEVE-1, which can follow short-horizon open-ended text and visual instructions in Minecraft. STEVE-1 is trained in two steps: adapting the pretrained VPT model to follow commands in MineCLIP's latent space, then training a prior to predict latent codes from text. This allows us to finetune VPT through self-supervised behavioral cloning and hindsight relabeling, reducing the need for costly human text annotations, and all for only $60 of compute. By leveraging pretrained models like VPT and MineCLIP and employing best practices from text-conditioned image generation, STEVE-1 sets a new bar for open-ended instruction following in Minecraft with low-level controls (mouse and keyboard) and raw pixel inputs, far outperforming previous baselines and robustly completing 12 of 13 tasks in our early-game evaluation suite. We provide experimental evidence highlighting key factors for downstream performance, including pretraining, classifier-free guidance, and data scaling. All resources, including our model weights, training scripts, and evaluation tools are made available for further research.

Neuron interpretation offers valuable insights into how knowledge is structured within a deep neural network model. While a number of neuron interpretation methods have been proposed in the literature, the field lacks a comprehensive comparison among these methods. This gap hampers progress due to the absence of standardized metrics and benchmarks. The commonly used evaluation metric has limitations, and creating ground truth annotations for neurons is impractical. Addressing these challenges, we propose an evaluation framework based on voting theory. Our hypothesis posits that neurons consistently identified by different methods carry more significant information. We rigorously assess our framework across a diverse array of neuron interpretation methods. Notable findings include: i) despite the theoretical differences among the methods, neuron ranking methods share over 60% of their rankings when identifying salient neurons, ii) the neuron interpretation methods are most sensitive to the last layer representations, iii) Probeless neuron ranking emerges as the most consistent method.

We study the problem of learning general (i.e., not necessarily homogeneous) halfspaces with Random Classification Noise under the Gaussian distribution. We establish nearly-matching algorithmic and Statistical Query (SQ) lower bound results revealing a surprising information-computation gap for this basic problem. Specifically, the sample complexity of this learning problem is $\widetilde{\Theta}(d/\epsilon)$, where $d$ is the dimension and $\epsilon$ is the excess error. Our positive result is a computationally efficient learning algorithm with sample complexity $\tilde{O}(d/\epsilon + d/\max(p, \epsilon))^2)$, where $p$ quantifies the bias of the target halfspace. On the lower bound side, we show that any efficient SQ algorithm (or low-degree test) for the problem requires sample complexity at least $\Omega(d^{1/2}/(\max(p, \epsilon))^2)$. Our lower bound suggests that this quadratic dependence on $1/\epsilon$ is inherent for efficient algorithms.