论文检索

Deep Learning/Algorithms

Bingrui Li, Jianfei Chen, Jun Zhu

Optimizer states are a major source of memory consumption for training neural networks, limiting the maximum trainable model within given memory budget. Compressing the optimizer states from 32-bit floating points to lower bitwidth is promising to reduce the training memory footprint, while the current lowest achievable bitwidth is 8-bit. In this work, we push optimizer states bitwidth down to 4-bit through a detailed empirical analysis of first and second moments. Specifically, we find that moments have complicated outlier patterns, that current block-wise quantization cannot accurately approximate. We use a smaller block size and propose to utilize both row-wise and column-wise information for better quantization. We further identify a zero point problem of quantizing the second moment, and solve this problem with a linear quantizer that excludes the zero point. Our 4-bit optimizers are evaluated on a wide variety of benchmarks including natural language understanding, machine translation, image classification, and instruction tuning. On all the tasks our optimizers can achieve comparable accuracy with their full-precision counterparts, while enjoying better memory efficiency.

Deep Learning/Theory

Berfin Simsek, Amire Bendjeddou, Wulfram Gerstner, Johanni Brea

Any continuous function $f^*$ can be approximated arbitrarily well by a neural network with sufficiently many neurons $k$. We consider the case when $f^*$ itself is a neural network with one hidden layer and $k$ neurons. Approximating $f^*$ with a neural network with $n< k$ neurons can thus be seen as fitting an under-parameterized "student" network with $n$ neurons to a "teacher" network with $k$ neurons. As the student has fewer neurons than the teacher, it is unclear, whether each of the $n$ student neurons should copy one of the teacher neurons or rather average a group of teacher neurons. For shallow neural networks with erf activation function and for the standard Gaussian input distribution, we prove that "copy-average" configurations are critical points if the teacher's incoming vectors are orthonormal and its outgoing weights are unitary. Moreover, the optimum among such configurations is reached when $n-1$ student neurons each copy one teacher neuron and the $n$-th student neuron averages the remaining $k-n+1$ teacher neurons. For the student network with $n=1$ neuron, we provide additionally a closed-form solution of the non-trivial critical point(s) for commonly used activation functions through solving an equivalent constrained optimization problem. Empirically, we find for the erf activation function that gradient flow converges either to the optimal copy-average critical point or to another point where each student neuron approximately copies a different teacher neuron. Finally, we find similar results for the ReLU activation function, suggesting that the optimal solution of underparameterized networks has a universal structure.

Applications/Physics

Kirill Neklyudov, Jannes Nys, Luca Thiede, Juan Carrasquilla, Qiang Liu, Max Welling, Alireza Makhzani

Solving the quantum many-body Schrödinger equation is a fundamental and challenging problem in the fields of quantum physics, quantum chemistry, and material sciences. One of the common computational approaches to this problem is Quantum Variational Monte Carlo (QVMC), in which ground-state solutions are obtained by minimizing the energy of the system within a restricted family of parameterized wave functions. Deep learning methods partially address the limitations of traditional QVMC by representing a rich family of wave functions in terms of neural networks. However, the optimization objective in QVMC remains notoriously hard to minimize and requires second-order optimization methods such as natural gradient. In this paper, we first reformulate energy functional minimization in the space of Born distributions corresponding to particle-permutation (anti-)symmetric wave functions, rather than the space of wave functions. We then interpret QVMC as the Fisher--Rao gradient flow in this distributional space, followed by a projection step onto the variational manifold. This perspective provides us with a principled framework to derive new QMC algorithms, by endowing the distributional space with better metrics, and following the projected gradient flow induced by those metrics. More specifically, we propose "Wasserstein Quantum Monte Carlo" (WQMC), which uses the gradient flow induced by the Wasserstein metric, rather than the Fisher--Rao metric, and corresponds to *transporting* the probability mass, rather than *teleporting* it. We demonstrate empirically that the dynamics of WQMC results in faster convergence to the ground state of molecular systems.

Deep Learning/Graph Neural Networks

Aseem Baranwal, Kimon Fountoulakis, Aukosh Jagannath

We study the node classification problem on feature-decorated graphs in the sparse setting, i.e., when the expected degree of a node is $O(1)$ in the number of nodes, in the fixed-dimensional asymptotic regime, i.e., the dimension of the feature data is fixed while the number of nodes is large. Such graphs are typically known to be locally tree-like. We introduce a notion of Bayes optimality for node classification tasks, called asymptotic local Bayes optimality, and compute the optimal classifier according to this criterion for a fairly general statistical data model with arbitrary distributions of the node features and edge connectivity. The optimal classifier is implementable using a message-passing graph neural network architecture. We then compute the generalization error of this classifier and compare its performance against existing learning methods theoretically on a well-studied statistical model with naturally identifiable signal-to-noise ratios (SNRs) in the data. We find that the optimal message-passing architecture interpolates between a standard MLP in the regime of low graph signal and a typical convolution in the regime of high graph signal. Furthermore, we prove a corresponding non-asymptotic result.

Khashayar Gatmiry, Zhiyuan Li, Tengyu Ma, Sashank Reddi, Stefanie Jegelka, Ching-Yao Chuang

Recent works on over-parameterized neural networks have shown that the stochasticity in optimizers has the implicit regularization effect of minimizing the sharpness of the loss function (in particular, the trace of its Hessian) over the family zero-loss solutions. More explicit forms of flatness regularization also empirically improve the generalization performance. However, it remains unclear why and when flatness regularization leads to better generalization. This work takes the first step towards understanding the inductive bias of the minimum trace of the Hessian solutions in an important setting: learning deep linear networks from linear measurements, also known as \emph{deep matrix factorization}. We show that with the standard Restricted Isometry Property (RIP) on the measurements, minimizing the trace of Hessian is approximately equivalent to minimizing the Schatten 1-norm of the corresponding end-to-end matrix parameters (i.e., the product of all layer matrices), which in turn leads to better generalization.

Deep Learning/Robustness

Anselm Krainovic, Mahdi Soltanolkotabi, Reinhard Heckel

Deep neural networks provide excellent performance for inverse problems such as denoising. However, neural networks can be sensitive to adversarial or worst-case perturbations. This raises the question of whether such networks can be trained efficiently to be worst-case robust. In this paper, we investigate whether jittering, a simple regularization technique that adds isotropic Gaussian noise during training, is effective for learning worst-case robust estimators for inverse problems. While well studied for prediction in classification tasks, the effectiveness of jittering for inverse problems has not been systematically investigated. In this paper, we present a novel analytical characterization of the optimal $\ell_2$-worst-case robust estimator for linear denoising and show that jittering yields optimal robust denoisers. Furthermore, we examine jittering empirically via training deep neural networks (U-nets) for natural image denoising, deconvolution, and accelerated magnetic resonance imaging (MRI). The results show that jittering significantly enhances the worst-case robustness, but can be suboptimal for inverse problems beyond denoising. Moreover, our results imply that training on real data which often contains slight noise is somewhat robustness enhancing.

Miscellaneous Aspects of Machine Learning/Transfer, Multitask and Meta-learning

Tung Nguyen, Sudhanshu Agrawal, Aditya Grover

Experimental design is a fundamental problem in many science and engineering fields. In this problem, sample efficiency is crucial due to the time, money, and safety costs of real-world design evaluations. Existing approaches either rely on active data collection or access to large, labeled datasets of past experiments, making them impractical in many real-world scenarios. In this work, we address the more challenging yet realistic setting of few-shot experimental design, where only a few labeled data points of input designs and their corresponding values are available. We approach this problem as a conditional generation task, where a model conditions on a few labeled examples and the desired output to generate an optimal input design. To this end, we introduce Experiment Pretrained Transformers (ExPT), a foundation model for few-shot experimental design that employs a novel combination of synthetic pretraining with in-context learning. In ExPT, we only assume knowledge of a finite collection of unlabelled data points from the input domain and pretrain a transformer neural network to optimize diverse synthetic functions defined over this domain. Unsupervised pretraining allows ExPT to adapt to any design task at test time in an in-context fashion by conditioning on a few labeled data points from the target task and generating the candidate optima. We evaluate ExPT on few-shot experimental design in challenging domains and demonstrate its superior generality and performance compared to existing methods. The source code is available at https://github.com/tung-nd/ExPT.git.

Deep Learning/Graph Neural Networks

Aditya Shahane, Saripilli Swapna Manjiri, Ankesh Jain, Sandeep Kumar

The design automation of analog circuits poses significant challenges in terms of the large design space, complex interdependencies between circuit specifications, and resource-intensive simulations. To address these challenges, this paper presents an innovative framework called the Graph of Circuits Explorer (GCX). Leveraging graph structure learning along with graph neural networks, GCX enables the creation of a surrogate model that facilitates efficient exploration of the optimal design space within a semi-supervised learning framework which reduces the need for large labelled datasets. The proposed approach comprises three key stages. First, we learn the geometric representation of circuits and enrich it with technology information to create a comprehensive feature vector. Subsequently, integrating feature-based graph learning with few-shot and zero-shot learning enhances the generalizability in predictions for unseen circuits. Finally, we introduce two algorithms namely, EASCO and ASTROG which upon integration with GCX optimize the available samples to yield the optimal circuit configuration meeting the designer's criteria. The effectiveness of the proposed approach is demonstrated through simulated performance evaluation of various circuits, using derived parameters in 180nm CMOS technology. Furthermore, the generalizability of the approach is extended to higher-order topologies and different technology nodes such as 65nm and 45nm CMOS process nodes.

Optimization/Non-Convex

Haochuan Li, Alexander Rakhlin, Ali Jadbabaie

In this paper, we provide a rigorous proof of convergence of the Adaptive Moment Estimate (Adam) algorithm for a wide class of optimization objectives. Despite the popularity and efficiency of the Adam algorithm in training deep neural networks, its theoretical properties are not yet fully understood, and existing convergence proofs require unrealistically strong assumptions, such as globally bounded gradients, to show the convergence to stationary points. In this paper, we show that Adam provably converges to $\epsilon$-stationary points with $\mathcal{O}(\epsilon^{-4})$ gradient complexity under far more realistic conditions. The key to our analysis is a new proof of boundedness of gradients along the optimization trajectory of Adam, under a generalized smoothness assumption according to which the local smoothness (i.e., Hessian norm when it exists) is bounded by a sub-quadratic function of the gradient norm. Moreover, we propose a variance-reduced version of Adam with an accelerated gradient complexity of $\mathcal{O}(\epsilon^{-3})$.

Social Aspects/Trustworthy Machine Learning

Russell Hart, Linlin Yu, Yifei Lou, Feng Chen

Deep neural networks have achieved significant success in the last decades, but they are not well-calibrated and often produce unreliable predictions. A large number of literature relies on uncertainty quantification to evaluate the reliability of a learning model, which is particularly important for applications of out-of-distribution (OOD) detection and misclassification detection. We are interested in uncertainty quantification for interdependent node-level classification. We start our analysis based on graph posterior networks (GPNs) that optimize the uncertainty cross-entropy (UCE)-based loss function. We describe the theoretical limitations of the widely-used UCE loss. To alleviate the identified drawbacks, we propose a distance-based regularization that encourages clustered OOD nodes to remain clustered in the latent space. We conduct extensive comparison experiments on eight standard datasets and demonstrate that the proposed regularization outperforms the state-of-the-art in both OOD detection and misclassification detection.

Deep Learning/Robustness

Nicholas Carlini, Milad Nasr, Christopher A. Choquette-Choo, Matthew Jagielski, Irena Gao, Pang Wei Koh, Daphne Ippolito, Florian Tramer, Ludwig Schmidt

Large language models are now tuned to align with the goals of their creators, namely to be "helpful and harmless." These models should respond helpfully to user questions, but refuse to answer requests that could cause harm. However, adversarial users can construct inputs which circumvent attempts at alignment. In this work, we study adversarial alignment, and ask to what extent these models remain aligned when interacting with an adversarial user who constructs worst-case inputs (adversarial examples). These inputs are designed to cause the model to emit harmful content that would otherwise be prohibited.We show that existing NLP-based optimization attacks are insufficiently powerful to reliably attack aligned text models: even when current NLP-based attacks fail, we can find adversarial inputs with brute force. As a result, the failure of current attacks should not be seen as proof that aligned text models remain aligned under adversarial inputs. However the recent trend in large-scale ML models is multimodal models that allow users to provide images that influence the text that is generated. We show these models can be easily attacked, i.e., induced to perform arbitrary un-aligned behavior through adversarial perturbation of the input image. We conjecture that improved NLP attacks may demonstrate this same level of adversarial control over text-only models.

Xiufeng Xie, Ning Zhou, Wentao Zhu, Ji Liu

The lightweight nature of IoT devices makes it challenging to run deep neural networks (DNNs) locally for applications like augmented reality. Recent advances in IoT communication like LTE-M have significantly boosted the link bandwidth, enabling IoT devices to stream visual data to edge servers running DNNs for inference. However, uncompressed visual data can still easily overload the IoT link, and the wireless spectrum is shared by numerous IoT devices, causing unstable link bandwidth. Mainstream codecs can reduce the traffic but at the cost of severe inference accuracy drops. Recent works on differentiable JPEG train the codec to tackle the damage to inference accuracy. But they rely on heuristic configurations in the loss function to balance the rate-accuracy tradeoff, providing no guarantee to meet the IoT bandwidth constraint. This paper presents AutoJPEG, a bandwidth-aware adaptive compression solution that learns the JPEG encoding parameters to optimize the DNN inference accuracy under bandwidth constraints. We model the compressed image size as a closed-form function of encoding parameters by analyzing the JPEG codec workflow. Furthermore, we formulate a constrained optimization framework to minimize the original DNN loss while ensuring the image size strictly meets the bandwidth constraint. Our evaluation validates AutoJPEG on various DNN models and datasets. In our experiments, AutoJPEG outperforms the mainstream codecs (like JPEG and WebP) and the state-of-the-art solutions that optimize the image codec for DNN inference.

Rahul Duggal, Hao Zhou, Shuo Yang, Jun Fang, Yuanjun Xiong, Wei Xia

With the shift towards on-device deep learning, ensuring a consistent behavior of an AI service across diverse compute platforms becomes tremendously important. Our work tackles the emergent problem of reducing predictive in-consistencies arising as negative flips: test samples that are correctly predicted by a less accurate on-device model, but incorrectly by a more accurate on-cloud one. We introduce REGression constrained Neural Architecture Search (REG-NAS) to design a family of highly accurate models that engender fewer negative flips. REG-NAS consists of two components: (1) A novel architecture constraint that enables a larger on-cloud model to contain all the weights of the smaller on-device one thus maximizing weight sharing. This idea stems from our observation that larger weight sharing among networks leads to similar sample-wise predictions and results in fewer negative flips; (2) A novel search reward that incorporates both Top-1 accuracy and negative flips in the architecture optimization metric. We demonstrate that REG-NAS can successfully find architecture with few negative flips, in three popular architecture search spaces. Compared to the existing state-of-the-art approach [29], REG-NAS leads to 33-48% relative reduction of negative flips.

Lianyu Hu, Liqing Gao, Zekang Liu, Wei Feng

Pooling methods are necessities for modern neural networks for increasing receptive fields and lowering down computational costs. However, commonly used hand-crafted pooling approaches, e.g. max pooling and average pooling, may not well preserve discriminative features. While many researchers have elaborately designed various pooling variants in spatial domain to handle these limitations with much progress, the temporal aspect is rarely visited where directly applying hand-crafted methods or these specialized spatial variants may not be optimal. In this paper, we derive temporal lift pooling (TLP) from the Lifting Scheme in signal processing to intelligently downsample features of different temporal hierarchies. The Lifting Scheme factorizes input signals into various sub-bands with different frequency, which can be viewed as different temporal movement patterns. Our TLP is a three-stage procedure, which performs signal decomposition, component weighting and information fusion to generate a refined downsized feature map. We select a typical temporal task with long sequences, i.e. continuous sign language recognition (CSLR), as our testbed to verify the effectiveness of TLP. Experiments on two large-scale datasets show TLP outperforms hand-crafted methods and specialized spatial variants by a large margin (1.5%) with similar computational overhead. As a robust feature extractor, TLP exhibits great generalizability upon multiple backbones on various datasets and achieves new state-of-the-art results on two large-scale CSLR datasets. Visualizations further demonstrate the mechanism of TLP in correcting gloss borders.

Jiangbei Yue, Dinesh Manocha, He Wang

Trajectory prediction has been widely pursued in many fields, and many model-based and model-free methods have been explored. The former include rule-based, geometric or optimization-based models, and the latter are mainly comprised of deep learning approaches. In this paper, we propose a new method combining both methodologies based on a new Neural Differential Equation model. Our new model (Neural Social Physics or NSP) is a deep neural network within which we use an explicit physics model with learnable parameters. The explicit physics model serves as a strong inductive bias in modeling pedestrian behaviors, while the rest of the network provides a strong data-fitting capability in terms of system parameter estimation and dynamics stochasticity modeling. We compare NSP with 15 recent deep learning methods on 6 datasets and improve the state-of-the-art performance by 5.56%-70%. Besides, we show that NSP has better generalizability in predicting plausible trajectories in drastically different scenarios where the density is 2-5 times as high as the testing data. Finally, we show that the physics model in NSP can provide plausible explanations for pedestrian behaviors, as opposed to black-box deep learning. Code is available: https://github.com/realcrane/Human- Trajectory-Prediction-via-Neural-Social-Physics.

Xingyi Yang, Jingwen Ye, Xinchao Wang

In this paper, we explore a novel and ambitious knowledge-transfer task, termed Knowledge Factorization (KF). The core idea of KF lies in the modularization and assemblability of knowledge: given a pretrained network model as input, KF aims to decompose it into several factor networks, each of which handles only a dedicated task and maintains task-specific knowledge factorized from the source network. Such factor networks are task-wise disentangled and can be directly assembled, without any fine-tuning, to produce the more competent combined-task networks. In other words, the factor networks serve as Lego-brick-like building blocks, allowing us to construct customized networks in a plug-and-play manner. Specifically, each factor network comprises two modules, a common-knowledge module that is task-agnostic and shared by all factor networks, alongside with a task-specific module dedicated to the factor network itself. We introduce an information-theoretic objective, InfoMax-Bottleneck (IMB), to carry out KF by optimizing the mutual information between the learned representations and input. Experiments across various benchmarks demonstrate that, the derived factor networks yield gratifying performances on not only the dedicated tasks but also disentanglement, while enjoying much better interpretability and modularity. Moreover, the learned common-knowledge representations give rise to impressive results on transfer learning. Our code is available at https://github.com/Adamdad/KnowledgeFactor.

Hyounguk Shon, Janghyeon Lee, Seung Hwan Kim, Junmo Kim

Pre-trained representation is one of the key elements in the success of modern deep learning. However, existing works on continual learning methods have mostly focused on learning models incrementally from scratch. In this paper, we explore an alternative framework to incremental learning where we continually fine-tune the model from a pre-trained representation. Our method takes advantage of linearization technique of a pre-trained neural network for simple and effective continual learning. We show that this allows us to design a linear model where quadratic parameter regularization method is placed as the optimal continual learning policy, and at the same time enjoying the high performance of neural networks. We also show that the proposed algorithm enables parameter regularization methods to be applied to class-incremental problems. Additionally, we provide a theoretical reason why the existing parameter-space regularization algorithms such as EWC underperform on neural networks trained with cross-entropy loss. We show that the proposed method can prevent forgetting while achieving high continual fine-tuning performance on image classification tasks. To show that our method can be applied to general continual learning settings, we evaluate our method in data-incremental, task-incremental, and class-incremental learning problems.

Jan U. Müller, Michael Weinmann, Reinhard Klein

We propose an efficient and GPU-accelerated sampling framework which enables unbiased gradient approximation for differentiable point cloud rendering based on surface splatting. Our framework models the contribution of a point to the rendered image as a probability distribution. We derive an unbiased approximative gradient for the rendering function within this model. To efficiently evaluate the proposed sample estimate, we introduce a tree-based data-structure which employs multi-pole methods to draw samples in near linear time. Our gradient estimator allows us to avoid regularization required by previous methods, leading to a more faithful shape recovery from images. Furthermore, we validate that these improvements are applicable to real-world applications by refining the camera poses and point cloud obtained from a real-time SLAM system. Finally, employing our framework in a neural rendering setting optimizes both the point cloud and network parameters, highlighting the framework’s ability to enhance data driven approaches.

Matteo Taiana, Matteo Toso, Stuart James, Alessio Del Bue

The estimation of the camera poses associated with a set of images commonly relies on feature matches between the images. In contrast, we are the first to address this challenge by using objectness regions to guide the pose estimation problem rather than explicit semantic object detections. We propose Pose Refiner Network (PoserNet) a light-weight Graph Neural Network to refine the approximate pair-wise relative camera poses. PoserNet exploits associations between the objectness regions - concisely expressed as bounding boxes - across multiple views to globally refine sparsely connected view graphs. We evaluate on the 7-Scenes dataset across varied sizes of graphs and show how this process can be beneficial to optimization-based motion averaging algorithms improving the median error on the rotation by 62 degrees with respect to the initial estimates obtained based on bounding boxes. Code and data are available at https://github.com/IIT-PAVIS/PoserNet.

Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, Yong-Jin Liu

Estimating the pose of a moving camera from monocular video is a challenging problem, especially due to the presence of moving objects in dynamic environments, where the performance of existing camera pose estimation methods are susceptible to pixels that are not geometrically consistent. To tackle this challenge, we present a robust dense indirect structure-from-motion method for videos that is based on dense correspondence initialized from pairwise optical flow. Our key idea is to optimize long range video correspondence as dense point trajectories and use it to learn robust estimation of motion segmentation. A novel neural network architecture is proposed for processing irregular point trajectory data. Camera poses are then estimated and optimized with global bundle adjustment over the portion of long-range point trajectories that are classified as static. Experiments on MPI Sintel dataset show that our system produces significantly more accurate camera trajectories compared to existing state-of-the-art methods. In addition, our method is able to retain reasonable accuracy of camera poses on fully static scenes, which consistently outperforms strong state-of-the-art dense correspondence based methods with end-to-end deep learning, demonstrating the potential of dense indirect methods based on optical flow and point trajectories. As the point trajectory representation is general, we further present results and comparisons on in-the-wild monocular videos with complex motion of dynamic objects. Code is available at https://github.com/bytedance/particle-sfm.