NeurIPS 2023 Papers

In this paper, we propose \textit{binary radiance fields} (BiRF), a storage-efficient radiance field representation employing binary feature encoding in a format of either $+1$ or $-1$. This binarization strategy lets us represent the feature grid with highly compact feature encoding and a dramatic reduction in storage size. Furthermore, our 2D-3D hybrid feature grid design enhances the compactness of feature encoding as the 3D grid includes main components while 2D grids capture details. In our experiments, binary radiance field representation successfully outperforms the reconstruction performance of state-of-the-art (SOTA) storage-efficient radiance field models with lower storage allocation. In particular, our model achieves impressive results in static scene reconstruction, with a PSNR of 32.03 dB for Synthetic-NeRF scenes, 34.48 dB for Synthetic-NSVF scenes, 28.20 dB for Tanks and Temples scenes while only utilizing 0.5 MB of storage space, respectively. We hope the proposed binary radiance field representation will make radiance fields more accessible without a storage bottleneck.

Complex-valued representation exists inherently in the time-sequential data that can be derived from the integration of harmonic waves. The non-stationary spectral kernel, realizing a complex-valued feature mapping, has shown its potential to analyze the time-varying statistical characteristics of the time-sequential data, as a result of the modeling frequency parameters. However, most existing spectral kernel-based methods eliminate the imaginary part, thereby limiting the representation power of the spectral kernel. To tackle this issue, we propose a generalized spectral kernel network, namely, \underline{Co}mplex-valued \underline{s}pectral kernel \underline{Net}work (CosNet), which includes spectral kernel mapping generalization (SKMG) module and complex-valued spectral kernel embedding (CSKE) module. Concretely, the SKMG module is devised to generalize the spectral kernel mapping in the real number domain to the complex number domain, recovering the inherent complex-valued representation for the real-valued data. Then a following CSKE module is further developed to combine the complex-valued spectral kernels and neural networks to effectively capture long-range or periodic relations of the data. Along with the CosNet, we study the effect of the complex-valued spectral kernel mapping via theoretically analyzing the bound of covering number and generalization error. Extensive experiments demonstrate that CosNet performs better than the mainstream kernel methods and complex-valued neural networks.

Deep learning methods have achieved state-of-the-art performance in most modeling tasks involving images, text and audio, however, they typically underperform tree-based methods on tabular data. In this paper, we hypothesize that a significant contributor to this performance gap is the interaction between irregular target functions resulting from the heterogeneous nature of tabular feature spaces, and the well-known tendency of neural networks to learn smooth functions. Utilizing tools from spectral analysis, we show that functions described by tabular datasets often have high irregularity, and that they can be smoothed by transformations such as scaling and ranking in order to improve performance. However, because these transformations tend to lose information or negatively impact the loss landscape during optimization, they need to be rigorously fine-tuned for each feature to achieve performance gains. To address these problems, we propose introducing frequency reduction as an inductive bias. We realize this bias as a neural network layer that promotes learning low-frequency representations of the input features, allowing the network to operate in a space where the target function is more regular. Our proposed method introduces less computational complexity than a fully connected layer, while significantly improving neural network performance, and speeding up its convergence on 14 tabular datasets.

The rapid scaling of language models is motivating research using low-bitwidth quantization. In this work, we propose a novel binarization technique for Transformers applied to machine translation (BMT), the first of its kind. We identify and address the problem of inflated dot-product variance when using one-bit weights and activations. Specifically, BMT leverages additional LayerNorms and residual connections to improve binarization quality. Experiments on the WMT dataset show that a one-bit weight-only Transformer can achieve the same quality as a float one, while being 16$\times$ smaller in size. One-bit activations incur varying degrees of quality drop, but mitigated by the proposed architectural changes. We further conduct a scaling law study using production-scale translation datasets, which shows that one-bit weight Transformers scale and generalize well in both in-domain and out-of-domain settings. Implementation in JAX/Flax will be open sourced.

Interacting particle systems have proven highly successful in various machine learning tasks, including approximate Bayesian inference and neural network optimization. However, the analysis of these systems often relies on the simplifying assumption of the \emph{mean-field} limit, where particle numbers approach infinity and infinitesimal step sizes are used. In practice, discrete time steps, finite particle numbers, and complex integration schemes are employed, creating a theoretical gap between continuous-time and discrete-time processes. In this paper, we present a novel framework that establishes a precise connection between these discrete-time schemes and their corresponding mean-field limits in terms of convergence properties and asymptotic behavior. By adopting a dynamical system perspective, our framework seamlessly integrates various numerical schemes that are typically analyzed independently. For example, our framework provides a unified treatment of optimizing an infinite-width two-layer neural network and sampling via Stein Variational Gradient descent, which were previously studied in isolation.

EXplainable AI (XAI) is an essential topic to improve human understanding of deep neural networks (DNNs) given their black-box internals. For computer vision tasks, mainstream pixel-based XAI methods explain DNN decisions by identifying important pixels, and emerging concept-based XAI explore forming explanations with concepts (e.g., a head in an image). However, pixels are generally hard to interpret and sensitive to the imprecision of XAI methods, whereas “concepts” in prior works require human annotation or are limited to pre-defined concept sets. On the other hand, driven by large-scale pre-training, Segment Anything Model (SAM) has been demonstrated as a powerful and promotable framework for performing precise and comprehensive instance segmentation, enabling automatic preparation of concept sets from a given image. This paper for the first time explores using SAM to augment concept-based XAI. We offer an effective and flexible concept-based explanation method, namely Explain Any Concept (EAC), which explains DNN decisions with any concept. While SAM is highly effective and offers an “out-of-the-box” instance segmentation, it is costly when being integrated into defacto XAI pipelines. We thus propose a lightweight per-input equivalent (PIE) scheme, enabling efficient explanation with a surrogate model. Our evaluation over two popular datasets (ImageNet and COCO) illustrate the highly encouraging performance of EAC over commonly-used XAI methods.

One of the most impactful findings in computational neuroscience over the past decade is that the object recognition accuracy of deep neural networks (DNNs) correlates with their ability to predict neural responses to natural images in the inferotemporal (IT) cortex. This discovery supported the long-held theory that object recognition is a core objective of the visual cortex, and suggested that more accurate DNNs would serve as better models of IT neuron responses to images. Since then, deep learning has undergone a revolution of scale: billion parameter-scale DNNs trained on billions of images are rivaling or outperforming humans at visual tasks including object recognition. Have today's DNNs become more accurate at predicting IT neuron responses to images as they have grown more accurate at object recognition? Surprisingly, across three independent experiments, we find that this is not the case. DNNs have become progressively worse models of IT as their accuracy has increased on ImageNet. To understand why DNNs experience this trade-off and evaluate if they are still an appropriate paradigm for modeling the visual system, we turn to recordings of IT that capture spatially resolved maps of neuronal activity elicited by natural images. These neuronal activity maps reveal that DNNs trained on ImageNet learn to rely on different visual features than those encoded by IT and that this problem worsens as their accuracy increases. We successfully resolved this issue with the neural harmonizer, a plug-and-play training routine for DNNs that aligns their learned representations with humans. Our results suggest that harmonized DNNs break the trade-off between ImageNet accuracy and neural prediction accuracy that assails current DNNs and offer a path to more accurate models of biological vision. Our work indicates that the standard approach for modeling IT with task-optimized DNNs needs revision, and other biological constraints, including human psychophysics data, are needed to accurately reverse-engineer the visual cortex.

Designing models that are both expressive and preserve known invariances of tasks is an increasingly hard problem. Existing solutions tradeoff invariance for computational or memory resources. In this work, we show how to leverage randomness and design models that are both expressive and invariant but use less resources. Inspired by randomized algorithms, our key insight is that accepting probabilistic notions of universal approximation and invariance can reduce our resource requirements. More specifically, we propose a class of binary classification models called Randomized Linear Classifiers (RLCs). We give parameter and sample size conditions in which RLCs can, with high probability, approximate any (smooth) function while preserving invariance to compact group transformations. Leveraging this result, we design three RLCs that are provably probabilistic invariant for classification tasks over sets, graphs, and spherical data. We show how these models can achieve probabilistic invariance and universality using less resources than (deterministic) neural networks and their invariant counterparts. Finally, we empirically demonstrate the benefits of this new class of models on invariant tasks where deterministic invariant neural networks are known to struggle.

This paper investigates methods for improving generative data augmentation for deep learning. Generative data augmentation leverages the synthetic samples produced by generative models as an additional dataset for classification with small dataset settings. A key challenge of generative data augmentation is that the synthetic data contain uninformative samples that degrade accuracy. This can be caused by the synthetic samples not perfectly representing class categories in real data and uniform sampling not necessarily providing useful samples for tasks. In this paper, we present a novel strategy for generative data augmentation called *meta generative regularization* (MGR). To avoid the degradation of generative data augmentation, MGR utilizes synthetic samples for regularizing feature extractors instead of training classifiers. These synthetic samples are dynamically determined to minimize the validation losses through meta-learning. We observed that MGR can avoid the performance degradation of naive generative data augmentation and boost the baselines. Experiments on six datasets showed that MGR is effective particularly when datasets are smaller and stably outperforms baselines by up to 7 percentage points on test accuracy.

Ensembles of machine learning models have been well established as a powerful method of improving performance over a single model. Traditionally, ensembling algorithms train their base learners independently or sequentially with the goal of optimizing their joint performance. In the case of deep ensembles of neural networks, we are provided with the opportunity to directly optimize the true objective: the joint performance of the ensemble as a whole. Surprisingly, however, directly minimizing the loss of the ensemble appears to rarely be applied in practice. Instead, most previous research trains individual models independently with ensembling performed _post hoc_. In this work, we show that this is for good reason - _joint optimization of ensemble loss results in degenerate behavior_. We approach this problem by decomposing the ensemble objective into the strength of the base learners and the diversity between them. We discover that joint optimization results in a phenomenon in which base learners collude to artificially inflate their apparent diversity. This pseudo-diversity fails to generalize beyond the training data, causing a larger generalization gap. We proceed to comprehensively demonstrate the practical implications of this effect on a range of standard machine learning tasks and architectures by smoothly interpolating between independent training and joint optimization.

Learning relies on coordinated synaptic changes in recurrently connected populations of neurons. Therefore, understanding the collective evolution of synaptic connectivity over learning is a key challenge in neuroscience and machine learning. In particular, recent work has shown that the weight matrices of task-trained RNNs are typically low rank, but how this low rank structure unfolds over learning is unknown. To address this, we investigate the rank of the 3-tensor formed by the weight matrices throughout learning. By fitting RNNs of varying rank to large-scale neural recordings during a motor learning task, we find that the inferred weights are low-tensor-rank and therefore evolve over a fixed low-dimensional subspace throughout the entire course of learning. We next validate the observation of low-tensor-rank learning on an RNN trained to solve the same task. Finally, we present a set of mathematical results bounding the matrix and tensor ranks of gradient descent learning dynamics which show that low-tensor-rank weights emerge naturally in RNNs trained to solve low-dimensional tasks. Taken together, our findings provide insight on the evolution of population connectivity over learning in both biological and artificial neural networks, and enable reverse engineering of learning-induced changes in recurrent dynamics from large-scale neural recordings.

The design of \emph{de novo} functional proteins that bind with specific ligand molecules is crucial in various domains like therapeutics and bio-engineering. One vital yet challenging step is to design the protein pocket, the cavity region of protein where the ligand binds with. Existing methods suffer from inefficient generation, insufficient context modeling (ligand molecule), and incapability of generating sidechain atoms. To overcome the limitations, we propose a \textbf{F}ull-\textbf{A}tom \textbf{I}terative \textbf{R}efinement framework (\textbf{FAIR}) for protein pocket sequence (i.e., residue types) and 3D structure co-design. Generally, FAIR consists of two steps that follow a coarse-to-fine pipeline (backbone atoms to full atoms including sidechain) for full-atom generation. For efficiency, all residue types and structures are updated together in each round (i.e., full-shot refinement). In the first step, the residue types and backbone coordinates are updated with a hierarchical context encoder and two structure refinement modules capturing inter-residue and pocket-ligand interactions. The second step further models the sidechain atoms of pockets and updates residue types to achieve sequence-structure consistency. The structure of the binding ligand is also updated along with the above refinement iterations accounting for its flexibility. Finally, extensive evaluations show that FAIR outperforms baselines in efficiently designing high-quality pocket sequences and structures. Specifically, the average improvements on AAR and RMSD are over 10$\%$.

Engineering design problems frequently require solving systems of partial differential equations with boundary conditions specified on object geometries in the form of a triangular mesh. These boundary geometries are provided by a designer and are problem dependent. The efficiency of the design process greatly benefits from fast turnaround times when repeatedly solving PDEs on various geometries. However, most current work that uses machine learning to speed up the solution process relies heavily on a fixed parameterization of the geometry, which cannot be changed after training. This severely limits the possibility of reusing a trained model across a variety of design problems. In this work, we propose a novel neural operator architecture which accepts boundary geometry, in the form of triangular meshes, as input and produces an approximate solution to a given PDE as output. Once trained, the model can be used to rapidly estimate the PDE solution over a new geometry, without the need for retraining or representation of the geometry to a pre-specified parameterization.

Multiplications are responsible for most of the computational cost involved in neural network training and inference. Recent research has thus looked for ways to reduce the cost associated with them. Inspired by Mogami 2020, we replace multiplication with a cheap piecewise affine approximation that is achieved by adding the bit representation of the floating point numbers together as integers. We show that transformers can be trained with the resulting modified matrix multiplications on both vision and language tasks with little to no performance impact, and without changes to the training hyperparameters. We further replace all non-linearities in the networks making them fully and jointly piecewise affine in both inputs and weights. Finally, we show that we can eliminate all multiplications in the entire training process, including operations in the forward pass, backward pass and optimizer update, demonstrating the first successful training of modern neural network architectures in a fully multiplication-free fashion.

Training on large-scale graphs has achieved remarkable results in graph representation learning, but its cost and storage have attracted increasing concerns. Existing graph condensation methods primarily focus on optimizing the feature matrices of condensed graphs while overlooking the impact of the structure information from the original graphs. To investigate the impact of the structure information, we conduct analysis from the spectral domain and empirically identify substantial Laplacian Energy Distribution (LED) shifts in previous works. Such shifts lead to poor performance in cross-architecture generalization and specific tasks, including anomaly detection and link prediction. In this paper, we propose a novel Structure-broadcasting Graph Dataset Distillation (\textbf{SGDD}) scheme for broadcasting the original structure information to the generation of the synthetic one, which explicitly prevents overlooking the original structure information. Theoretically, the synthetic graphs by SGDD are expected to have smaller LED shifts than previous works, leading to superior performance in both cross-architecture settings and specific tasks. We validate the proposed SGDD~across 9 datasets and achieve state-of-the-art results on all of them: for example, on YelpChi dataset, our approach maintains 98.6\% test accuracy of training on the original graph dataset with 1,000 times saving on the scale of the graph. Moreover, we empirically evaluate there exist 17.6\% $\sim$ 31.4\% reductions in LED shift crossing 9 datasets. Extensive experiments and analysis verify the effectiveness and necessity of the proposed designs. The code will be made public.

Graph learning from signals is a core task in graph signal processing (GSP). A significant subclass of graph signals called the stationary graph signals that broadens the concept of stationarity of data defined on regular domains to signals on graphs is gaining increasing popularity in the GSP community. The most commonly used model to learn graphs from these stationary signals is SpecT, which forms the foundation for nearly all the subsequent, more advanced models. Despite its strengths, the practical formulation of the model, known as rSpecT, has been identified to be susceptible to the choice of hyperparameters. More critically, it may suffer from infeasibility as an optimization problem. In this paper, we introduce the first condition that ensures the infeasibility of rSpecT and design a novel model called LogSpecT, along with its practical formulation rLogSpecT to overcome this issue. Contrary to rSpecT, our novel practical model rLogSpecT is always feasible. Furthermore, we provide recovery guarantees of rLogSpecT from modern optimization tools related to epi-convergence, which could be of independent interest and significant for various learning problems. To demonstrate the practical advantages of rLogSpecT, a highly efficient algorithm based on the linearized alternating direction method of multipliers (L-ADMM) that allows closed-form solutions for each subproblem is proposed with convergence guarantees. Extensive numerical results on both synthetic and real networks not only corroborate the stability of our proposed methods, but also highlight their comparable and even superior performance than existing models.

Federated Learning (FL) aims to train a global inference model from remotely distributed clients, gaining popularity due to its benefit of improving data privacy. However, traditional FL often faces challenges in practical applications, including model overfitting and divergent local models due to limited and non-IID data among clients. To address these issues, we introduce a novel Bayesian meta-learning approach called meta-variational dropout (MetaVD). MetaVD learns to predict client-dependent dropout rates via a shared hypernetwork, enabling effective model personalization of FL algorithms in limited non-IID data settings. We also emphasize the posterior adaptation view of meta-learning and the posterior aggregation view of Bayesian FL via the conditional dropout posterior. We conducted extensive experiments on various sparse and non-IID FL datasets. MetaVD demonstrated excellent classification accuracy and uncertainty calibration performance, especially for out-of-distribution (OOD) clients. MetaVD compresses the local model parameters needed for each client, mitigating model overfitting and reducing communication costs. Code is available at https://github.com/insujeon/MetaVD.

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.

Recent studies have demonstrated the susceptibility of deep neural networks to backdoor attacks. Given a backdoored model, its prediction of a poisoned sample with trigger will be dominated by the trigger information, though trigger information and benign information coexist. Inspired by the mechanism of the optical polarizer that a polarizer could pass light waves with particular polarizations while filtering light waves with other polarizations, we propose a novel backdoor defense method by inserting a learnable neural polarizer into the backdoored model as an intermediate layer, in order to purify the poisoned sample via filtering trigger information while maintaining benign information. The neural polarizer is instantiated as one lightweight linear transformation layer, which is learned through solving a well designed bi-level optimization problem, based on a limited clean dataset. Compared to other fine-tuning-based defense methods which often adjust all parameters of the backdoored model, the proposed method only needs to learn one additional layer, such that it is more efficient and requires less clean data. Extensive experiments demonstrate the effectiveness and efficiency of our method in removing backdoors across various neural network architectures and datasets, especially in the case of very limited clean data. Codes are available at \href{https://github.com/SCLBD/BackdoorBench}{https://github.com/SCLBD/BackdoorBench} (PyTorch) and \href{https://github.com/JulieCarlon/NPD-MindSpore}{https://github.com/JulieCarlon/NPD-MindSpore} (MindSpore).