NeurIPS 2024 Papers

Iterative algorithms solve problems by taking steps until a solution is reached. Models in the form of Deep Thinking (DT) networks have been demonstrated to learn iterative algorithms in a way that can scale to different sized problems at inference time using recurrent computation and convolutions. However, they are often unstable during training, and have no guarantees of convergence/termination at the solution. This paper addresses the problem of instability by analyzing the growth in intermediate representations, allowing us to build models (referred to as Deep Thinking with Lipschitz Constraints (DT-L)) with many fewer parameters and providing more reliable solutions. Additionally our DT-L formulation provides guarantees of convergence of the learned iterative procedure to a unique solution at inference time. We demonstrate DT-L is capable of robustly learning algorithms which extrapolate to harder problems than in the training set. We benchmark on the traveling salesperson problem to evaluate the capabilities of the modified system in an NP-hard problem where DT fails to learn.

Cognitive processes usually take place at a macroscopic scale in systems characterised by emergent properties, which make the whole ‘more than the sum of its parts.’ While recent proposals have provided quantitative, information-theoretic metrics to detect emergence in time series data, it is often highly non-trivial to identify the relevant macroscopic variables a priori. In this paper we leverage recent advances in representation learning and differentiable information estimators to put forward a data-driven method to find emergent variables. The proposed method successfully detects emergent variables and recovers the ground-truth emergence values in a synthetic dataset. Furthermore, we show the method can be extended to learn multiple independent features, extracting a diverse set of emergent quantities. We finally show that a modified method scales to real experimental data from primate brain activity, paving the ground for future analyses uncovering the emergent structure of cognitive representations in biological and artificial intelligence systems.

Due to the recurrent structure of RNN, the long information propagation path poses limitations in capturing long-term dependencies, gradient explosion/vanishing issues, and inefficient sequential execution. Based on this, we propose a novel paradigm called Parallel Gated Network (PGN) as the new successor to RNN. PGN directly captures information from previous time steps through the designed Historical Information Extraction (HIE) layer and leverages gated mechanisms to select and fuse it with the current time step information. This reduces the information propagation path to $\mathcal{O}(1)$, effectively addressing the limitations of RNN. To enhance PGN's performance in long-range time series forecasting tasks, we propose a novel temporal modeling framework called Temporal PGN (TPGN). TPGN incorporates two branches to comprehensively capture the semantic information of time series. One branch utilizes PGN to capture long-term periodic patterns while preserving their local characteristics. The other branch employs patches to capture short-term information and aggregate the global representation of the series. TPGN achieves a theoretical complexity of $\mathcal{O}(\sqrt{L})$, ensuring efficiency in its operations. Experimental results on five benchmark datasets demonstrate the state-of-the-art (SOTA) performance and high efficiency of TPGN, further confirming the effectiveness of PGN as the new successor to RNN in long-range time series forecasting. The code is available in this repository: https://github.com/Water2sea/TPGN.

Multi-label image recognition aims to predict all objects present in an input image. A common belief is that modeling the correlations between objects is beneficial for multi-label recognition. However, this belief has been recently challenged as label correlations may mislead the classifier in testing, due to the possible contextual bias in training. Accordingly, a few of recent works not only discarded label correlation modeling, but also advocated to remove contextual information for multi-label image recognition. This work explicitly explores label correlations for multi-label image recognition based on a principled causal intervention approach. With causal intervention, we pursue causal label correlations and suppress spurious label correlations, as the former tend to convey useful contextual cues while the later may mislead the classifier. Specifically, we decouple label-specific features with a Transformer decoder attached to the backbone network, and model the confounders which may give rise to spurious correlations by clustering spatial features of all training images. Based on label-specific features and confounders, we employ a cross-attention module to implement causal intervention, quantifying the causal correlations from all object categories to each predicted object category. Finally, we obtain image labels by combining the predictions from decoupled features and causal label correlations. Extensive experiments clearly validate the effectiveness of our approach for multi-label image recognition in both common and cross-dataset settings.

Sparse Mixture of Experts (SMoE) has become the key to unlocking unparalleled scalability in deep learning. SMoE has the potential to exponentially increase in parameter count while maintaining the efficiency of the model by only activating a small subset of these parameters for a given sample. However, it has been observed that SMoE suffers from unstable training and has difficulty adapting to new distributions, leading to the model's lack of robustness to data contamination. To overcome these limitations, we first establish a connection between the dynamics of the expert representations in SMoEs and gradient descent on a multi-objective optimization problem. Leveraging our framework, we then integrate momentum into SMoE and propose a new family of SMoEs, named MomentumSMoE. We theoretically prove and numerically validate that MomentumSMoE is more stable and robust than SMoE. In particular, we verify the advantages of MomentumSMoE over SMoE on a variety of practical tasks including ImageNet-1K object recognition and WikiText-103 language modeling. We demonstrate the applicability of MomentumSMoE to many types of SMoE models, including those in the Sparse MoE model for vision (V-MoE) and the Generalist Language Model (GLaM). We also show that other advanced momentum-based optimization methods, such as Adam, can be easily incorporated into the MomentumSMoE framework for designing new SMoE models with even better performance, almost negligible additional computation cost, and simple implementations.

Transformers with linear attention (i.e., linear transformers) and state-space models have recently been suggested as a viable linear-time alternative to transformers with softmax attention. However, these models still underperform transformers especially on tasks that require in-context retrieval. While more expressive variants of linear transformers which replace the additive update in linear transformers with the delta rule (DeltaNet) have been found to be more effective at associative recall, existing algorithms for training such models do not parallelize over sequence length and are thus inefficient to train on modern hardware. This work describes a hardware-efficient algorithm for training linear transformers with the delta rule, which exploits a memory-efficient representation for computing products of Householder matrices. This algorithm allows us to scale up DeltaNet to standard language modeling settings. We train a 1.3B model for 100B tokens and find that it outperforms recent linear-time baselines such as Mamba and GLA in terms of perplexity and zero-shot performance on downstream tasks. We also experiment with two hybrid models which combine DeltaNet layers with (1) sliding-window attention layers every other layer or (2) two global attention layers, and find that these hybrids outperform strong transformer baselines.

Neural networks (NN) are extensively studied in cutting-edge soft sensor models due to their feature extraction and function approximation capabilities. Current research into network-based methods primarily focuses on models' offline accuracy. Notably, in industrial soft sensor context, online optimizing stability and interpretability are prioritized, followed by accuracy. This requires a clearer understanding of network's training process. To bridge this gap, we propose a novel NN named the Approximated Orthogonal Projection Unit (AOPU) which has solid mathematical basis and presents superior training stability. AOPU truncates the gradient backpropagation at dual parameters, optimizes the trackable parameters updates, and enhances the robustness of training. We further prove that AOPU attains minimum variance estimation in NN, wherein the truncated gradient approximates the natural gradient. Empirical results on two chemical process datasets clearly show that AOPU outperforms other models in achieving stable convergence, marking a significant advancement in soft sensor field.

The past neural network design has largely focused on feature \textit{representation space} dimension and its capacity scaling (e.g., width, depth), but overlooked the feature \textit{interaction space} scaling. Recent advancements have shown shifted focus towards element-wise multiplication to facilitate higher-dimensional feature interaction space for better information transformation. Despite this progress, multiplications predominantly capture low-order interactions, thus remaining confined to a finite-dimensional interaction space. To transcend this limitation, classic kernel methods emerge as a promising solution to engage features in an infinite-dimensional space. We introduce InfiNet, a model architecture that enables feature interaction within an infinite-dimensional space created by RBF kernel. Our experiments reveal that InfiNet achieves new state-of-the-art, owing to its capability to leverage infinite-dimensional interactions, significantly enhancing model performance.

CRATE, a white-box transformer architecture designed to learn compressed and sparse representations, offers an intriguing alternative to standard vision transformers (ViTs) due to its inherent mathematical interpretability. Despite extensive investigations into the scaling behaviors of language and vision transformers, the scalability of CRATE remains an open question which this paper aims to address. Specifically, we propose CRATE-$\alpha$, featuring strategic yet minimal modifications to the sparse coding block in the CRATE architecture design, and a light training recipe designed to improve the scalability of CRATE. Through extensive experiments, we demonstrate that CRATE-$\alpha$ can effectively scale with larger model sizes and datasets. For example, our CRATE-$\alpha$-B substantially outperforms the prior best CRATE-B model accuracy on ImageNet classification by 3.7%, achieving an accuracy of 83.2%. Meanwhile, when scaling further, our CRATE-$\alpha$-L obtains an ImageNet classification accuracy of 85.1%. More notably, these model performance improvements are achieved while preserving, and potentially even enhancing the interpretability of learned CRATE models, as we demonstrate through showing that the learned token representations of increasingly larger trained CRATE-$\alpha$ models yield increasingly higher-quality unsupervised object segmentation of images.

Implicit neural representations (INRs) have proven effective in various tasks including image, shape, audio, and video reconstruction. These INRs typically learn the implicit field from sampled input points. This is often done using a single network for the entire domain, imposing many global constraints on a single function. In this paper, we propose a mixture of experts (MoE) implicit neural representation approach that enables learning local piece-wise continuous functions that simultaneously learns to subdivide the domain and fit it locally. We show that incorporating a mixture of experts architecture into existing INR formulations provides a boost in speed, accuracy, and memory requirements. Additionally, we introduce novel conditioning and pretraining methods for the gating network that improves convergence to the desired solution. We evaluate the effectiveness of our approach on multiple reconstruction tasks, including surface reconstruction, image reconstruction, and audio signal reconstruction and show improved performance compared to non-MoE methods. Code is available at our project page https://sitzikbs.github.io/neural-experts-projectpage/ .

This work presents a novel approach to neural architecture search (NAS) that aims to increase carbon efficiency for the model design process. The proposed framework CE-NAS addresses the key challenge of high carbon cost associated with NAS by exploring the carbon emission variations of energy and energy differences of different NAS algorithms. At the high level, CE-NAS leverages a reinforcement-learning agent to dynamically adjust GPU resources based on carbon intensity, predicted by a time-series transformer, to balance energy-efficient sampling and energy-intensive evaluation tasks. Furthermore, CE-NAS leverages a recently proposed multi-objective optimizer to effectively reduce the NAS search space. We demonstrate the efficacy of CE-NAS in lowering carbon emissions while achieving SOTA results for both NAS datasets and open-domain NAS tasks. For example, on the HW-NasBench dataset, CE-NAS reduces carbon emissions by up to 7.22X while maintaining a search efficiency comparable to vanilla NAS. For open-domain NAS tasks, CE-NAS achieves SOTA results with 97.35% top-1 accuracy on CIFAR-10 with only 1.68M parameters and a carbon consumption of 38.53 lbs of CO2. On ImageNet, our searched model achieves 80.6% top-1 accuracy with a 0.78 ms TensorRT latency using FP16 on NVIDIA V100, consuming only 909.86 lbs of CO2, making it comparable to other one-shot-based NAS baselines. Our code is available at https://github.com/cake-lab/CE-NAS.

Complex visual reasoning and question answering (VQA) is a challenging task that requires compositional multi-step processing and higher-level reasoning capabilities beyond the immediate recognition and localization of objects and events. Here, we introduce a fully neural Iterative and Parallel Reasoning Mechanism (IPRM) that combines two distinct forms of computation -- iterative and parallel -- to better address complex VQA scenarios. Specifically, IPRM's "iterative" computation facilitates compositional step-by-step reasoning for scenarios wherein individual operations need to be computed, stored, and recalled dynamically (e.g. when computing the query “determine the color of pen to the left of the child in red t-shirt sitting at the white table”). Meanwhile, its "parallel'' computation allows for the simultaneous exploration of different reasoning paths and benefits more robust and efficient execution of operations that are mutually independent (e.g. when counting individual colors for the query: "determine the maximum occurring color amongst all t-shirts'"). We design IPRM as a lightweight and fully-differentiable neural module that can be conveniently applied to both transformer and non-transformer vision-language backbones. It notably outperforms prior task-specific methods and transformer-based attention modules across various image and video VQA benchmarks testing distinct complex reasoning capabilities such as compositional spatiotemporal reasoning (AGQA), situational reasoning (STAR), multi-hop reasoning generalization (CLEVR-Humans) and causal event linking (CLEVRER-Humans). Further, IPRM's internal computations can be visualized across reasoning steps, aiding interpretability and diagnosis of its errors.

Deep Neural Networks (DNNs) have revolutionized artificial intelligence, achieving impressive results on diverse data types, including images, videos, and texts. However, DNNs still lag behind Gradient Boosting Decision Trees (GBDT) on tabular data, a format extensively utilized across various domains. This paper introduces DOFEN, which stands for Deep Oblivious Forest ENsemble. DOFEN is a novel DNN architecture inspired by oblivious decision trees and achieves on-off sparse selection of columns. DOFEN surpasses other DNNs on tabular data, achieving state-of-the-art performance on the well-recognized benchmark: Tabular Benchmark, which includes 73 total datasets spanning a wide array of domains. The code of DOFEN is available at: https://github.com/Sinopac-Digital-Technology-Division/DOFEN

Human Mesh Recovery (HMR) is the task of estimating a parameterized 3D human mesh from an image. There is a kind of methods first training a regression model for this problem, then further optimizing the pretrained regression model for any specific sample individually at test time. However, the pretrained model may not provide an ideal optimization starting point for the test-time optimization. Inspired by meta-learning, we incorporate the test-time optimization into training, performing a step of test-time optimization for each sample in the training batch before really conducting the training optimization over all the training samples. In this way, we obtain a meta-model, the meta-parameter of which is friendly to the test-time optimization. At test time, after several test-time optimization steps starting from the meta-parameter, we obtain much higher HMR accuracy than the test-time optimization starting from the simply pretrained regression model. Furthermore, we find test-time HMR objectives are different from training-time objectives, which reduces the effectiveness of the learning of the meta-model. To solve this problem, we propose a dual-network architecture that unifies the training-time and test-time objectives. Our method, armed with meta-learning and the dual networks, outperforms state-of-the-art regression-based and optimization-based HMR approaches, as validated by the extensive experiments. The codes are available at https://github.com/fmx789/Meta-HMR.

Given the ubiquitous presence of time series data across various domains, precise forecasting of time series holds significant importance and finds widespread real-world applications such as energy, weather, healthcare, etc. While numerous forecasters have been proposed using different network architectures, the Transformer-based models have state-of-the-art performance in time series forecasting. However, forecasters based on Transformers are still suffering from vulnerability to high-frequency signals, efficiency in computation, and bottleneck in full-spectrum utilization, which essentially are the cornerstones for accurately predicting time series with thousands of points. In this paper, we explore a novel perspective of enlightening signal processing for deep time series forecasting. Inspired by the filtering process, we introduce one simple yet effective network, namely FilterNet, built upon our proposed learnable frequency filters to extract key informative temporal patterns by selectively passing or attenuating certain components of time series signals. Concretely, we propose two kinds of learnable filters in the FilterNet: (i) Plain shaping filter, that adopts a universal frequency kernel for signal filtering and temporal modeling; (ii) Contextual shaping filter, that utilizes filtered frequencies examined in terms of its compatibility with input signals for dependency learning. Equipped with the two filters, FilterNet can approximately surrogate the linear and attention mappings widely adopted in time series literature, while enjoying superb abilities in handling high-frequency noises and utilizing the whole frequency spectrum that is beneficial for forecasting. Finally, we conduct extensive experiments on eight time series forecasting benchmarks, and experimental results have demonstrated our superior performance in terms of both effectiveness and efficiency compared with state-of-the-art methods. Our code is available at$^1$.

Numerous industrial sectors necessitate models capable of providing robust forecasts across various horizons. Despite the recent strides in crafting specific architectures for time-series forecasting and developing pre-trained universal models, a comprehensive examination of their capability in accommodating varied-horizon forecasting during inference is still lacking. This paper bridges this gap through the design and evaluation of the Elastic Time-Series Transformer (ElasTST). The ElasTST model incorporates a non-autoregressive design with placeholders and structured self-attention masks, warranting future outputs that are invariant to adjustments in inference horizons. A tunable version of rotary position embedding is also integrated into ElasTST to capture time-series-specific periods and enhance adaptability to different horizons. Additionally, ElasTST employs a multi-scale patch design, effectively integrating both fine-grained and coarse-grained information. During the training phase, ElasTST uses a horizon reweighting strategy that approximates the effect of random sampling across multiple horizons with a single fixed horizon setting. Through comprehensive experiments and comparisons with state-of-the-art time-series architectures and contemporary foundation models, we demonstrate the efficacy of ElasTST's unique design elements. Our findings position ElasTST as a robust solution for the practical necessity of varied-horizon forecasting.

Popular parameter-efficient fine-tuning (PEFT) methods, such as LoRA and its variants, freeze pre-trained model weights $\(\mathbf{W}\)$ and inject learnable matrices $\(\mathbf{\Delta W}\)$. These $\(\mathbf{\Delta W}\)$ matrices are structured for efficient parameterization, often using techniques like low-rank approximations or scaling vectors. However, these methods typically exhibit a performance gap compared to full fine-tuning. While recent PEFT methods have narrowed this gap, they do so at the expense of additional learnable parameters. We propose SVFT, a *simple* approach that structures $\(\mathbf{\Delta W}\)$ based on the specific weight matrix $\(\mathbf{W}\)$. SVFT updates $\(\mathbf{W}\)$ as a sparse combination $\(M\)$ of outer products of its singular vectors, training only the coefficients of these combinations. Crucially, we make additional off-diagonal elements in $M$ learnable, enabling a smooth trade-off between trainable parameters and expressivity—an aspect that distinctly sets our approach apart from previous works leveraging singular values. Extensive experiments on language and vision benchmarks show that SVFT recovers up to **96%** of full fine-tuning performance while training only **0.006 to 0.25%** of parameters, outperforming existing methods that achieve only up to **{85\%}** performance with **0.03 to 0.8%** of the trainable parameter budget.

As the key component in multimodal large language models (MLLMs), the ability of the visual encoder greatly affects MLLM's understanding on diverse image content. Although some large-scale pretrained vision encoders such as vision encoders in CLIP and DINOv2 have brought promising performance, we found that there is still no single vision encoder that can dominate various image content understanding, e.g., the CLIP vision encoder leads to outstanding results on general image understanding but poor performance on document or chart content. To alleviate the bias of CLIP vision encoder, we first delve into the inherent behavior of different pre-trained vision encoders and then propose the MoVA, a powerful and novel MLLM, adaptively routing and fusing task-specific vision experts with a coarse-to-fine mechanism. In the coarse-grained stage, we design a context-aware expert routing strategy to dynamically select the most suitable vision experts according to the user instruction, input image, and expertise of vision experts. This benefits from the powerful model function understanding ability of the large language model (LLM). In the fine-grained stage, we elaborately conduct the mixture-of-vision-expert adapter (MoV-Adapter) to extract and fuse task-specific knowledge from various experts. This coarse-to-fine paradigm effectively leverages representations from experts based on multimodal context and model expertise, further enhancing the generalization ability. We conduct extensive experiments to evaluate the effectiveness of the proposed approach. Without any bells and whistles, MoVA can achieve significant performance gains over current state-of-the-art methods in a wide range of challenging multimodal benchmarks.

Linear RNN architectures, like Mamba, can be competitive with Transformer models in language modeling while having advantageous deployment characteristics. Given the focus on training large-scale Transformer models, we consider the challenge of converting these pretrained models for deployment. We demonstrate that it is feasible to distill large Transformers into linear RNNs by reusing the linear projection weights from attention layers with academic GPU resources. The resulting hybrid model, which incorporates a quarter of the attention layers, achieves performance comparable to the original Transformer in chat benchmarks and outperforms open-source hybrid Mamba models trained from scratch with trillions of tokens in both chat benchmarks and general benchmarks. Moreover, we introduce a hardware-aware speculative decoding algorithm that accelerates the inference speed of Mamba and hybrid models. Overall we show how, with limited computation resources, we can remove many of the original attention layers and generate from the resulting model more efficiently. Our top-performing model, distilled from Llama3-8B-Instruct, achieves a 29.61 length-controlled win rate on AlpacaEval 2 against GPT-4 and 7.35 on MT-Bench, surpassing the best 8B scale instruction-tuned linear RNN model. We also find that the distilled model has natural length extrapolation, showing almost perfect accuracy in the needle-in-a-haystack test at 20x the distillation length. Code and pre-trained checkpoints are open-sourced at [MambaInLlama](https://github.com/jxiw/MambaInLlama) for distillation and [SpeculativeMamba](https://github.com/itsdaniele/speculative\_mamba) for speculative decoding.