Linear Recurrent Neural Networks (LRNNs) such as Mamba, RWKV, GLA, mLSTM, and DeltaNet have emerged as efficient alternatives to Transformers for long sequences. However, both Transformers and LRNNs struggle to perform state-tracking, which may impair performance in tasks such as code evaluation. In one forward pass, current architectures are unable to solve even parity, the simplest state-tracking task, which non-linear RNNs can handle effectively. Recently, Sarrof et al. (2024) demonstrated that the failure of LRNNs like Mamba to solve parity stems from restricting the value range of their diagonal state-transition matrices to $[0, 1]$ and that incorporating negative values can resolve this issue. We extend this result to non-diagonal LRNNs such as DeltaNet. We prove that finite precision LRNNs with state-transition matrices having only positive eigenvalues cannot solve parity, while non-triangular matrices are needed to count modulo $3$. Notably, we also prove that LRNNs can learn any regular language when their state-transition matrices are products of identity minus vector outer product matrices, each with eigenvalues in the range $[-1, 1]$. Our experiments confirm that extending the eigenvalue range of Mamba and DeltaNet to include negative values not only enables them to solve parity but consistently improves their performance on state-tracking tasks. We also show that state-tracking enabled LRNNs can be pretrained stably and efficiently at scale (1.3B parameters), achieving competitive performance on language modeling and showing promise on code and math tasks.

Pareto front profiling in multi-objective optimization (MOO), i.e., finding a diverse set of Pareto optimal solutions, is challenging, especially with expensive objectives that require training a neural network. Typically, in MOO for neural architecture search (NAS), we aim to balance performance and hardware metrics across devices. Prior NAS approaches simplify this task by incorporating hardware constraints into the objective function, but profiling the Pareto front necessitates a computationally expensive search for each constraint. In this work, we propose a novel NAS algorithm that encodes user preferences to trade-off performance and hardware metrics, yielding representative and diverse architectures across multiple devices in just a single search run. To this end, we parameterize the joint architectural distribution across devices and multiple objectives via a hypernetwork that can be conditioned on hardware features and preference vectors, enabling zero-shot transferability to new devices. Extensive experiments involving up to 19 hardware devices and 3 different objectives demonstrate the effectiveness and scalability of our method. Finally, we show that, without any additional costs, our method outperforms existing MOO NAS methods across a broad range of qualitatively different search spaces and datasets, including MobileNetV3 on ImageNet-1k, an encoder-decoder transformer space for machine translation and a decoder-only space for language modelling.

Linear Recurrent Neural Networks (linear RNNs) have emerged as competitive alternatives to Transformers for sequence modeling, offering efficient training and linear-time inference. However, existing architectures face a fundamental trade-off between expressivity and efficiency, dictated by the structure of their state-transition matrices. Diagonal matrices, used in models such as Mamba, GLA, or mLSTM, yield fast runtime but have limited expressivity. To address this, recent architectures such as DeltaNet and RWKV-7 adopted a diagonal plus rank-1 structure, which allows simultaneous token and channel mixing, improving associative recall and, as recently shown, state-tracking when allowing negative eigenvalues in the state-transition matrices. Building on the interpretation of DeltaNet's recurrence as performing one step of online gradient descent per token on an associative recall loss, we introduce DeltaProduct, which instead takes multiple ($n_h$) steps per token. This naturally leads to diagonal plus rank-$n_h$ state-transition matrices, formed as products of $n_h$ generalized Householder transformations, providing a tunable mechanism to balance expressivity and efficiency. We provide a detailed theoretical characterization of the state-tracking capability of DeltaProduct in finite precision and how it improves by increasing $n_h$. Our extensive experiments demonstrate that DeltaProduct outperforms DeltaNet in both state-tracking and language modeling, while also showing significantly improved length extrapolation capabilities.

The increasing size of language models necessitates a thorough analysis across multiple dimensions to assess trade-offs among crucial hardware metrics such as latency, energy consumption, GPU memory usage, and performance. Identifying optimal model configurations under specific hardware constraints is becoming essential but remains challenging due to the computational load of exhaustive training and evaluation on multiple devices. To address this, we introduce HW-GPT-Bench, a hardware-aware benchmark that utilizes surrogate predictions to approximate various hardware metrics across 13 devices of architectures in the GPT-2 family, with architectures containing up to 1.55B parameters. Our surrogates, via calibrated predictions and reliable uncertainty estimates, faithfully model the heteroscedastic noise inherent in the energy and latency measurements. To estimate perplexity, we employ weight-sharing techniques from Neural Architecture Search (NAS), inheriting pretrained weights from the largest GPT-2 model. Finally, we demonstrate the utility of HW-GPT-Bench by simulating optimization trajectories of various multi-objective optimization algorithms in just a few seconds.

Abstract The success of deep learning in recent years has lead to a rising demand for neural network architecture engineering. As a consequence, neural architecture search (NAS), which aims at automatically designing neural network architectures in a data-driven manner rather than manually, has evolved as a popular field of research. With the advent of weight sharing strategies across architectures, NAS has become applicable to a much wider range of problems. In particular, there are now many publications for dense prediction tasks in computer vision that require pixel-level predictions, such as semantic segmentation or object detection. These tasks come with novel challenges, such as higher memory footprints due to high-resolution data, learning multi-scale representations, longer training times, and more complex and larger neural architectures. In this manuscript, we provide an overview of NAS for dense prediction tasks by elaborating on these novel challenges and surveying ways to address them to ease future research and application of existing methods to novel problems.

The past few years have seen the development of many benchmarks for Neural Architecture Search (NAS), fueling rapid progress in NAS research. However, recent work, which shows that good hyperparameter settings can be more important than using the best architecture, calls for a shift in focus towards Joint Architecture and Hyperparameter Search (JAHS). Therefore, we present JAHS-Bench-201, the first collection of surrogate benchmarks for JAHS, built to also facilitate research on multi-objective, cost-aware and (multi) multi-fidelity optimization algorithms. To the best of our knowledge, JAHS-Bench-201 is based on the most extensive dataset of neural network performance data in the public domain. It is composed of approximately 161 million data points and 20 performance metrics for three deep learning tasks, while featuring a 14-dimensional search and fidelity space that extends the popular NAS-Bench-201 space. With JAHS-Bench-201, we hope to democratize research on JAHS and lower the barrier to entry of an extremely compute intensive field, e.g., by reducing the compute time to run a JAHS algorithm from 5 days to only a few seconds.

Zero-cost proxies (ZC proxies) are a recent architecture performance prediction technique aiming to significantly speed up algorithms for neural architecture search (NAS). Recent work has shown that these techniques show great promise, but certain aspects, such as evaluating and exploiting their complementary strengths, are under-studied. In this work, we create NAS-Bench-Suite: we evaluate 13 ZC proxies across 28 tasks, creating by far the largest dataset (and unified codebase) for ZC proxies, enabling orders-of-magnitude faster experiments on ZC proxies, while avoiding confounding factors stemming from different implementations. To demonstrate the usefulness of NAS-Bench-Suite, we run a large-scale analysis of ZC proxies, including a bias analysis, and the first information-theoretic analysis which concludes that ZC proxies capture substantial complementary information. Motivated by these findings, we present a procedure to improve the performance of ZC proxies by reducing biases such as cell size, and we also show that incorporating all 13 ZC proxies into the surrogate models used by NAS algorithms can improve their predictive performance by up to 42%. Our code and datasets are available at https://github.com/automl/naslib/tree/zerocost.

Neural architecture search (NAS) has recently been addressed from various directions, including discrete, sampling-based methods and efficient differentiable approaches. While the former are notoriously expensive, the latter suffer from imposing strong constraints on the search space. Architecture optimization from a learned embedding space for example through graph neural network based variational autoencoders builds a middle ground and leverages advantages from both sides. Such approaches have recently shown good performance on several benchmarks. Yet, their stability and predictive power heavily depends on their capacity to reconstruct networks from the embedding space. In this paper, we propose a two-sided variational graph autoencoder, which allows to smoothly encode and accurately reconstruct neural architectures from various search spaces. We evaluate the proposed approach on neural architectures defined by the ENAS approach, the NAS-Bench-101 and the NAS-Bench-201 search space and show that our smooth embedding space allows to directly extrapolate the performance prediction to architectures outside the seen domain (e.g. with more operations). Thus, it facilitates to predict good network architectures even without expensive Bayesian optimization or reinforcement learning.

Early methods in the rapidly developing field of neural architecture search (NAS) required fully training thousands of neural networks. To reduce this extreme computational cost, dozens of techniques have since been proposed to predict the final performance of neural architectures. Despite the success of such performance prediction methods, it is not well-understood how different families of techniques compare to one another, due to the lack of an agreed-upon evaluation metric and optimization for different constraints on the initialization time and query time. In this work, we give the first large-scale study of performance predictors by analyzing 31 techniques ranging from learning curve extrapolation, to weight-sharing, to supervised learning, to zero-cost proxies. We test a number of correlation- and rank-based performance measures in a variety of settings, as well as the ability of each technique to speed up predictor-based NAS frameworks. Our results act as recommendations for the best predictors to use in different settings, and we show that certain families of predictors can be combined to achieve even better predictive power, opening up promising research directions. We release our code, featuring a library of 31 performance predictors.

Ensembles of neural networks achieve superior performance compared to standalone networks in terms of accuracy, uncertainty calibration and robustness to dataset shift. Deep ensembles, a state-of-the-art method for uncertainty estimation, only ensemble random initializations of a fixed architecture. Instead, we propose two methods for automatically constructing ensembles with varying architectures, which implicitly trade-off individual architectures’ strengths against the ensemble’s diversity and exploit architectural variation as a source of diversity. On a variety of classification tasks and modern architecture search spaces, we show that the resulting ensembles outperform deep ensembles not only in terms of accuracy but also uncertainty calibration and robustness to dataset shift. Our further analysis and ablation studies provide evidence of higher ensemble diversity due to architectural variation, resulting in ensembles that can outperform deep ensembles, even when having weaker average base learners. To foster reproducibility, our code is available: https://github.com/automl/nes

This paper reports the results and post-challenge analyses of ChaLearn's AutoDL challenge series, which helped sorting out a profusion of AutoML solutions for Deep Learning (DL) that had been introduced in a variety of settings, but lacked fair comparisons. All input data modalities (time series, images, videos, text, tabular) were formatted as tensors and all tasks were multi-label classification problems. Code submissions were executed on hidden tasks, with limited time and computational resources, pushing solutions that get results quickly. In this setting, DL methods dominated, though popular Neural Architecture Search (NAS) was impractical. Solutions relied on fine-tuned pre-trained networks, with architectures matching data modality. Post-challenge tests did not reveal improvements beyond the imposed time limit. While no component is particularly original or novel, a high level modular organization emerged featuring a "meta-learner", "data ingestor", "model selector", "model/learner", and "evaluator". This modularity enabled ablation studies, which revealed the importance of (off-platform) meta-learning, ensembling, and efficient data management. Experiments on heterogeneous module combinations further confirm the (local) optimality of the winning solutions. Our challenge legacy includes an ever-lasting benchmark (http://autodl.chalearn.org), the open-sourced code of the winners, and a free "AutoDL self-service."

Much research work in computer vision is being spent on optimizing existing network architectures to obtain a few more percentage points on benchmarks. Recent AutoML approaches promise to relieve us from this effort. However, they are mainly designed for comparatively small-scale classification tasks. In this work, we show how to use and extend existing AutoML techniques to efficiently optimize large-scale U-Net-like encoder-decoder architectures. In particular, we leverage gradient-based neural architecture search and Bayesian optimization for hyperparameter search. The resulting optimization does not require a large-scale compute cluster. We show results on disparity estimation that clearly outperform the manually optimized baseline and reach state-of-the-art performance.