NeurIPS 2023 Papers

Neural network based computer vision systems are typically built on a backbone, a pretrained or randomly initialized feature extractor. Several years ago, the default option was an ImageNet-trained convolutional neural network. However, the recent past has seen the emergence of countless backbones pretrained using various algorithms and datasets. While this abundance of choice has led to performance increases for a range of systems, it is difficult for practitioners to make informed decisions about which backbone to choose. Battle of the Backbones (BoB) makes this choice easier by benchmarking a diverse suite of pretrained models, including vision-language models, those trained via self-supervised learning, and the Stable Diffusion backbone, across a diverse set of computer vision tasks ranging from classification to object detection to OOD generalization and more. Furthermore, BoB sheds light on promising directions for the research community to advance computer vision by illuminating strengths and weakness of existing approaches through a comprehensive analysis conducted on more than 1500 training runs. While vision transformers (ViTs) and self-supervised learning (SSL) are increasingly popular, we find that convolutional neural networks pretrained in a supervised fashion on large training sets still perform best on most tasks among the models we consider. Moreover, in apples-to-apples comparisons on the same architectures and similarly sized pretraining datasets, we find that SSL backbones are highly competitive, indicating that future works should perform SSL pretraining with advanced architectures and larger pretraining datasets. We release the raw results of our experiments along with code that allows researchers to put their own backbones through the gauntlet here: https://github.com/hsouri/Battle-of-the-Backbones.

The meteoric rise in the adoption of deep neural networks as computational models of vision has inspired efforts to ``align” these models with humans. One dimension of interest for alignment includes behavioral choices, but moving beyond characterizing choice patterns to capturing temporal aspects of visual decision-making has been challenging. Here, we sketch a general-purpose methodology to construct computational accounts of reaction times from a stimulus-computable, task-optimized model. Specifically, we introduce a novel metric leveraging insights from subjective logic theory summarizing evidence accumulation in recurrent vision models. We demonstrate that our metric aligns with patterns of human reaction times for stimulus manipulations across four disparate visual decision-making tasks spanning perceptual grouping, mental simulation, and scene categorization. This work paves the way for exploring the temporal alignment of model and human visual strategies in the context of various other cognitive tasks toward generating testable hypotheses for neuroscience. Links to the code and data can be found on the project page: https://serre-lab.github.io/rnn_rts_site/.

Program synthesis has been long studied with recent approaches focused on directly using the power of Large Language Models (LLMs) to generate code. Programming benchmarks, with curated synthesis problems and test-cases, are used to measure the performance of various LLMs on code synthesis. However, these test-cases can be limited in both quantity and quality for fully assessing the functional correctness of the generated code. Such limitation in the existing benchmarks begs the following question: In the era of LLMs, is the code generated really correct? To answer this, we propose EvalPlus – a code synthesis evaluation framework to rigorously benchmark the functional correctness of LLM-synthesized code. EvalPlus augments a given evaluation dataset with large amounts of test-cases newly produced by an automatic test input generator, powered by both LLM- and mutation-based strategies. While EvalPlus is general, we extend the test-cases of the popular HumanEval benchmark by 80x to build HumanEval+. Our extensive evaluation across 26 popular LLMs (e.g., GPT-4 and ChatGPT) demonstrates that HumanEval+ is able to catch significant amounts of previously undetected wrong code synthesized by LLMs, reducing the pass@k by up-to 19.3-28.9%. We also surprisingly found that test insufficiency can lead to mis-ranking. For example, both WizardCoder-CodeLlama and Phind-CodeLlama now outperform ChatGPT on HumanEval+, while none of them could on HumanEval. Our work not only indicates that prior popular code synthesis evaluation results do not accurately reflect the true performance of LLMs for code synthesis, but also opens up a new direction to improve such programming benchmarks through automated testing. We have open-sourced our tools, enhanced datasets as well as all LLM-generated code at https://github.com/evalplus/evalplus to facilitate and accelerate future LLM-for-code research.

Visual reasoning requires multimodal perception and commonsense cognition of the world. Recently, multiple vision-language models (VLMs) have been proposed with excellent commonsense reasoning ability in various domains. However, how to harness the collective power of these complementary VLMs is rarely explored. Existing methods like ensemble still struggle to aggregate these models with the desired higher-order communications. In this work, we propose Cola, a novel paradigm that coordinates multiple VLMs for visual reasoning. Our key insight is that a large language model (LLM) can efficiently coordinate multiple VLMs by facilitating natural language communication that leverages their distinct and complementary capabilities. Extensive experiments demonstrate that our instruction tuning variant, Cola-FT, achieves state-of-the-art performance on visual question answering (VQA), outside knowledge VQA, visual entailment, and visual spatial reasoning tasks. Moreover, we show that our in-context learning variant, Cola-Zero, exhibits competitive performance in zero and few-shot settings, without finetuning. Through systematic ablation studies and visualizations, we validate that a coordinator LLM indeed comprehends the instruction prompts as well as the separate functionalities of VLMs; it then coordinates them to enable impressive visual reasoning capabilities.

It has been well established that increasing scale in deep transformer networks leads to improved quality and performance. However, this increase in scale often comes with prohibitive increases in compute cost and inference latency. We introduce Alternating Updates (AltUp), a simple-to-implement method to increase a model's capacity without the computational burden. AltUp enables the widening of the learned representation, i.e., the token embedding, while only incurring a negligible increase in latency. AltUp achieves this by working on a subblock of the widened representation at each layer and using a predict-and-correct mechanism to update the inactivated blocks. We present extensions of AltUp, such as its applicability to the sequence dimension, and demonstrate how AltUp can be synergistically combined with existing approaches, such as Sparse Mixture-of-Experts models, to obtain efficient models with even higher capacity. Our experiments on benchmark transformer models and language tasks demonstrate the consistent effectiveness of AltUp on a diverse set of scenarios. Notably, on SuperGLUE and SQuAD benchmarks, AltUp enables up to $87\%$ speedup relative to the dense baselines at the same accuracy.

Large Language Models (LLMs) with strong abilities in natural language processing tasks have emerged and have been applied in various kinds of areas such as science, finance and software engineering. However, the capability of LLMs to advance the field of chemistry remains unclear. In this paper, rather than pursuing state-of-the-art performance, we aim to evaluate capabilities of LLMs in a wide range of tasks across the chemistry domain. We identify three key chemistry-related capabilities including understanding, reasoning and explaining to explore in LLMs and establish a benchmark containing eight chemistry tasks. Our analysis draws on widely recognized datasets facilitating a broad exploration of the capacities of LLMs within the context of practical chemistry. Five LLMs (GPT-4,GPT-3.5, Davinci-003, Llama and Galactica) are evaluated for each chemistry task in zero-shot and few-shot in-context learning settings with carefully selected demonstration examples and specially crafted prompts. Our investigation found that GPT-4 outperformed other models and LLMs exhibit different competitive levels in eight chemistry tasks. In addition to the key findings from the comprehensive benchmark analysis, our work provides insights into the limitation of current LLMs and the impact of in-context learning settings on LLMs’ performance across various chemistry tasks. The code and datasets used in this study are available at https://github.com/ChemFoundationModels/ChemLLMBench.

In recent years, most exciting inputs (MEIs) synthesized from encoding models of neuronal activity have become an established method for studying tuning properties of biological and artificial visual systems. However, as we move up the visual hierarchy, the complexity of neuronal computations increases. Consequently, it becomes more challenging to model neuronal activity, requiring more complex models. In this study, we introduce a novel readout architecture inspired by the mechanism of visual attention. This new architecture, which we call attention readout, together with a data-driven convolutional core outperforms previous task-driven models in predicting the activity of neurons in macaque area V4. However, as our predictive network becomes deeper and more complex, synthesizing MEIs via straightforward gradient ascent (GA) can struggle to produce qualitatively good results and overfit to idiosyncrasies of a more complex model, potentially decreasing the MEI's model-to-brain transferability. To solve this problem, we propose a diffusion-based method for generating MEIs via Energy Guidance (EGG). We show that for models of macaque V4, EGG generates single neuron MEIs that generalize better across varying model architectures than the state-of-the-art GA, while at the same time reducing computational costs by a factor of 4.7x, facilitating experimentally challenging closed-loop experiments. Furthermore, EGG diffusion can be used to generate other neurally exciting images, like most exciting naturalistic images that are on par with a selection of highly activating natural images, or image reconstructions that generalize better across architectures. Finally, EGG is simple to implement, requires no retraining of the diffusion model, and can easily be generalized to provide other characterizations of the visual system, such as invariances. Thus, EGG provides a general and flexible framework to study the coding properties of the visual system in the context of natural images.

Recently, learned image compression (LIC) has garnered increasing interest with its rapidly improving performance surpassing conventional codecs. A key ingredient of LIC is a hyperprior-based entropy model, where the underlying joint probability of the latent image features is modeled as a product of Gaussian distributions from each latent element. Since latents from the actual images are not spatially independent, autoregressive (AR) context based entropy models were proposed to handle the discrepancy between the assumed distribution and the actual distribution. Though the AR-based models have proven effective, the computational complexity is significantly increased due to the inherent sequential nature of the algorithm. In this paper, we present a novel alternative to the AR-based approach that can provide a significantly better trade-off between performance and complexity. To minimize the discrepancy, we introduce a correlation loss that forces the latents to be spatially decorrelated and better fitted to the independent probability model. Our correlation loss is proved to act as a general plug-in for the hyperprior (HP) based learned image compression methods. The performance gain from our correlation loss is ‘free’ in terms of computation complexity for both inference time and decoding time. To our knowledge, our method gives the best trade-off between the complexity and performance: combined with the Checkerboard-CM, it attains **90%** and when combined with ChARM-CM, it attains **98%** of the AR-based BD-Rate gains yet is around **50 times** and **30 times** faster than AR-based methods respectively

Multimodal learning assumes all modality combinations of interest are available during training to learn cross-modal correspondences. In this paper, we challenge this modality-complete assumption for multimodal learning and instead strive for generalization to unseen modality combinations during inference. We pose the problem of unseen modality interaction and introduce a first solution. It exploits a module that projects the multidimensional features of different modalities into a common space with rich information preserved. This allows the information to be accumulated with a simple summation operation across available modalities. To reduce overfitting to less discriminative modality combinations during training, we further improve the model learning with pseudo-supervision indicating the reliability of a modality’s prediction. We demonstrate that our approach is effective for diverse tasks and modalities by evaluating it for multimodal video classification, robot state regression, and multimedia retrieval. Project website: https://xiaobai1217.github.io/Unseen-Modality-Interaction/.

Effectively modeling long spatiotemporal sequences is challenging due to the need to model complex spatial correlations and long-range temporal dependencies simultaneously. ConvLSTMs attempt to address this by updating tensor-valued states with recurrent neural networks, but their sequential computation makes them slow to train. In contrast, Transformers can process an entire spatiotemporal sequence, compressed into tokens, in parallel. However, the cost of attention scales quadratically in length, limiting their scalability to longer sequences. Here, we address the challenges of prior methods and introduce convolutional state space models (ConvSSM) that combine the tensor modeling ideas of ConvLSTM with the long sequence modeling approaches of state space methods such as S4 and S5. First, we demonstrate how parallel scans can be applied to convolutional recurrences to achieve subquadratic parallelization and fast autoregressive generation. We then establish an equivalence between the dynamics of ConvSSMs and SSMs, which motivates parameterization and initialization strategies for modeling long-range dependencies. The result is ConvS5, an efficient ConvSSM variant for long-range spatiotemporal modeling. ConvS5 significantly outperforms Transformers and ConvLSTM on a long horizon Moving-MNIST experiment while training $3\times$ faster than ConvLSTM and generating samples $400\times$ faster than Transformers. In addition, ConvS5 matches or exceeds the performance of state-of-the-art methods on challenging DMLab, Minecraft and Habitat prediction benchmarks and enables new directions for modeling long spatiotemporal sequences.

Modern language models excel at integrating across long temporal scales needed to encode linguistic meaning and show non-trivial similarities to biological neural systems. Prior work suggests that human brain responses to language exhibit hierarchically organized "integration windows" that substantially constrain the overall influence of an input token (e.g., a word) on the neural response. However, little prior work has attempted to use integration windows to characterize computations in large language models (LLMs). We developed a simple word-swap procedure for estimating integration windows from black-box language models that does not depend on access to gradients or knowledge of the model architecture (e.g., attention weights). Using this method, we show that trained LLMs exhibit stereotyped integration windows that are well-fit by a convex combination of an exponential and a power-law function, with a partial transition from exponential to power-law dynamics across network layers. We then introduce a metric for quantifying the extent to which these integration windows vary with structural boundaries (e.g., sentence boundaries), and using this metric, we show that integration windows become increasingly yoked to structure at later network layers. None of these findings were observed in an untrained model, which as expected integrated uniformly across its input. These results suggest that LLMs learn to integrate information in natural language using a stereotyped pattern: integrating across position-yoked, exponential windows at early layers, followed by structure-yoked, power-law windows at later layers. The methods we describe in this paper provide a general-purpose toolkit for understanding temporal integration in language models, facilitating cross-disciplinary research at the intersection of biological and artificial intelligence.

Successful detection of Out-of-Distribution (OoD) data is becoming increasingly important to ensure safe deployment of neural networks. One of the main challenges in OoD detection is that neural networks output overconfident predictions on OoD data, make it difficult to determine OoD-ness of data solely based on their predictions. Outlier exposure addresses this issue by introducing an additional loss that encourages low-confidence predictions on OoD data during training. While outlier exposure has shown promising potential in improving OoD detection performance, all previous studies on outlier exposure have been limited to utilizing visual outliers. Drawing inspiration from the recent advancements in vision-language pre-training, this paper venture out to the uncharted territory of textual outlier exposure. First, we uncover the benefits of using textual outliers by replacing real or virtual outliers in the image-domain with textual equivalents. Then, we propose various ways of generating preferable textual outliers. Our extensive experiments demonstrate that generated textual outliers achieve competitive performance on large-scale OoD and hard OoD benchmarks. Furthermore, we conduct empirical analyses of textual outliers to provide primary criteria for designing advantageous textual outliers: near-distribution, descriptiveness, and inclusion of visual semantics.