NeurIPS 2023 Papers

The application of deep learning to nursing procedure activity understanding has the potential to greatly enhance the quality and safety of nurse-patient interactions. By utilizing the technique, we can facilitate training and education, improve quality control, and enable operational compliance monitoring. However, the development of automatic recognition systems in this field is currently hindered by the scarcity of appropriately labeled datasets. The existing video datasets pose several limitations: 1) these datasets are small-scale in size to support comprehensive investigations of nursing activity; 2) they primarily focus on single procedures, lacking expert-level annotations for various nursing procedures and action steps; and 3) they lack temporally localized annotations, which prevents the effective localization of targeted actions within longer video sequences. To mitigate these limitations, we propose NurViD, a large video dataset with expert-level annotation for nursing procedure activity understanding. NurViD consists of over 1.5k videos totaling 144 hours, making it approximately four times longer than the existing largest nursing activity datasets. Notably, it encompasses 51 distinct nursing procedures and 177 action steps, providing a much more comprehensive coverage compared to existing datasets that primarily focus on limited procedures. To evaluate the efficacy of current deep learning methods on nursing activity understanding, we establish three benchmarks on NurViD: procedure recognition on untrimmed videos, procedure and action recognition on trimmed videos, and action detection. Our benchmark and code will be available at https://github.com/minghu0830/NurViD-benchmark.

Recently, growing interest has been aroused in extending the multimodal capability of large language models (LLMs), e.g., vision-language (VL) learning, which is regarded as the next milestone of artificial general intelligence. However, existing solutions are prohibitively expensive, which not only need to optimize excessive parameters, but also require another large-scale pre-training before VL instruction tuning. In this paper, we propose a novel and affordable solution for the effective VL adaption of LLMs, called Mixture-of-Modality Adaptation (MMA). Instead of using large neural networks to connect the image encoder and LLM, MMA adopts lightweight modules, i.e., adapters, to bridge the gap between LLMs and VL tasks, which also enables the joint optimization of the image and language models. Meanwhile, MMA is also equipped with a routing algorithm to help LLMs achieve an automatic shift between single- and multi-modal instructions without compromising their ability of natural language understanding. To validate MMA, we apply it to a recent LLM called LLaMA and term this formed large vision-language instructed model as LaVIN. To validate MMA and LaVIN, we conduct extensive experiments under two setups, namely multimodal science question answering and multimodal dialogue. The experimental results not only demonstrate the competitive performance and the superior training efficiency of LaVIN than existing multimodal LLMs, but also confirm its great potential as a general-purpose chatbot. More importantly, the actual expenditure of LaVIN is extremely cheap, e.g., only 1.4 training hours with 3.8M trainable parameters, greatly confirming the effectiveness of MMA. Our code is anonymously released at: https://anonymous.4open.science/r/LaVIN--1067.

This paper introduces a novel task called Cross Modal Generalization (CMG), which addresses the challenge of learning a unified discrete representation from paired multimodal data during pre-training. Then in downstream tasks, the model can achieve zero-shot generalization ability in other modalities when only one modal is labeled. Existing approaches in multimodal representation learning focus more on coarse-grained alignment or rely on the assumption that information from different modalities is completely aligned, which is impractical in real-world scenarios. To overcome this limitation, we propose \textbf{Uni-Code}, which contains two key contributions: the Dual Cross-modal Information Disentangling (DCID) module and the Multi-Modal Exponential Moving Average (MM-EMA). These methods facilitate bidirectional supervision between modalities and align semantically equivalent information in a shared discrete latent space, enabling fine-grained unified representation of multimodal sequences. During pre-training, we investigate various modality combinations, including audio-visual, audio-text, and the tri-modal combination of audio-visual-text. Extensive experiments on various downstream tasks, i.e., cross-modal event classification, localization, cross-modal retrieval, query-based video segmentation, and cross-dataset event localization, demonstrate the effectiveness of our proposed methods. The code is available at https://github.com/haihuangcode/CMG.

Unsupervised video-based object-centric learning is a promising avenue to learn structured representations from large, unlabeled video collections, but previous approaches have only managed to scale to real-world datasets in restricted domains. Recently, it was shown that the reconstruction of pre-trained self-supervised features leads to object-centric representations on unconstrained real-world image datasets. Building on this approach, we propose a novel way to use such pre-trained features in the form of a temporal feature similarity loss. This loss encodes semantic and temporal correlations between image patches and is a natural way to introduce a motion bias for object discovery. We demonstrate that this loss leads to state-of-the-art performance on the challenging synthetic MOVi datasets. When used in combination with the feature reconstruction loss, our model is the first object-centric video model that scales to unconstrained video datasets such as YouTube-VIS. https://martius-lab.github.io/videosaur/

We introduce new methods for 1) accelerating and 2) stabilizing training for large language-vision models. 1) For acceleration, we introduce SwitchBack, a linear layer for int8 quantized training which provides a speed-up of 13-25% while matching the performance of bfloat16 training within 0.1 percentage points for the 1B parameter CLIP ViT-Huge---the largest int8 training to date. Our main focus is int8 as GPU support for float8 is rare, though we also analyze float8 training through simulation. While SwitchBack proves effective for float8, we show that standard techniques are also successful if the network is trained and initialized so that large feature magnitudes are discouraged, which we accomplish via layer-scale initialized with zeros. 2) For stability, we analyze loss spikes and find they consistently occur 1-8 iterations after the squared gradients become under-estimated by their AdamW second moment estimator. As a result, we recommend an AdamW-Adafactor hybrid which avoids loss spikes when training a CLIP ViT-Huge model and outperforms gradient clipping at the scales we test.

Recent works such as VisProg and ViperGPT have smartly composed foundation models for visual reasoning—using large language models (LLMs) to produce programs that can be executed by pre-trained vision-language models. However, they operate in limited domains, such as 2D images, not fully exploiting the generalization of language: abstract concepts like “*left*” can also be grounded in 3D, temporal, and action data, as in moving to your *left*. This limited generalization stems from these inference-only methods’ inability to learn or adapt pre-trained models to a new domain. We propose the **L**ogic-**E**nhanced **F**ounda**T**ion Model (**LEFT**), a unified framework that *learns* to ground and reason with concepts across domains with a differentiable, domain-independent, first-order logic-based program executor. LEFT has an LLM interpreter that outputs a program represented in a general, logic-based reasoning language, which is shared across all domains and tasks. LEFT’s executor then executes the program with trainable domain-specific grounding modules. We show that LEFT flexibly learns concepts in four domains: 2D images, 3D scenes, human motions, and robotic manipulation. It exhibits strong reasoning ability in a wide variety of tasks, including those that are complex and not seen during training, and can be easily applied to new domains.

Vision and text have been fully explored in contemporary video-text foundational models, while other modalities such as audio and subtitles in videos have not received sufficient attention. In this paper, we resort to establish connections between multi-modality video tracks, including Vision, Audio, and Subtitle, and Text by exploring an automatically generated large-scale omni-modality video caption dataset called VAST-27M. Specifically, we first collect 27 million open-domain video clips and separately train a vision and an audio captioner to generate vision and audio captions. Then, we employ an off-the-shelf Large Language Model (LLM) to integrate the generated captions, together with subtitles and instructional prompts into omni-modality captions. Based on the proposed VAST-27M dataset, we train an omni-modality video-text foundational model named VAST, which can perceive and process vision, audio, and subtitle modalities from video, and better support various tasks including vision-text, audio-text, and multi-modal video-text tasks (retrieval, captioning and QA). Extensive experiments have been conducted to demonstrate the effectiveness of our proposed VAST-27M corpus and VAST foundation model. VAST achieves 22 new state-of-the-art results on various cross-modality benchmarks.

Due to its promising performance, deep hashing has become a prevalent method for approximate nearest neighbors search (ANNs). However, most of current deep hashing methods are validated on relatively small-scale datasets, leaving potential threats when are applied to large-scale real-world scenarios. Specifically, they can be constrained either by the computational cost due to the large number of training categories and samples, or unsatisfactory accuracy. To tackle those issues, we propose a novel deep hashing framework based on product quantization (PQ). It uses a softmax-based differentiable PQ branch to learn a set of predefined PQ codes of the classes. Our method is easy to implement, does not involve large-scale matrix operations, and learns highly discriminate compact codes. We validate our method on multiple large-scaled datasets, including ImageNet100, ImageNet1K, and Glint360K, where the category size scales from 100 to 360K and sample number scales from 10K to 17 million, respectively. Extensive experiments demonstrate the superiority of our method. Code is available at https://github.com/yuleung/FPPQ.

Modern deep networks are highly complex and their inferential outcome very hard to interpret. This is a serious obstacle to their transparent deployment in safety-critical or bias-aware applications. This work contributes to *post-hoc* interpretability, and specifically Network Dissection. Our goal is to present a framework that makes it easier to *discover* the individual functionality of each neuron in a network trained on a vision task; discovery is performed in terms of textual description generation. To achieve this objective, we leverage: (i) recent advances in multimodal vision-text models and (ii) network layers founded upon the novel concept of stochastic local competition between linear units. In this setting, only a *small subset* of layer neurons are activated *for a given input*, leading to extremely high activation sparsity (as low as only $\approx 4\%$). Crucially, our proposed method infers (sparse) neuron activation patterns that enables the neurons to activate/specialize to inputs with specific characteristics, diversifying their individual functionality. This capacity of our method supercharges the potential of dissection processes: human understandable descriptions are generated only for the very few active neurons, thus facilitating the direct investigation of the network's decision process. As we experimentally show, our approach: (i) yields Vision Networks that retain or improve classification performance, and (ii) realizes a principled framework for text-based description and examination of the generated neuronal representations.

Multimodal fusion has become an important research technique in neuroscience that completes downstream tasks by extracting complementary information from multiple modalities. Existing multimodal research on brain networks mainly focuses on two modalities, structural connectivity (SC) and functional connectivity (FC). Recently, extensive literature has shown that the relationship between SC and FC is complex and not a simple one-to-one mapping. The coupling of structure and function at the regional level is heterogeneous. However, all previous studies have neglected the modal regional heterogeneity between SC and FC and fused their representations via "simple patterns", which are inefficient ways of multimodal fusion and affect the overall performance of the model. In this paper, to alleviate the issue of regional heterogeneity of multimodal brain networks, we propose a novel Regional Heterogeneous multimodal Brain networks Fusion Strategy (RH-BrainFS). Briefly, we introduce a brain subgraph networks module to extract regional characteristics of brain networks, and further use a new transformer-based fusion bottleneck module to alleviate the issue of regional heterogeneity between SC and FC. To the best of our knowledge, this is the first paper to explicitly state the issue of structural-functional modal regional heterogeneity and to propose a solution. Extensive experiments demonstrate that the proposed method outperforms several state-of-the-art methods in a variety of neuroscience tasks.

Learnt deep neural networks for image super-resolution fail easily if the assumed degradation model in training mismatches that of the real degradation source at the inference stage. Instead of attempting to exhaust all degradation variants in simulation, which is unwieldy and impractical, we propose a novel adversarial neural degradation (AND) model that can, when trained in conjunction with a deep restoration neural network under a minmax criterion, generate a wide range of highly nonlinear complex degradation effects without any explicit supervision. The AND model has a unique advantage over the current state of the art in that it can generalize much better to unseen degradation variants and hence deliver significantly improved restoration performance on real-world images.

Though modern neural networks have achieved impressive performance in both vision and language tasks, we know little about the functions that they implement. One possibility is that neural networks implicitly break down complex tasks into subroutines, implement modular solutions to these subroutines, and compose them into an overall solution to a task --- a property we term structural compositionality. Another possibility is that they may simply learn to match new inputs to learned templates, eliding task decomposition entirely. Here, we leverage model pruning techniques to investigate this question in both vision and language across a variety of architectures, tasks, and pretraining regimens. Our results demonstrate that models oftentimes implement solutions to subroutines via modular subnetworks, which can be ablated while maintaining the functionality of other subnetworks. This suggests that neural networks may be able to learn compositionality, obviating the need for specialized symbolic mechanisms.

Most recurrent neural networks (RNNs) do not include a fundamental constraint of real neural circuits: Dale's Law, which implies that neurons must be excitatory (E) or inhibitory (I). Dale's Law is generally absent from RNNs because simply partitioning a standard network's units into E and I populations impairs learning. However, here we extend a recent feedforward bio-inspired EI network architecture, named Dale's ANNs, to recurrent networks, and demonstrate that good performance is possible while respecting Dale's Law. This begs the question: What makes some forms of EI network learn poorly and others learn well? And, why does the simple approach of incorporating Dale's Law impair learning? Historically the answer was thought to be the sign constraints on EI network parameters, and this was a motivation behind Dale's ANNs. However, here we show the spectral properties of the recurrent weight matrix at initialisation are more impactful on network performance than sign constraints. We find that simple EI partitioning results in a singular value distribution that is multimodal and dispersed, whereas standard RNNs have an unimodal, more clustered singular value distribution, as do recurrent Dale's ANNs. We also show that the spectral properties and performance of partitioned EI networks are worse for small networks with fewer I units, and we present normalised SVD entropy as a measure of spectrum pathology that correlates with performance. Overall, this work sheds light on a long-standing mystery in neuroscience-inspired AI and computational neuroscience, paving the way for greater alignment between neural networks and biology.

Decentralized learning enables the training of deep learning models over large distributed datasets generated at different locations, without the need for a central server. However, in practical scenarios, the data distribution across these devices can be significantly different, leading to a degradation in model performance. In this paper, we focus on designing a decentralized learning algorithm that is less susceptible to variations in data distribution across devices. We propose Global Update Tracking (GUT), a novel tracking-based method that aims to mitigate the impact of heterogeneous data in decentralized learning without introducing any communication overhead. We demonstrate the effectiveness of the proposed technique through an exhaustive set of experiments on various Computer Vision datasets (CIFAR-10, CIFAR-100, Fashion MNIST, and ImageNette), model architectures, and network topologies. Our experiments show that the proposed method achieves state-of-the-art performance for decentralized learning on heterogeneous data via a 1-6% improvement in test accuracy compared to other existing techniques.

Despite their immense success as a model of macaque visual cortex, deep convolutional neural networks (CNNs) have struggled to predict activity in visual cortex of the mouse, which is thought to be strongly dependent on the animal’s behavioral state. Furthermore, most computational models focus on predicting neural responses to static images presented under head fixation, which are dramatically different from the dynamic, continuous visual stimuli that arise during movement in the real world. Consequently, it is still unknown how natural visual input and different behavioral variables may integrate over time to generate responses in primary visual cortex (V1). To address this, we introduce a multimodal recurrent neural network that integrates gaze-contingent visual input with behavioral and temporal dynamics to explain V1 activity in freely moving mice. We show that the model achieves state-of-the-art predictions of V1 activity during free exploration and demonstrate the importance of each component in an extensive ablation study. Analyzing our model using maximally activating stimuli and saliency maps, we reveal new insights into cortical function, including the prevalence of mixed selectivity for behavioral variables in mouse V1. In summary, our model offers a comprehensive deep-learning framework for exploring the computational principles underlying V1 neurons in freely-moving animals engaged in natural behavior.

Despite recent success in large language model (LLM) reasoning, LLMs struggle with hierarchical multi-step reasoning tasks like generating complex programs. For these tasks, humans often start with a high-level algorithmic design and implement each part gradually. We introduce Parsel, a framework enabling automatic implementation and validation of complex algorithms with code LLMs. With Parsel, we automatically decompose algorithmic tasks into hierarchical natural language function descriptions and then search over combinations of possible function implementations using tests. We show that Parsel can be used across domains requiring hierarchical reasoning, including program synthesis and robotic planning. We find that, using Parsel, LLMs solve more competition-level problems in the APPS dataset, resulting in pass rates over 75\% higher than prior results from directly sampling AlphaCode and Codex, while often using a smaller sample budget. Moreover, with automatically generated tests, we find that Parsel can improve the state-of-the-art pass@1 performance on HumanEval from 67\% to 85\%. We also find that LLM-generated robotic plans using Parsel are more than twice as likely to be considered accurate than directly generated plans. Lastly, we explore how Parsel addresses LLM limitations and discuss how Parsel may be useful for human programmers. We release our code at https://github.com/ezelikman/parsel.

The stunning qualitative improvement of text-to-image models has led to their widespread attention and adoption. However, we lack a comprehensive quantitative understanding of their capabilities and risks. To fill this gap, we introduce a new benchmark, Holistic Evaluation of Text-to-Image Models (HEIM). Whereas previous evaluations focus mostly on image-text alignment and image quality, we identify 12 aspects, including text-image alignment, image quality, aesthetics, originality, reasoning, knowledge, bias, toxicity, fairness, robustness, multilinguality, and efficiency. We curate 62 scenarios encompassing these aspects and evaluate 26 state-of-the-art text-to-image models on this benchmark. Our results reveal that no single model excels in all aspects, with different models demonstrating different strengths. We release the generated images and human evaluation results for full transparency at https://crfm.stanford.edu/heim/latest and the code at https://github.com/stanford-crfm/helm, which is integrated with the HELM codebase

StressID is a new dataset specifically designed for stress identification from unimodal and multimodal data. It contains videos of facial expressions, audio recordings, and physiological signals. The video and audio recordings are acquired using an RGB camera with an integrated microphone. The physiological data is composed of electrocardiography (ECG), electrodermal activity (EDA), and respiration signals that are recorded and monitored using a wearable device. This experimental setup ensures a synchronized and high-quality multimodal data col- lection. Different stress-inducing stimuli, such as emotional video clips, cognitive tasks including mathematical or comprehension exercises, and public speaking scenarios, are designed to trigger a diverse range of emotional responses. The final dataset consists of recordings from 65 participants who performed 11 tasks, as well as their ratings of perceived relaxation, stress, arousal, and valence levels. StressID is one of the largest datasets for stress identification that features three different sources of data and varied classes of stimuli, representing more than 39 hours of annotated data in total. StressID offers baseline models for stress classification including a cleaning, feature extraction, and classification phase for each modality. Additionally, we provide multimodal predictive models combining video, audio, and physiological inputs. The data and the code for the baselines are available at https://project.inria.fr/stressid/.

Although existing fMRI-to-image reconstruction methods could predict high-quality images, they do not explicitly consider the semantic gap between training and testing data, resulting in reconstruction with unstable and uncertain semantics. This paper addresses the problem of generalized fMRI-to-image reconstruction by explicitly alleviates the semantic gap. Specifically, we leverage the pre-trained CLIP model to map the training data to a compact feature representation, which essentially extends the sparse semantics of training data to dense ones, thus alleviating the semantic gap of the instances nearby known concepts (i.e., inside the training super-classes). Inspired by the robust low-level representation in fMRI data, which could help alleviate the semantic gap for instances that far from the known concepts (i.e., outside the training super-classes), we leverage structural information as a general cue to guide image reconstruction. Further, we quantify the semantic uncertainty based on probability density estimation and achieve Generalized fMRI-to-image reconstruction by adaptively integrating Expanded Semantics and Structural information (GESS) within a diffusion process. Experimental results demonstrate that the proposed GESS model outperforms state-of-the-art methods, and we propose a generalized scenario split strategy to evaluate the advantage of GESS in closing the semantic gap.