Recently, DALL-E, a multimodal transformer language model, and its variants, including diffusion models, have shown high-quality text-to-image generation capabilities. However, despite the realistic image generation results, there has not been a detailed analysis of how to evaluate such models. In this work, we investigate the visual reasoning capabilities and social biases of different text-to-image models, covering both multimodal transformer language models and diffusion models. First, we measure three visual reasoning skills: object recognition, object counting, and spatial relation understanding. For this, we propose PaintSkills, a compositional diagnostic evaluation dataset that measures these skills. Despite the high-fidelity image generation capability, a large gap exists between the performance of recent models and the upper bound accuracy in object counting and spatial relation understanding skills. Second, we assess the gender and skin tone biases by measuring the gender/skin tone distribution of generated images across various professions and attributes. We demonstrate that recent text-to-image generation models learn specific biases about gender and skin tone from web image-text pairs. We hope our work will help guide future progress in improving text-to-image generation models on visual reasoning skills and learning socially unbiased representations. Code and data: https://github.com/j-min/DallEval
The success of deep neural networks in image classification and learning can be partly attributed to the features they extract from images. It is often speculated about the properties of a low-dimensional manifold that models extract and learn from images. However, there is not sufficient understanding about this low-dimensional space based on theory or empirical evidence. For image classification models, their last hidden layer is the one where images of each class is separated from other classes and it also has the least number of features. Here, we develop methods and formulations to study that feature space for any model. We study the partitioning of the domain in feature space, identify regions guaranteed to have certain classifications, and investigate its implications for the pixel space. We observe that geometric arrangements of decision boundaries in feature space is significantly different compared to pixel space, providing insights about adversarial vulnerabilities, image morphing, extrapolation, ambiguity in classification, and the mathematical understanding of image classification models.
We analyze the role of rotational equivariance in convolutional neural networks (CNNs) applied to spherical images. We compare the performance of the group equivariant networks known as S2CNNs and standard non-equivariant CNNs trained with an increasing amount of data augmentation. The chosen architectures can be considered baseline references for the respective design paradigms. Our models are trained and evaluated on single or multiple items from the MNIST or FashionMNIST dataset projected onto the sphere. For the task of image classification, which is inherently rotationally invariant, we find that by considerably increasing the amount of data augmentation and the size of the networks, it is possible for the standard CNNs to reach at least the same performance as the equivariant network. In contrast, for the inherently equivariant task of semantic segmentation, the non-equivariant networks are consistently outperformed by the equivariant networks with significantly fewer parameters. We also analyze and compare the inference latency and training times of the different networks, enabling detailed tradeoff considerations between equivariant architectures and data augmentation for practical problems. The equivariant spherical networks used in the experiments are available at https://github.com/JanEGerken/sem_seg_s2cnn .
Though remarkable progress has been achieved in various vision tasks, deep neural networks still suffer obvious performance degradation when tested in out-of-distribution scenarios. We argue that the feature statistics (mean and standard deviation), which carry the domain characteristics of the training data, can be properly manipulated to improve the generalization ability of deep learning models. Common methods often consider the feature statistics as deterministic values measured from the learned features and do not explicitly consider the uncertain statistics discrepancy caused by potential domain shifts during testing. In this paper, we improve the network generalization ability by modeling the uncertainty of domain shifts with synthesized feature statistics during training. Specifically, we hypothesize that the feature statistic, after considering the potential uncertainties, follows a multivariate Gaussian distribution. Hence, each feature statistic is no longer a deterministic value, but a probabilistic point with diverse distribution possibilities. With the uncertain feature statistics, the models can be trained to alleviate the domain perturbations and achieve better robustness against potential domain shifts. Our method can be readily integrated into networks without additional parameters. Extensive experiments demonstrate that our proposed method consistently improves the network generalization ability on multiple vision tasks, including image classification, semantic segmentation, and instance retrieval. The code can be available at https://github.com/lixiaotong97/DSU.
Machine Vision Components (MVC) are becoming safety-critical. Assuring their quality, including safety, is essential for their successful deployment. Assurance relies on the availability of precisely specified and, ideally, machine-verifiable requirements. MVCs with state-of-the-art performance rely on machine learning (ML) and training data but largely lack such requirements. In this paper, we address the need for defining machine-verifiable reliability requirements for MVCs against transformations that simulate the full range of realistic and safety-critical changes in the environment. Using human performance as a baseline, we define reliability requirements as: 'if the changes in an image do not affect a human's decision, neither should they affect the MVC's.' To this end, we provide: (1) a class of safety-related image transformations; (2) reliability requirement classes to specify correctness-preservation and prediction-preservation for MVCs; (3) a method to instantiate machine-verifiable requirements from these requirements classes using human performance experiment data; (4) human performance experiment data for image recognition involving eight commonly used transformations, from about 2000 human participants; and (5) a method for automatically checking whether an MVC satisfies our requirements. Further, we show that our reliability requirements are feasible and reusable by evaluating our methods on 13 state-of-the-art pre-trained image classification models. Finally, we demonstrate that our approach detects reliability gaps in MVCs that other existing methods are unable to detect.
Machine Learning has attracted considerable attention throughout the past decade due to its potential to solve far-reaching tasks, such as image classification, object recognition, anomaly detection, and data forecasting. A standard approach to tackle such applications is based on supervised learning, which is assisted by large sets of labeled data and is conducted by the so-called classifiers, such as Logistic Regression, Decision Trees, Random Forests, and Support Vector Machines, among others. An alternative to traditional classifiers is the parameterless Optimum-Path Forest (OPF), which uses a graph-based methodology and a distance measure to create arcs between nodes and hence sets of trees, responsible for conquering the nodes, defining their labels, and shaping the forests. Nevertheless, its performance is strongly associated with an appropriate distance measure, which may vary according to the dataset's nature. Therefore, this work proposes a comparative study over a wide range of distance measures applied to the supervised Optimum-Path Forest classification. The experimental results are conducted using well-known literature datasets and compared across benchmarking classifiers, illustrating OPF's ability to adapt to distinct domains.
Traditional methods of identifying pathologies in X-ray images rely heavily on skilled human interpretation and are often time-consuming. The advent of deep learning techniques has enabled the development of automated disease diagnosis systems. Still, the performance of such systems is opaque to end-users and limited to detecting a single pathology. In this paper, we propose a multi-label disease prediction model that allows the detection of more than one pathology at a given test time. We use a dense convolutional neural network (DenseNet) for disease diagnosis. Our proposed model achieved the highest AUC score of 0.896 for the condition Cardiomegaly with an accuracy of 0.826, while the lowest AUC score was obtained for Nodule, at 0.655 with an accuracy of 0.66. To build trust in decision-making, we generated heatmaps on X-rays to visualize the regions where the model paid attention to make certain predictions. Our proposed automated disease prediction model obtained highly confident high-performance metrics in multi-label disease prediction tasks.
Recent research has shown the existence of significant redundancy in large Transformer models. One can prune the redundant parameters without significantly sacrificing the generalization performance. However, we question whether the redundant parameters could have contributed more if they were properly trained. To answer this question, we propose a novel training strategy that encourages all parameters to be trained sufficiently. Specifically, we adaptively adjust the learning rate for each parameter according to its sensitivity, a robust gradient-based measure reflecting this parameter's contribution to the model performance. A parameter with low sensitivity is redundant, and we improve its fitting by increasing its learning rate. In contrast, a parameter with high sensitivity is well-trained, and we regularize it by decreasing its learning rate to prevent further overfitting. We conduct extensive experiments on natural language understanding, neural machine translation, and image classification to demonstrate the effectiveness of the proposed schedule. Analysis shows that the proposed schedule indeed reduces the redundancy and improves generalization performance.
Many datasets are underspecified: there exist multiple equally viable solutions to a given task. Underspecification can be problematic for methods that learn a single hypothesis because different functions that achieve low training loss can focus on different predictive features and thus produce widely varying predictions on out-of-distribution data. We propose DivDis, a simple two-stage framework that first learns a diverse collection of hypotheses for a task by leveraging unlabeled data from the test distribution. We then disambiguate by selecting one of the discovered hypotheses using minimal additional supervision, in the form of additional labels or inspection of function visualization. We demonstrate the ability of DivDis to find hypotheses that use robust features in image classification and natural language processing problems with underspecification.
Transfer learning has witnessed remarkable progress in recent years, for example, with the introduction of augmentation-based contrastive self-supervised learning methods. While a number of large-scale empirical studies on the transfer performance of such models have been conducted, there is not yet an agreed-upon set of control baselines, evaluation practices, and metrics to report, which often hinders a nuanced and calibrated understanding of the real efficacy of the methods. We share an evaluation standard that aims to quantify and communicate transfer learning performance in an informative and accessible setup. This is done by baking a number of simple yet critical control baselines in the evaluation method, particularly the blind-guess (quantifying the dataset bias), scratch-model (quantifying the architectural contribution), and maximal-supervision (quantifying the upper-bound). To demonstrate how the evaluation standard can be employed, we provide an example empirical study investigating a few basic questions about self-supervised learning. For example, using this standard, the study shows the effectiveness of existing self-supervised pre-training methods is skewed towards image classification tasks versus dense pixel-wise predictions. In general, we encourage using/reporting the suggested control baselines in evaluating transfer learning in order to gain a more meaningful and informative understanding.
Recent Spiking Neural Networks (SNNs) works focus on an image classification task, therefore various coding techniques have been proposed to convert an image into temporal binary spikes. Among them, rate coding and direct coding are regarded as prospective candidates for building a practical SNN system as they show state-of-the-art performance on large-scale datasets. Despite their usage, there is little attention to comparing these two coding schemes in a fair manner. In this paper, we conduct a comprehensive analysis of the two codings from three perspectives: accuracy, adversarial robustness, and energy-efficiency. First, we compare the performance of two coding techniques with various architectures and datasets. Then, we measure the robustness of the coding techniques on two adversarial attack methods. Finally, we compare the energy-efficiency of two coding schemes on a digital hardware platform. Our results show that direct coding can achieve better accuracy especially for a small number of timesteps. In contrast, rate coding shows better robustness to adversarial attacks owing to the non-differentiable spike generation process. Rate coding also yields higher energy-efficiency than direct coding which requires multi-bit precision for the first layer. Our study explores the characteristics of two codings, which is an important design consideration for building SNNs. The code is made available at https://github.com/Intelligent-Computing-Lab-Yale/Rate-vs-Direct.
In this work, we pursue a unified paradigm for multimodal pretraining to break the scaffolds of complex task/modality-specific customization. We propose OFA, a Task-Agnostic and Modality-Agnostic framework that supports Task Comprehensiveness. OFA unifies a diverse set of cross-modal and unimodal tasks, including image generation, visual grounding, image captioning, image classification, language modeling, etc., in a simple sequence-to-sequence learning framework. OFA follows the instruction-based learning in both pretraining and finetuning stages, requiring no extra task-specific layers for downstream tasks. In comparison with the recent state-of-the-art vision & language models that rely on extremely large cross-modal datasets, OFA is pretrained on only 20M publicly available image-text pairs. Despite its simplicity and relatively small-scale training data, OFA achieves new SOTAs in a series of cross-modal tasks while attaining highly competitive performances on uni-modal tasks. Our further analysis indicates that OFA can also effectively transfer to unseen tasks and unseen domains. Our code and models are publicly available at https://github.com/OFA-Sys/OFA.
Recent studies have demonstrated that gradient matching-based dataset synthesis, or dataset condensation (DC), methods can achieve state-of-the-art performance when applied to data-efficient learning tasks. However, in this study, we prove that the existing DC methods can perform worse than the random selection method when task-irrelevant information forms a significant part of the training dataset. We attribute this to the lack of participation of the contrastive signals between the classes resulting from the class-wise gradient matching strategy. To address this problem, we propose Dataset Condensation with Contrastive signals (DCC) by modifying the loss function to enable the DC methods to effectively capture the differences between classes. In addition, we analyze the new loss function in terms of training dynamics by tracking the kernel velocity. Furthermore, we introduce a bi-level warm-up strategy to stabilize the optimization. Our experimental results indicate that while the existing methods are ineffective for fine-grained image classification tasks, the proposed method can successfully generate informative synthetic datasets for the same tasks. Moreover, we demonstrate that the proposed method outperforms the baselines even on benchmark datasets such as SVHN, CIFAR-10, and CIFAR-100. Finally, we demonstrate the high applicability of the proposed method by applying it to continual learning tasks.
The advent of autonomous driving and advanced driver assistance systems necessitates continuous developments in computer vision for 3D scene understanding. Self-supervised monocular depth estimation, a method for pixel-wise distance estimation of objects from a single camera without the use of ground truth labels, is an important task in 3D scene understanding. However, existing methods for this task are limited to convolutional neural network (CNN) architectures. In contrast with CNNs that use localized linear operations and lose feature resolution across the layers, vision transformers process at constant resolution with a global receptive field at every stage. While recent works have compared transformers against their CNN counterparts for tasks such as image classification, no study exists that investigates the impact of using transformers for self-supervised monocular depth estimation. Here, we first demonstrate how to adapt vision transformers for self-supervised monocular depth estimation. Thereafter, we compare the transformer and CNN-based architectures for their performance on KITTI depth prediction benchmarks, as well as their robustness to natural corruptions and adversarial attacks, including when the camera intrinsics are unknown. Our study demonstrates how transformer-based architecture, though lower in run-time efficiency, achieves comparable performance while being more robust and generalizable.
We introduce Corrupted Image Modeling (CIM) for self-supervised visual pre-training. CIM uses an auxiliary generator with a small trainable BEiT to corrupt the input image instead of using artificial [MASK] tokens, where some patches are randomly selected and replaced with plausible alternatives sampled from the BEiT output distribution. Given this corrupted image, an enhancer network learns to either recover all the original image pixels, or predict whether each visual token is replaced by a generator sample or not. The generator and the enhancer are simultaneously trained and synergistically updated. After pre-training, the enhancer can be used as a high-capacity visual encoder for downstream tasks. CIM is a general and flexible visual pre-training framework that is suitable for various network architectures. For the first time, CIM demonstrates that both ViT and CNN can learn rich visual representations using a unified, non-Siamese framework. Experimental results show that our approach achieves compelling results in vision benchmarks, such as ImageNet classification and ADE20K semantic segmentation.
Camera traps have transformed how ecologists study wildlife species distributions, activity patterns, and interspecific interactions. Although camera traps provide a cost-effective method for monitoring species, the time required for data processing can limit survey efficiency. Thus, the potential of Artificial Intelligence (AI), specifically Deep Learning (DL), to process camera-trap data has gained considerable attention. Using DL for these applications involves training algorithms, such as Convolutional Neural Networks (CNNs), to automatically detect objects and classify species. To overcome technical challenges associated with training CNNs, several research communities have recently developed platforms that incorporate DL in easy-to-use interfaces. We review key characteristics of four AI-powered platforms --Wildlife Insights (WI), MegaDetector (MD), Machine Learning for Wildlife Image Classification (MLWIC2), and Conservation AI-- including data management tools and AI features. We also provide R code in an open-source GitBook, to demonstrate how users can evaluate model performance, and incorporate AI output in semi-automated workflows. We found that species classifications from WI and MLWIC2 generally had low recall values (animals that were present in the images often were not classified to the correct species). Yet, the precision of WI and MLWIC2 classifications for some species was high (i.e., when classifications were made, they were generally accurate). MD, which classifies images using broader categories (e.g., "blank" or "animal"), also performed well. Thus, we conclude that, although species classifiers were not accurate enough to automate image processing, DL could be used to improve efficiencies by accepting classifications with high confidence values for certain species or by filtering images containing blanks.
We propose a novel algorithm called Backpropagation Neural Tree (BNeuralT), which is a stochastic computational dendritic tree. BNeuralT takes random repeated inputs through its leaves and imposes dendritic nonlinearities through its internal connections like a biological dendritic tree would do. Considering the dendritic-tree like plausible biological properties, BNeuralT is a single neuron neural tree model with its internal sub-trees resembling dendritic nonlinearities. BNeuralT algorithm produces an ad hoc neural tree which is trained using a stochastic gradient descent optimizer like gradient descent (GD), momentum GD, Nesterov accelerated GD, Adagrad, RMSprop, or Adam. BNeuralT training has two phases, each computed in a depth-first search manner: the forward pass computes neural tree's output in a post-order traversal, while the error backpropagation during the backward pass is performed recursively in a pre-order traversal. A BNeuralT model can be considered a minimal subset of a neural network (NN), meaning it is a "thinned" NN whose complexity is lower than an ordinary NN. Our algorithm produces high-performing and parsimonious models balancing the complexity with descriptive ability on a wide variety of machine learning problems: classification, regression, and pattern recognition.
With the recent advances in A.I. methodologies and their application to medical imaging, there has been an explosion of related research programs utilizing these techniques to produce state-of-the-art classification performance. Ultimately, these research programs culminate in submission of their work for consideration in peer reviewed journals. To date, the criteria for acceptance vs. rejection is often subjective; however, reproducible science requires reproducible review. The Machine Learning Education Sub-Committee of SIIM has identified a knowledge gap and a serious need to establish guidelines for reviewing these studies. Although there have been several recent papers with this goal, this present work is written from the machine learning practitioners standpoint. In this series, the committee will address the best practices to be followed in an A.I.-based study and present the required sections in terms of examples and discussion of what should be included to make the studies cohesive, reproducible, accurate, and self-contained. This first entry in the series focuses on the task of image classification. Elements such as dataset curation, data pre-processing steps, defining an appropriate reference standard, data partitioning, model architecture and training are discussed. The sections are presented as they would be detailed in a typical manuscript, with content describing the necessary information that should be included to make sure the study is of sufficient quality to be considered for publication. The goal of this series is to provide resources to not only help improve the review process for A.I.-based medical imaging papers, but to facilitate a standard for the information that is presented within all components of the research study. We hope to provide quantitative metrics in what otherwise may be a qualitative review process.
Recent advances in deep learning have relied on large, labelled datasets to train high-capacity models. However, collecting large datasets in a time- and cost-efficient manner often results in label noise. We present a method for learning from noisy labels that leverages similarities between training examples in feature space, encouraging the prediction of each example to be similar to its nearest neighbours. Compared to training algorithms that use multiple models or distinct stages, our approach takes the form of a simple, additional regularization term. It can be interpreted as an inductive version of the classical, transductive label propagation algorithm. We thoroughly evaluate our method on datasets evaluating both synthetic (CIFAR-10, CIFAR-100) and realistic (mini-WebVision, WebVision, Clothing1M, mini-ImageNet-Red) noise, and achieve competitive or state-of-the-art accuracies across all of them.
Convolutional neural networks typically contain several downsampling operators, such as strided convolutions or pooling layers, that progressively reduce the resolution of intermediate representations. This provides some shift-invariance while reducing the computational complexity of the whole architecture. A critical hyperparameter of such layers is their stride: the integer factor of downsampling. As strides are not differentiable, finding the best configuration either requires cross-validation or discrete optimization (e.g. architecture search), which rapidly become prohibitive as the search space grows exponentially with the number of downsampling layers. Hence, exploring this search space by gradient descent would allow finding better configurations at a lower computational cost. This work introduces DiffStride, the first downsampling layer with learnable strides. Our layer learns the size of a cropping mask in the Fourier domain, that effectively performs resizing in a differentiable way. Experiments on audio and image classification show the generality and effectiveness of our solution: we use DiffStride as a drop-in replacement to standard downsampling layers and outperform them. In particular, we show that introducing our layer into a ResNet-18 architecture allows keeping consistent high performance on CIFAR10, CIFAR100 and ImageNet even when training starts from poor random stride configurations. Moreover, formulating strides as learnable variables allows us to introduce a regularization term that controls the computational complexity of the architecture. We show how this regularization allows trading off accuracy for efficiency on ImageNet.