This work focuses on tackling the challenging but realistic visual task of Incremental Few-Shot Learning (IFSL), which requires a model to continually learn novel classes from only a few examples while not forgetting the base classes on which it was pre-trained. Our study reveals that the challenges of IFSL lie in both inter-class separation and novel-class representation. Dur to intra-class variation, a novel class may implicitly leverage the knowledge from multiple base classes to construct its feature representation. Hence, simply reusing the pre-trained embedding space could lead to a scattered feature distribution and result in category confusion. To address such issues, we propose a two-step learning strategy referred to as \textbf{Im}planting and \textbf{Co}mpressing (\textbf{IMCO}), which optimizes both feature space partition and novel class reconstruction in a systematic manner. Specifically, in the \textbf{Implanting} step, we propose to mimic the data distribution of novel classes with the assistance of data-abundant base set, so that a model could learn semantically-rich features that are beneficial for discriminating between the base and other unseen classes. In the \textbf{Compressing} step, we adapt the feature extractor to precisely represent each novel class for enhancing intra-class compactness, together with a regularized parameter updating rule for preventing aggressive model updating. Finally, we demonstrate that IMCO outperforms competing baselines with a significant margin, both in image classification task and more challenging object detection task.
For image classification problems, various neural network models are commonly used due to their success in yielding high accuracies. Convolutional Neural Network (CNN) is one of the most frequently used deep learning methods for image classification applications. It may produce extraordinarily accurate results with regard to its complexity. However, the more complex the model is the longer it takes to train. In this paper, an acceleration design that uses the power of FPGA is given for a basic CNN model which consists of one convolutional layer and one fully connected layer for the training phase of the fully connected layer. Nonetheless, inference phase is also accelerated automatically due to the fact that training phase includes inference. In this design, the convolutional layer is calculated by the host computer and the fully connected layer is calculated by an FPGA board. It should be noted that the training of convolutional layer is not taken into account in this design and is left for future research. The results are quite encouraging as this FPGA design tops the performance of some of the state-of-the-art deep learning platforms such as Tensorflow on the host computer approximately 2 times in both training and inference.
Despite abundant data on the spatial distribution of contemporary human settlements, historical data on the long-term evolution of human settlements at fine spatial and temporal granularity is scarce, limiting our quantitative understanding of long-term changes of built-up areas. This is because commonly used mapping methods (e.g., image classification) and suitable data sources (i.e., aerial imagery, multi-spectral remote sensing data, LiDAR) have only been available in recent decades. However, there are alternative data sources such as cadastral records that are digitally available, containing relevant information such as building age information, allowing for an approximate, digital reconstruction of past building distributions. We conducted a non-exhaustive search of open and publicly available data resources from administrative institutions in the United States and gathered, integrated, and harmonized cadastral parcel data, tax assessment data, and building footprint data for 33 counties, wherever building footprint geometries and building construction year information was available. The result of this effort is a unique dataset which we call the Multi-Temporal Building Footprint Dataset for 33 U.S. Counties (MTBF-33). MTBF-33 contains over 6.2 million building footprints including their construction year, and can be used to derive retrospective depictions of built-up areas from 1900 to 2015, at fine spatial and temporal grain and can be used for data validation purposes, or to train statistical learning approaches aiming to extract historical information on human settlements from remote sensing data, historical maps, or similar data sources. MTBF-33 is available at http://doi.org/10.17632/w33vbvjtdy.
Recent joint embedding-based self-supervised methods have surpassed standard supervised approaches on various image recognition tasks such as image classification. These self-supervised methods aim at maximizing agreement between features extracted from two differently transformed views of the same image, which results in learning an invariant representation with respect to appearance and geometric image transformations. However, the effectiveness of these approaches remains unclear in the context of gaze estimation, a structured regression task that requires equivariance under geometric transformations (e.g., rotations, horizontal flip). In this work, we propose SwAT, an equivariant version of the online clustering-based self-supervised approach SwAV, to learn more informative representations for gaze estimation. We demonstrate that SwAT, with ResNet-50 and supported with uncurated unlabeled face images, outperforms state-of-the-art gaze estimation methods and supervised baselines in various experiments. In particular, we achieve up to 57% and 25% improvements in cross-dataset and within-dataset evaluation tasks on existing benchmarks (ETH-XGaze, Gaze360, and MPIIFaceGaze).
Most of the existing Out-Of-Distribution (OOD) detection algorithms depend on single input source: the feature, the logit, or the softmax probability. However, the immense diversity of the OOD examples makes such methods fragile. There are OOD samples that are easy to identify in the feature space while hard to distinguish in the logit space and vice versa. Motivated by this observation, we propose a novel OOD scoring method named Virtual-logit Matching (ViM), which combines the class-agnostic score from feature space and the In-Distribution (ID) class-dependent logits. Specifically, an additional logit representing the virtual OOD class is generated from the residual of the feature against the principal space, and then matched with the original logits by a constant scaling. The probability of this virtual logit after softmax is the indicator of OOD-ness. To facilitate the evaluation of large-scale OOD detection in academia, we create a new OOD dataset for ImageNet-1K, which is human-annotated and is 8.8x the size of existing datasets. We conducted extensive experiments, including CNNs and vision transformers, to demonstrate the effectiveness of the proposed ViM score. In particular, using the BiT-S model, our method gets an average AUROC 90.91% on four difficult OOD benchmarks, which is 4% ahead of the best baseline. Code and dataset are available at https://github.com/haoqiwang/vim.
Image signal processors (ISPs) are historically grown legacy software systems for reconstructing color images from noisy raw sensor measurements. They are usually composited of many heuristic blocks for denoising, demosaicking, and color restoration. Color reproduction in this context is of particular importance, since the raw colors are often severely distorted, and each smart phone manufacturer has developed their own characteristic heuristics for improving the color rendition, for example of skin tones and other visually important colors. In recent years there has been strong interest in replacing the historically grown ISP systems with deep learned pipelines. Much progress has been made in approximating legacy ISPs with such learned models. However, so far the focus of these efforts has been on reproducing the structural features of the images, with less attention paid to color rendition. Here we present CRISPnet, the first learned ISP model to specifically target color rendition accuracy relative to a complex, legacy smart phone ISP. We achieve this by utilizing both image metadata (like a legacy ISP would), as well as by learning simple global semantics based on image classification -- similar to what a legacy ISP does to determine the scene type. We also contribute a new ISP image dataset consisting of both high dynamic range monitor data, as well as real-world data, both captured with an actual cell phone ISP pipeline under a variety of lighting conditions, exposure times, and gain settings.
Transformers exhibit great advantages in handling computer vision tasks. They model image classification tasks by utilizing a multi-head attention mechanism to process a series of patches consisting of split images. However, for complex tasks, Transformer in computer vision not only requires inheriting a bit of dynamic attention and global context, but also needs to introduce features concerning noise reduction, shifting, and scaling invariance of objects. Therefore, here we take a step forward to study the structural characteristics of Transformer and convolution and propose an architecture search method-Vision Transformer with Convolutions Architecture Search (VTCAS). The high-performance backbone network searched by VTCAS introduces the desirable features of convolutional neural networks into the Transformer architecture while maintaining the benefits of the multi-head attention mechanism. The searched block-based backbone network can extract feature maps at different scales. These features are compatible with a wider range of visual tasks, such as image classification (32 M parameters, 82.0% Top-1 accuracy on ImageNet-1K) and object detection (50.4% mAP on COCO2017). The proposed topology based on the multi-head attention mechanism and CNN adaptively associates relational features of pixels with multi-scale features of objects. It enhances the robustness of the neural network for object recognition, especially in the low illumination indoor scene.
For robots to be generally useful, they must be able to find arbitrary objects described by people (i.e., be language-driven) even without expensive navigation training on in-domain data (i.e., perform zero-shot inference). We explore these capabilities in a unified setting: language-driven zero-shot object navigation (L-ZSON). Inspired by the recent success of open-vocabulary models for image classification, we investigate a straightforward framework, CLIP on Wheels (CoW), to adapt open-vocabulary models to this task without fine-tuning. To better evaluate L-ZSON, we introduce the Pasture benchmark, which considers finding uncommon objects, objects described by spatial and appearance attributes, and hidden objects described relative to visible objects. We conduct an in-depth empirical study by directly deploying 21 CoW baselines across Habitat, RoboTHOR, and Pasture. In total, we evaluate over 90k navigation episodes and find that (1) CoW baselines often struggle to leverage language descriptions, but are proficient at finding uncommon objects. (2) A simple CoW, with CLIP-based object localization and classical exploration -- and no additional training -- matches the navigation efficiency of a state-of-the-art ZSON method trained for 500M steps on Habitat MP3D data. This same CoW provides a 15.6 percentage point improvement in success over a state-of-the-art RoboTHOR ZSON model.
Current visual detectors, though impressive within their training distribution, often fail to parse out-of-distribution scenes into their constituent entities. Recent test-time adaptation methods use auxiliary self-supervised losses to adapt the network parameters to each test example independently and have shown promising results towards generalization outside the training distribution for the task of image classification. In our work, we find evidence that these losses are insufficient for the task of scene decomposition, without also considering architectural inductive biases. Recent slot-centric generative models attempt to decompose scenes into entities in a self-supervised manner by reconstructing pixels. Drawing upon these two lines of work, we propose Slot-TTA, a semi-supervised slot-centric scene decomposition model that at test time is adapted per scene through gradient descent on reconstruction or cross-view synthesis objectives. We evaluate Slot-TTA across multiple input modalities, images or 3D point clouds, and show substantial out-of-distribution performance improvements against state-of-the-art supervised feed-forward detectors, and alternative test-time adaptation methods.
The ability to discover new transients via image differencing without direct human intervention is an important task in observational astronomy. For these kind of image classification problems, machine Learning techniques such as Convolutional Neural Networks (CNNs) have shown remarkable success. In this work, we present the results of an automated transient identification on images with CNNs for an extant dataset from the Dark Energy Survey Supernova program (DES-SN), whose main focus was on using Type Ia supernovae for cosmology. By performing an architecture search of CNNs, we identify networks that efficiently select non-artifacts (e.g. supernovae, variable stars, AGN, etc.) from artifacts (image defects, mis-subtractions, etc.), achieving the efficiency of previous work performed with random Forests, without the need to expend any effort in feature identification. The CNNs also help us identify a subset of mislabeled images. Performing a relabeling of the images in this subset, the resulting classification with CNNs is significantly better than previous results.
After their initial success in natural language processing, transformer architectures have rapidly gained traction in computer vision, providing state-of-the-art results for tasks such as image classification, detection, segmentation, and video analysis. We offer three insights based on simple and easy to implement variants of vision transformers. (1) The residual layers of vision transformers, which are usually processed sequentially, can to some extent be processed efficiently in parallel without noticeably affecting the accuracy. (2) Fine-tuning the weights of the attention layers is sufficient to adapt vision transformers to a higher resolution and to other classification tasks. This saves compute, reduces the peak memory consumption at fine-tuning time, and allows sharing the majority of weights across tasks. (3) Adding MLP-based patch pre-processing layers improves Bert-like self-supervised training based on patch masking. We evaluate the impact of these design choices using the ImageNet-1k dataset, and confirm our findings on the ImageNet-v2 test set. Transfer performance is measured across six smaller datasets.
To generalize well, classifiers must learn to be invariant to nuisance transformations that do not alter an input's class. Many problems have "class-agnostic" nuisance transformations that apply similarly to all classes, such as lighting and background changes for image classification. Neural networks can learn these invariances given sufficient data, but many real-world datasets are heavily class imbalanced and contain only a few examples for most of the classes. We therefore pose the question: how well do neural networks transfer class-agnostic invariances learned from the large classes to the small ones? Through careful experimentation, we observe that invariance to class-agnostic transformations is still heavily dependent on class size, with the networks being much less invariant on smaller classes. This result holds even when using data balancing techniques, and suggests poor invariance transfer across classes. Our results provide one explanation for why classifiers generalize poorly on unbalanced and long-tailed distributions. Based on this analysis, we show how a generative approach for learning the nuisance transformations can help transfer invariances across classes and improve performance on a set of imbalanced image classification benchmarks. Source code for our experiments is available at https://github.com/AllanYangZhou/generative-invariance-transfer.
We present a general-purpose framework for image modelling and vision tasks based on probabilistic frame prediction. Our approach unifies a broad range of tasks, from image segmentation, to novel view synthesis and video interpolation. We pair this framework with an architecture we term Transframer, which uses U-Net and Transformer components to condition on annotated context frames, and outputs sequences of sparse, compressed image features. Transframer is the state-of-the-art on a variety of video generation benchmarks, is competitive with the strongest models on few-shot view synthesis, and can generate coherent 30 second videos from a single image without any explicit geometric information. A single generalist Transframer simultaneously produces promising results on 8 tasks, including semantic segmentation, image classification and optical flow prediction with no task-specific architectural components, demonstrating that multi-task computer vision can be tackled using probabilistic image models. Our approach can in principle be applied to a wide range of applications that require learning the conditional structure of annotated image-formatted data.
Machine learning based image classification algorithms, such as deep neural network approaches, will be increasingly employed in critical settings such as quality control in industry, where transparency and comprehensibility of decisions are crucial. Therefore, we aim to extend the defect detection task towards an interactive human-in-the-loop approach that allows us to integrate rich background knowledge and the inference of complex relationships going beyond traditional purely data-driven approaches. We propose an approach for an interactive support system for classifications in an industrial quality control setting that combines the advantages of both (explainable) knowledge-driven and data-driven machine learning methods, in particular inductive logic programming and convolutional neural networks, with human expertise and control. The resulting system can assist domain experts with decisions, provide transparent explanations for results, and integrate feedback from users; thus reducing workload for humans while both respecting their expertise and without removing their agency or accountability.
Modern deep neural networks for classification usually jointly learn a backbone for representation and a linear classifier to output the logit of each class. A recent study has shown a phenomenon called neural collapse that the within-class means of features and the classifier vectors converge to the vertices of a simplex equiangular tight frame (ETF) at the terminal phase of training on a balanced dataset. Since the ETF geometric structure maximally separates the pair-wise angles of all classes in the classifier, it is natural to raise the question, why do we spend an effort to learn a classifier when we know its optimal geometric structure? In this paper, we study the potential of learning a neural network for classification with the classifier randomly initialized as an ETF and fixed during training. Our analytical work based on the layer-peeled model indicates that the feature learning with a fixed ETF classifier naturally leads to the neural collapse state even when the dataset is imbalanced among classes. We further show that in this case the cross entropy (CE) loss is not necessary and can be replaced by a simple squared loss that shares the same global optimality but enjoys a better convergence property. Our experimental results show that our method is able to bring significant improvements with faster convergence on multiple imbalanced datasets.
Deep learning methods have been successful in solving tasks in machine learning and have made breakthroughs in many sectors owing to their ability to automatically extract features from unstructured data. However, their performance relies on manual trial-and-error processes for selecting an appropriate network architecture, hyperparameters for training, and pre-/post-procedures. Even though it has been shown that network architecture plays a critical role in learning feature representation feature from data and the final performance, searching for the best network architecture is computationally intensive and heavily relies on researchers' experience. Automated machine learning (AutoML) and its advanced techniques i.e. Neural Architecture Search (NAS) have been promoted to address those limitations. Not only in general computer vision tasks, but NAS has also motivated various applications in multiple areas including medical imaging. In medical imaging, NAS has significant progress in improving the accuracy of image classification, segmentation, reconstruction, and more. However, NAS requires the availability of large annotated data, considerable computation resources, and pre-defined tasks. To address such limitations, meta-learning has been adopted in the scenarios of few-shot learning and multiple tasks. In this book chapter, we first present a brief review of NAS by discussing well-known approaches in search space, search strategy, and evaluation strategy. We then introduce various NAS approaches in medical imaging with different applications such as classification, segmentation, detection, reconstruction, etc. Meta-learning in NAS for few-shot learning and multiple tasks is then explained. Finally, we describe several open problems in NAS.
Research on the generalization ability of deep neural networks (DNNs) has recently attracted a great deal of attention. However, due to their complex architectures and large numbers of parameters, measuring the generalization ability of specific DNN models remains an open challenge. In this paper, we propose to use multiple factors to measure and rank the relative generalization of DNNs based on a new concept of confidence dimension (CD). Furthermore, we provide a feasible framework in our CD to theoretically calculate the upper bound of generalization based on the conventional Vapnik-Chervonenk dimension (VC-dimension) and Hoeffding's inequality. Experimental results on image classification and object detection demonstrate that our CD can reflect the relative generalization ability for different DNNs. In addition to full-precision DNNs, we also analyze the generalization ability of binary neural networks (BNNs), whose generalization ability remains an unsolved problem. Our CD yields a consistent and reliable measure and ranking for both full-precision DNNs and BNNs on all the tasks.
The paradigm of training models on massive data without label through self-supervised learning (SSL) and finetuning on many downstream tasks has become a trend recently. However, due to the high training costs and the unconsciousness of downstream usages, most self-supervised learning methods lack the capability to correspond to the diversities of downstream scenarios, as there are various data domains, different vision tasks and latency constraints on models. Neural architecture search (NAS) is one universally acknowledged fashion to conquer the issues above, but applying NAS on SSL seems impossible as there is no label or metric provided for judging model selection. In this paper, we present DATA, a simple yet effective NAS approach specialized for SSL that provides Domain-Aware and Task-Aware pre-training. Specifically, we (i) train a supernet which could be deemed as a set of millions of networks covering a wide range of model scales without any label, (ii) propose a flexible searching mechanism compatible with SSL that enables finding networks of different computation costs, for various downstream vision tasks and data domains without explicit metric provided. Instantiated With MoCo v2, our method achieves promising results across a wide range of computation costs on downstream tasks, including image classification, object detection and semantic segmentation. DATA is orthogonal to most existing SSL methods and endows them the ability of customization on downstream needs. Extensive experiments on other SSL methods demonstrate the generalizability of the proposed method. Code is released at https://github.com/GAIA-vision/GAIA-ssl
The ever-growing volume of satellite imagery data presents a challenge for industry and governments making data-driven decisions based on the timely analysis of very large data sets. Commonly used deep learning algorithms for automatic classification of satellite images are time and resource-intensive to train. The cost of retraining in the context of Big Data presents a practical challenge when new image data and/or classes are added to a training corpus. Recognizing the need for an adaptable, accurate, and scalable satellite image chip classification scheme, in this research we present an ensemble of: i) a slow to train but high accuracy vision transformer; and ii) a fast to train, low-parameter convolutional neural network. The vision transformer model provides a scalable and accurate foundation model. The high-speed CNN provides an efficient means of incorporating newly labelled data into analysis, at the expense of lower accuracy. To simulate incremental data, the very large (~400,000 images) So2Sat LCZ42 satellite image chip dataset is divided into four intervals, with the high-speed CNN retrained every interval and the vision transformer trained every half interval. This experimental setup mimics an increase in data volume and diversity over time. For the task of automated land-cover/land-use classification, the ensemble models for each data increment outperform each of the component models, with best accuracy of 65% against a holdout test partition of the So2Sat dataset. The proposed ensemble and staggered training schedule provide a scalable and cost-effective satellite image classification scheme that is optimized to process very large volumes of satellite data.
Sponge examples are test-time inputs optimized to increase energy consumption and prediction latency of deep networks deployed on hardware accelerators. By increasing the fraction of neurons activated during classification, these attacks reduce sparsity in network activation patterns, worsening the performance of hardware accelerators. In this work, we present a novel training-time attack, named sponge poisoning, which aims to worsen energy consumption and prediction latency of neural networks on any test input without affecting classification accuracy. To stage this attack, we assume that the attacker can control only a few model updates during training -- a likely scenario, e.g., when model training is outsourced to an untrusted third party or distributed via federated learning. Our extensive experiments on image classification tasks show that sponge poisoning is effective, and that fine-tuning poisoned models to repair them poses prohibitive costs for most users, highlighting that tackling sponge poisoning remains an open issue.