Multimodal Large Language Models (MLLMs) have demonstrated notable capabilities in general visual understanding and reasoning tasks. However, their deployment is hindered by substantial computational costs in both training and inference, limiting accessibility to the broader research and user communities. A straightforward solution is to leverage smaller pre-trained vision and language models, which inevitably cause significant performance drops. In this paper, we demonstrate the possibility of training a smaller but better MLLM with high-quality training data. Specifically, we introduce Bunny, a family of lightweight MLLMs with flexible vision and language backbones for efficient multimodal learning from selected training data. Experiments show that our Bunny-4B/8B outperforms the state-of-the-art large MLLMs on multiple benchmarks. We expect that this work can provide the community with a clean and flexible open-source tool for further research and development. The code, models, and data can be found in https://github.com/BAAI-DCAI/Bunny.
Vision Transformer (ViT) self-attention mechanism is characterized by feature collapse in deeper layers, resulting in the vanishing of low-level visual features. However, such features can be helpful to accurately represent and identify elements within an image and increase the accuracy and robustness of vision-based recognition systems. Following this rationale, we propose a novel residual attention learning method for improving ViT-based architectures, increasing their visual feature diversity and model robustness. In this way, the proposed network can capture and preserve significant low-level features, providing more details about the elements within the scene being analyzed. The effectiveness and robustness of the presented method are evaluated on five image classification benchmarks, including ImageNet1k, CIFAR10, CIFAR100, Oxford Flowers-102, and Oxford-IIIT Pet, achieving improved performances. Additionally, experiments on the COCO2017 dataset show that the devised approach discovers and incorporates semantic and spatial relationships for object detection and instance segmentation when implemented into spatial-aware transformer models.
We present a novel bi-directional Transformer architecture (BiXT) which scales linearly with input size in terms of computational cost and memory consumption, but does not suffer the drop in performance or limitation to only one input modality seen with other efficient Transformer-based approaches. BiXT is inspired by the Perceiver architectures but replaces iterative attention with an efficient bi-directional cross-attention module in which input tokens and latent variables attend to each other simultaneously, leveraging a naturally emerging attention-symmetry between the two. This approach unlocks a key bottleneck experienced by Perceiver-like architectures and enables the processing and interpretation of both semantics ('what') and location ('where') to develop alongside each other over multiple layers -- allowing its direct application to dense and instance-based tasks alike. By combining efficiency with the generality and performance of a full Transformer architecture, BiXT can process longer sequences like point clouds, text or images at higher feature resolutions and achieves competitive performance across a range of tasks like point cloud part segmentation, semantic image segmentation, image classification, hierarchical sequence modeling and document retrieval. Our experiments demonstrate that BiXT models outperform larger competitors by leveraging longer sequences more efficiently on vision tasks like classification and segmentation, and perform on par with full Transformer variants on sequence modeling and document retrieval -- but require $28\%$ fewer FLOPs and are up to $8.4\times$ faster.
Crowdsourced machine learning on competition platforms such as Kaggle is a popular and often effective method for generating accurate models. Typically, teams vie for the most accurate model, as measured by overall error on a holdout set, and it is common towards the end of such competitions for teams at the top of the leaderboard to ensemble or average their models outside the platform mechanism to get the final, best global model. In arXiv:2201.10408, the authors developed an alternative crowdsourcing framework in the context of fair machine learning, in order to integrate community feedback into models when subgroup unfairness is present and identifiable. There, unlike in classical crowdsourced ML, participants deliberately specialize their efforts by working on subproblems, such as demographic subgroups in the service of fairness. Here, we take a broader perspective on this work: we note that within this framework, participants may both specialize in the service of fairness and simply to cater to their particular expertise (e.g., focusing on identifying bird species in an image classification task). Unlike traditional crowdsourcing, this allows for the diversification of participants' efforts and may provide a participation mechanism to a larger range of individuals (e.g. a machine learning novice who has insight into a specific fairness concern). We present the first medium-scale experimental evaluation of this framework, with 46 participating teams attempting to generate models to predict income from American Community Survey data. We provide an empirical analysis of teams' approaches, and discuss the novel system architecture we developed. From here, we give concrete guidance for how best to deploy such a framework.
Whole-slide image (WSI) classification is a challenging task because 1) patches from WSI lack annotation, and 2) WSI possesses unnecessary variability, e.g., stain protocol. Recently, Multiple-Instance Learning (MIL) has made significant progress, allowing for classification based on slide-level, rather than patch-level, annotations. However, existing MIL methods ignore that all patches from normal slides are normal. Using this free annotation, we introduce a semi-supervision signal to de-bias the inter-slide variability and to capture the common factors of variation within normal patches. Because our method is orthogonal to the MIL algorithm, we evaluate our method on top of the recently proposed MIL algorithms and also compare the performance with other semi-supervised approaches. We evaluate our method on two public WSI datasets including Camelyon-16 and TCGA lung cancer and demonstrate that our approach significantly improves the predictive performance of existing MIL algorithms and outperforms other semi-supervised algorithms. We release our code at https://github.com/AITRICS/pathology_mil.
Oftentimes, machine learning applications using neural networks involve solving discrete optimization problems, such as in pruning, parameter-isolation-based continual learning and training of binary networks. Still, these discrete problems are combinatorial in nature and are also not amenable to gradient-based optimization. Additionally, classical approaches used in discrete settings do not scale well to large neural networks, forcing scientists and empiricists to rely on alternative methods. Among these, two main distinct sources of top-down information can be used to lead the model to good solutions: (1) extrapolating gradient information from points outside of the solution set (2) comparing evaluations between members of a subset of the valid solutions. We take continuation path (CP) methods to represent using purely the former and Monte Carlo (MC) methods to represent the latter, while also noting that some hybrid methods combine the two. The main goal of this work is to compare both approaches. For that purpose, we first overview the two classes while also discussing some of their drawbacks analytically. Then, on the experimental section, we compare their performance, starting with smaller microworld experiments, which allow more fine-grained control of problem variables, and gradually moving towards larger problems, including neural network regression and neural network pruning for image classification, where we additionally compare against magnitude-based pruning.
Class-Incremental Learning (CIL) is a practical and challenging problem for achieving general artificial intelligence. Recently, Pre-Trained Models (PTMs) have led to breakthroughs in both visual and natural language processing tasks. Despite recent studies showing PTMs' potential ability to learn sequentially, a plethora of work indicates the necessity of alleviating the catastrophic forgetting of PTMs. Through a pilot study and a causal analysis of CIL, we reveal that the crux lies in the imbalanced causal effects between new and old data. Specifically, the new data encourage models to adapt to new classes while hindering the adaptation of old classes. Similarly, the old data encourages models to adapt to old classes while hindering the adaptation of new classes. In other words, the adaptation process between new and old classes conflicts from the causal perspective. To alleviate this problem, we propose Balancing the Causal Effects (BaCE) in CIL. Concretely, BaCE proposes two objectives for building causal paths from both new and old data to the prediction of new and classes, respectively. In this way, the model is encouraged to adapt to all classes with causal effects from both new and old data and thus alleviates the causal imbalance problem. We conduct extensive experiments on continual image classification, continual text classification, and continual named entity recognition. Empirical results show that BaCE outperforms a series of CIL methods on different tasks and settings.
Informally, a model is calibrated if its predictions are correct with a probability that matches the confidence of the prediction. By far the most common method in the literature for measuring calibration is the expected calibration error (ECE). Recent work, however, has pointed out drawbacks of ECE, such as the fact that it is discontinuous in the space of predictors. In this work, we ask: how fundamental are these issues, and what are their impacts on existing results? Towards this end, we completely characterize the discontinuities of ECE with respect to general probability measures on Polish spaces. We then use the nature of these discontinuities to motivate a novel continuous, easily estimated miscalibration metric, which we term Logit-Smoothed ECE (LS-ECE). By comparing the ECE and LS-ECE of pre-trained image classification models, we show in initial experiments that binned ECE closely tracks LS-ECE, indicating that the theoretical pathologies of ECE may be avoidable in practice.
Purpose: We apply federated learning to train an OCT image classifier simulating a realistic scenario with multiple clients and statistical heterogeneous data distribution where data in the clients lack samples of some categories entirely. Methods: We investigate the effectiveness of FedAvg and FedProx to train an OCT image classification model in a decentralized fashion, addressing privacy concerns associated with centralizing data. We partitioned a publicly available OCT dataset across multiple clients under IID and Non-IID settings and conducted local training on the subsets for each client. We evaluated two federated learning methods, FedAvg and FedProx for these settings. Results: Our experiments on the dataset suggest that under IID settings, both methods perform on par with training on a central data pool. However, the performance of both algorithms declines as we increase the statistical heterogeneity across the client data, while FedProx consistently performs better than FedAvg in the increased heterogeneity settings. Conclusion: Despite the effectiveness of federated learning in the utilization of private data across multiple medical institutions, the large number of clients and heterogeneous distribution of labels deteriorate the performance of both algorithms. Notably, FedProx appears to be more robust to the increased heterogeneity.
Hyper spectral images have drawn the attention of the researchers for its complexity to classify. It has nonlinear relation between the materials and the spectral information provided by the HSI image. Deep learning methods have shown superiority in learning this nonlinearity in comparison to traditional machine learning methods. Use of 3-D CNN along with 2-D CNN have shown great success for learning spatial and spectral features. However, it uses comparatively large number of parameters. Moreover, it is not effective to learn inter layer information. Hence, this paper proposes a neural network combining 3-D CNN, 2-D CNN and Bi-LSTM. The performance of this model has been tested on Indian Pines(IP) University of Pavia(PU) and Salinas Scene(SA) data sets. The results are compared with the state of-the-art deep learning-based models. This model performed better in all three datasets. It could achieve 99.83, 99.98 and 100 percent accuracy using only 30 percent trainable parameters of the state-of-art model in IP, PU and SA datasets respectively.
Satellite missions and Earth Observation (EO) systems represent fundamental assets for environmental monitoring and the timely identification of catastrophic events, long-term monitoring of both natural resources and human-made assets, such as vegetation, water bodies, forests as well as buildings. Different EO missions enables the collection of information on several spectral bandwidths, such as MODIS, Sentinel-1 and Sentinel-2. Thus, given the recent advances of machine learning, computer vision and the availability of labeled data, researchers demonstrated the feasibility and the precision of land-use monitoring systems and remote sensing image classification through the use of deep neural networks. Such systems may help domain experts and governments in constant environmental monitoring, enabling timely intervention in case of catastrophic events (e.g., forest wildfire in a remote area). Despite the recent advances in the field of computer vision, many works limit their analysis on Convolutional Neural Networks (CNNs) and, more recently, to vision transformers (ViTs). Given the recent successes of Graph Neural Networks (GNNs) on non-graph data, such as time-series and images, we investigate the performances of a recent Vision GNN architecture (ViG) applied to the task of land cover classification. The experimental results show that ViG achieves state-of-the-art performances in multiclass and multilabel classification contexts, surpassing both ViT and ResNet on large-scale benchmarks.
Deep Learning (DL) is undergoing a paradigm shift with the emergence of foundation models, aptly named by their crucial, yet incomplete nature. In this work, we focus on Contrastive Language-Image Pre-training (CLIP), an open-vocabulary foundation model, which achieves high accuracy across many image classification tasks and is often competitive with a fully supervised baseline without being explicitly trained. Nevertheless, there are still domains where zero-shot CLIP performance is far from optimal, such as Remote Sensing (RS) and medical imagery. These domains do not only exhibit fundamentally different distributions compared to natural images, but also commonly rely on complementary modalities, beyond RGB, to derive meaningful insights. To this end, we propose a methodology for the purpose of aligning distinct RS imagery modalities with the visual and textual modalities of CLIP. Our two-stage procedure, comprises of robust fine-tuning CLIP in order to deal with the distribution shift, accompanied by the cross-modal alignment of a RS modality encoder, in an effort to extend the zero-shot capabilities of CLIP. We ultimately demonstrate our method on the tasks of RS imagery classification and cross-modal retrieval. We empirically show that both robust fine-tuning and cross-modal alignment translate to significant performance gains, across several RS benchmark datasets. Notably, these enhancements are achieved without the reliance on textual descriptions, without introducing any task-specific parameters, without training from scratch and without catastrophic forgetting.
Convolutional neural networks (CNNs) for image processing tend to focus on localized texture patterns, commonly referred to as texture bias. While most of the previous works in the literature focus on the task of image classification, we go beyond this and study the texture bias of CNNs in semantic segmentation. In this work, we propose to train CNNs on pre-processed images with less texture to reduce the texture bias. Therein, the challenge is to suppress image texture while preserving shape information. To this end, we utilize edge enhancing diffusion (EED), an anisotropic image diffusion method initially introduced for image compression, to create texture reduced duplicates of existing datasets. Extensive numerical studies are performed with both CNNs and vision transformer models trained on original data and EED-processed data from the Cityscapes dataset and the CARLA driving simulator. We observe strong texture-dependence of CNNs and moderate texture-dependence of transformers. Training CNNs on EED-processed images enables the models to become completely ignorant with respect to texture, demonstrating resilience with respect to texture re-introduction to any degree. Additionally we analyze the performance reduction in depth on a level of connected components in the semantic segmentation and study the influence of EED pre-processing on domain generalization as well as adversarial robustness.
Identifying how much a model ${\widehat{p}}_{\theta}(Y|X)$ knows about the stochastic real-world process $p(Y|X)$ it was trained on is important to ensure it avoids producing incorrect or "hallucinated" answers or taking unsafe actions. But this is difficult for generative models because probabilistic predictions do not distinguish between per-response noise (aleatoric uncertainty) and lack of knowledge about the process (epistemic uncertainty), and existing epistemic uncertainty quantification techniques tend to be overconfident when the model underfits. We propose a general strategy for teaching a model to both approximate $p(Y|X)$ and also estimate the remaining gaps between ${\widehat{p}}_{\theta}(Y|X)$ and $p(Y|X)$: train it to predict pairs of independent responses drawn from the true conditional distribution, allow it to "cheat" by observing one response while predicting the other, then measure how much it cheats. Remarkably, we prove that being good at cheating (i.e. cheating whenever it improves your prediction) is equivalent to being second-order calibrated, a principled extension of ordinary calibration that allows us to construct provably-correct frequentist confidence intervals for $p(Y|X)$ and detect incorrect responses with high probability. We demonstrate empirically that our approach accurately estimates how much models don't know across ambiguous image classification, (synthetic) language modeling, and partially-observable navigation tasks, outperforming existing techniques.
Deep neural networks are extensively applied to real-world tasks, such as face recognition and medical image classification, where privacy and data protection are critical. Image data, if not protected, can be exploited to infer personal or contextual information. Existing privacy preservation methods, like encryption, generate perturbed images that are unrecognizable to even humans. Adversarial attack approaches prohibit automated inference even for authorized stakeholders, limiting practical incentives for commercial and widespread adaptation. This pioneering study tackles an unexplored practical privacy preservation use case by generating human-perceivable images that maintain accurate inference by an authorized model while evading other unauthorized black-box models of similar or dissimilar objectives, and addresses the previous research gaps. The datasets employed are ImageNet, for image classification, Celeba-HQ dataset, for identity classification, and AffectNet, for emotion classification. Our results show that the generated images can successfully maintain the accuracy of a protected model and degrade the average accuracy of the unauthorized black-box models to 11.97%, 6.63%, and 55.51% on ImageNet, Celeba-HQ, and AffectNet datasets, respectively.
Exponential Moving Average (EMA) is a widely used weight averaging (WA) regularization to learn flat optima for better generalizations without extra cost in deep neural network (DNN) optimization. Despite achieving better flatness, existing WA methods might fall into worse final performances or require extra test-time computations. This work unveils the full potential of EMA with a single line of modification, i.e., switching the EMA parameters to the original model after each epoch, dubbed as Switch EMA (SEMA). From both theoretical and empirical aspects, we demonstrate that SEMA can help DNNs to reach generalization optima that better trade-off between flatness and sharpness. To verify the effectiveness of SEMA, we conduct comparison experiments with discriminative, generative, and regression tasks on vision and language datasets, including image classification, self-supervised learning, object detection and segmentation, image generation, video prediction, attribute regression, and language modeling. Comprehensive results with popular optimizers and networks show that SEMA is a free lunch for DNN training by improving performances and boosting convergence speeds.
In today's machine learning landscape, fine-tuning pretrained transformer models has emerged as an essential technique, particularly in scenarios where access to task-aligned training data is limited. However, challenges surface when data sharing encounters obstacles due to stringent privacy regulations or user apprehension regarding personal information disclosure. Earlier works based on secure multiparty computation (SMC) and fully homomorphic encryption (FHE) for privacy-preserving machine learning (PPML) focused more on privacy-preserving inference than privacy-preserving training. In response, we introduce BlindTuner, a privacy-preserving fine-tuning system that enables transformer training exclusively on homomorphically encrypted data for image classification. Our extensive experimentation validates BlindTuner's effectiveness by demonstrating comparable accuracy to non-encrypted models. Notably, our findings highlight a substantial speed enhancement of 1.5x to 600x over previous work in this domain.
Recent research has seen many behavioral comparisons between humans and deep neural networks (DNNs) in the domain of image classification. Often, comparison studies focus on the end-result of the learning process by measuring and comparing the similarities in the representations of object categories once they have been formed. However, the process of how these representations emerge -- that is, the behavioral changes and intermediate stages observed during the acquisition -- is less often directly and empirically compared. Here we report a detailed investigation of the learning dynamics in human observers and various classic and state-of-the-art DNNs. We develop a constrained supervised learning environment to align learning-relevant conditions such as starting point, input modality, available input data and the feedback provided. Across the whole learning process we evaluate and compare how well learned representations can be generalized to previously unseen test data. Comparisons across the entire learning process indicate that DNNs demonstrate a level of data efficiency comparable to human learners, challenging some prevailing assumptions in the field. However, our results also reveal representational differences: while DNNs' learning is characterized by a pronounced generalisation lag, humans appear to immediately acquire generalizable representations without a preliminary phase of learning training set-specific information that is only later transferred to novel data.
Recent works have demonstrated that it is possible to reconstruct training images and their labels from gradients of an image-classification model when its architecture is known. Unfortunately, there is still an incomplete theoretical understanding of the efficacy and failure of these gradient-leakage attacks. In this paper, we propose a novel framework to analyse training-data leakage from gradients that draws insights from both analytic and optimisation-based gradient-leakage attacks. We formulate the reconstruction problem as solving a linear system from each layer iteratively, accompanied by corrections using gradient matching. Under this framework, we claim that the solubility of the reconstruction problem is primarily determined by that of the linear system at each layer. As a result, we are able to partially attribute the leakage of the training data in a deep network to its architecture. We also propose a metric to measure the level of security of a deep learning model against gradient-based attacks on the training data.
Activation function is a pivotal component of deep learning, facilitating the extraction of intricate data patterns. While classical activation functions like ReLU and its variants are extensively utilized, their static nature and simplicity, despite being advantageous, often limit their effectiveness in specialized tasks. The trainable activation functions also struggle sometimes to adapt to the unique characteristics of the data. Addressing these limitations, we introduce a novel trainable activation function, adaptive piecewise approximated activation linear unit (APALU), to enhance the learning performance of deep learning across a broad range of tasks. It presents a unique set of features that enable it to maintain stability and efficiency in the learning process while adapting to complex data representations. Experiments reveal significant improvements over widely used activation functions for different tasks. In image classification, APALU increases MobileNet and GoogleNet accuracy by 0.37% and 0.04%, respectively, on the CIFAR10 dataset. In anomaly detection, it improves the average area under the curve of One-CLASS Deep SVDD by 0.8% on the MNIST dataset, 1.81% and 1.11% improvements with DifferNet, and knowledge distillation, respectively, on the MVTech dataset. Notably, APALU achieves 100% accuracy on a sign language recognition task with a limited dataset. For regression tasks, APALU enhances the performance of deep neural networks and recurrent neural networks on different datasets. These improvements highlight the robustness and adaptability of APALU across diverse deep-learning applications.