Multimodal encoders like CLIP excel in tasks such as zero-shot image classification and cross-modal retrieval. However, they require excessive training data. We propose canonical similarity analysis (CSA), which uses two unimodal encoders to replicate multimodal encoders using limited data. CSA maps unimodal features into a multimodal space, using a new similarity score to retain only the multimodal information. CSA only involves the inference of unimodal encoders and a cubic-complexity matrix decomposition, eliminating the need for extensive GPU-based model training. Experiments show that CSA outperforms CLIP while requiring $50,000\times$ fewer multimodal data pairs to bridge the modalities given pre-trained unimodal encoders on ImageNet classification and misinformative news caption detection. CSA surpasses the state-of-the-art method to map unimodal features to multimodal features. We also demonstrate the ability of CSA with modalities beyond image and text, paving the way for future modality pairs with limited paired multimodal data but abundant unpaired unimodal data, such as lidar and text.
Machine unlearning aims to selectively remove specific knowledge from a model. Current methods, such as task arithmetic, rely on fine-tuning models on the forget set, generating a task vector, and subtracting it from the original model. However, we argue the effectiveness of this approach is highly sensitive to hyperparameter selection, necessitating careful validation to identify the best model among many fine-tuned candidates. In this paper, we propose a novel method that leverages all given fine-tuned models rather than selecting a single one. By constructing task vectors from models trained with varied hyperparameters and merging only the components of the task vectors with consistent signs, we perform unlearning by negating the merged task vector from the original model. Given that existing methods also utilize multiple fine-tuned models, our approach delivers more effective unlearning without incurring additional computational costs. We demonstrate the effectiveness of our method on both vision-language models and standard image classification models, showing improved unlearning performance with minimal degradation on the retain set, outperforming state-of-the-art techniques.
Graph Neural Networks (GNNs) have gained popularity in various learning tasks, with successful applications in fields like molecular biology, transportation systems, and electrical grids. These fields naturally use graph data, benefiting from GNNs' message-passing framework. However, the potential of GNNs in more general data representations, especially in the image domain, remains underexplored. Leveraging the manifold hypothesis, which posits that high-dimensional data lies in a low-dimensional manifold, we explore GNNs' potential in this context. We construct an image manifold using variational autoencoders, then sample the manifold to generate graphs where each node is an image. This approach reduces data dimensionality while preserving geometric information. We then train a GNN to predict node labels corresponding to the image labels in the classification task, and leverage convergence of GNNs to manifold neural networks to analyze GNN generalization. Experiments on MNIST and CIFAR10 datasets demonstrate that GNNs generalize effectively to unseen graphs, achieving competitive accuracy in classification tasks.
Conformal prediction has recently emerged as a promising strategy for quantifying the uncertainty of a predictive model; these algorithms modify the model to output sets of labels that are guaranteed to contain the true label with high probability. However, existing conformal prediction algorithms have largely targeted classification and regression settings, where the structure of the prediction set has a simple form as a level set of the scoring function. However, for complex structured outputs such as text generation, these prediction sets might include a large number of labels and therefore be hard for users to interpret. In this paper, we propose a general framework for conformal prediction in the structured prediction setting, that modifies existing conformal prediction algorithms to output structured prediction sets that implicitly represent sets of labels. In addition, we demonstrate how our approach can be applied in domains where the prediction sets can be represented as a set of nodes in a directed acyclic graph; for instance, for hierarchical labels such as image classification, a prediction set might be a small subset of coarse labels implicitly representing the prediction set of all their more fine-descendants. We demonstrate how our algorithm can be used to construct prediction sets that satisfy a desired coverage guarantee in several domains.
Recent work developed convolutional deep kernel machines, achieving 92.7% test accuracy on CIFAR-10 using a ResNet-inspired architecture, which is SOTA for kernel methods. However, this still lags behind neural networks, which easily achieve over 94% test accuracy with similar architectures. In this work we introduce several modifications to improve the convolutional deep kernel machine's generalisation, including stochastic kernel regularisation, which adds noise to the learned Gram matrices during training. The resulting model achieves 94.5% test accuracy on CIFAR-10. This finding has important theoretical and practical implications, as it demonstrates that the ability to perform well on complex tasks like image classification is not unique to neural networks. Instead, other approaches including deep kernel methods can achieve excellent performance on such tasks, as long as they have the capacity to learn representations from data.
In this empirical article, we introduce INNAprop, an optimization algorithm that combines the INNA method with the RMSprop adaptive gradient scaling. It leverages second-order information and rescaling while keeping the memory requirements of standard DL methods as AdamW or SGD with momentum. After giving geometrical insights, we evaluate INNAprop on CIFAR-10, Food101, and ImageNet with ResNets, VGG, DenseNet, and ViT, and on GPT-2 (OpenWebText) train from scratch and with LoRA fine-tuning (E2E). INNAprop consistently matches or outperforms AdamW both in training speed and accuracy, with minimal hyperparameter tuning in large-scale settings. Our code is publicly available at \url{https://github.com/innaprop/innaprop}.
In this work, we propose a mean-squared error-based risk that enables the comparison and optimization of estimators of squared calibration errors in practical settings. Improving the calibration of classifiers is crucial for enhancing the trustworthiness and interpretability of machine learning models, especially in sensitive decision-making scenarios. Although various calibration (error) estimators exist in the current literature, there is a lack of guidance on selecting the appropriate estimator and tuning its hyperparameters. By leveraging the bilinear structure of squared calibration errors, we reformulate calibration estimation as a regression problem with independent and identically distributed (i.i.d.) input pairs. This reformulation allows us to quantify the performance of different estimators even for the most challenging calibration criterion, known as canonical calibration. Our approach advocates for a training-validation-testing pipeline when estimating a calibration error on an evaluation dataset. We demonstrate the effectiveness of our pipeline by optimizing existing calibration estimators and comparing them with novel kernel ridge regression-based estimators on standard image classification tasks.
Deploying large and complex deep neural networks on resource-constrained edge devices poses significant challenges due to their computational demands and the complexities of non-convex optimization. Traditional compression methods such as distillation and pruning often retain non-convexity that complicates fine-tuning in real-time on such devices. Moreover, these methods often necessitate extensive end-to-end network fine-tuning after compression to preserve model performance, which is not only time-consuming but also requires fully annotated datasets, thus potentially negating the benefits of efficient network compression. In this paper, we introduce a novel distillation technique that efficiently compresses the model via convex optimization -- eliminating intermediate non-convex activation functions and using only intermediate activations from the original model. Our approach enables distillation in a label-free data setting and achieves performance comparable to the original model without requiring any post-compression fine-tuning. We demonstrate the effectiveness of our method for image classification models on multiple standard datasets, and further show that in the data limited regime, our method can outperform standard non-convex distillation approaches. Our method promises significant advantages for deploying high-efficiency, low-footprint models on edge devices, making it a practical choice for real-world applications. We show that convex neural networks, when provided with rich feature representations from a large pre-trained non-convex model, can achieve performance comparable to their non-convex counterparts, opening up avenues for future research at the intersection of convex optimization and deep learning.
Deep networks are frequently tuned to novel tasks and continue learning from ongoing data streams. Such sequential training requires consolidation of new and past information, a challenge predominantly addressed by retaining the most important data points - formally known as coresets. Traditionally, these coresets consist of entire samples, such as images or sentences. However, recent transformer architectures operate on tokens, leading to the famous assertion that an image is worth 16x16 words. Intuitively, not all of these tokens are equally informative or memorable. Going beyond coresets, we thus propose to construct a deeper-level data summary on the level of tokens. Our respectively named core tokensets both select the most informative data points and leverage feature attribution to store only their most relevant features. We demonstrate that core tokensets yield significant performance retention in incremental image classification, open-ended visual question answering, and continual image captioning with significantly reduced memory. In fact, we empirically find that a core tokenset of 1\% of the data performs comparably to at least a twice as large and up to 10 times larger coreset.
Predicting the neural response to natural images in the visual cortex requires extracting relevant features from the images and relating those feature to the observed responses. In this work, we optimize the feature extraction in order to maximize the information shared between the image features and the neural response across voxels in a given region of interest (ROI) extracted from the BOLD signal measured by fMRI. We adapt contrastive learning (CL) to fine-tune a convolutional neural network, which was pretrained for image classification, such that a mapping of a given image's features are more similar to the corresponding fMRI response than to the responses to other images. We exploit the recently released Natural Scenes Dataset (Allen et al., 2022) as organized for the Algonauts Project (Gifford et al., 2023), which contains the high-resolution fMRI responses of eight subjects to tens of thousands of naturalistic images. We show that CL fine-tuning creates feature extraction models that enable higher encoding accuracy in early visual ROIs as compared to both the pretrained network and a baseline approach that uses a regression loss at the output of the network to tune it for fMRI response encoding. We investigate inter-subject transfer of the CL fine-tuned models, including subjects from another, lower-resolution dataset (Gong et al., 2023). We also pool subjects for fine-tuning to further improve the encoding performance. Finally, we examine the performance of the fine-tuned models on common image classification tasks, explore the landscape of ROI-specific models by applying dimensionality reduction on the Bhattacharya dissimilarity matrix created using the predictions on those tasks (Mao et al., 2024), and investigate lateralization of the processing for early visual ROIs using salience maps of the classifiers built on the CL-tuned models.
Although it is traditionally believed that lossy image compression, such as JPEG compression, has a negative impact on the performance of deep neural networks (DNNs), it is shown by recent works that well-crafted JPEG compression can actually improve the performance of deep learning (DL). Inspired by this, we propose JPEG-DL, a novel DL framework that prepends any underlying DNN architecture with a trainable JPEG compression layer. To make the quantization operation in JPEG compression trainable, a new differentiable soft quantizer is employed at the JPEG layer, and then the quantization operation and underlying DNN are jointly trained. Extensive experiments show that in comparison with the standard DL, JPEG-DL delivers significant accuracy improvements across various datasets and model architectures while enhancing robustness against adversarial attacks. Particularly, on some fine-grained image classification datasets, JPEG-DL can increase prediction accuracy by as much as 20.9%. Our code is available on https://github.com/JpegInspiredDl/JPEG-Inspired-DL.git.
Recent advancements in State Space Models, notably Mamba, have demonstrated superior performance over the dominant Transformer models, particularly in reducing the computational complexity from quadratic to linear. Yet, difficulties in adapting Mamba from language to vision tasks arise due to the distinct characteristics of visual data, such as the spatial locality and adjacency within images and large variations in information granularity across visual tokens. Existing vision Mamba approaches either flatten tokens into sequences in a raster scan fashion, which breaks the local adjacency of images, or manually partition tokens into windows, which limits their long-range modeling and generalization capabilities. To address these limitations, we present a new vision Mamba model, coined QuadMamba, that effectively captures local dependencies of varying granularities via quadtree-based image partition and scan. Concretely, our lightweight quadtree-based scan module learns to preserve the 2D locality of spatial regions within learned window quadrants. The module estimates the locality score of each token from their features, before adaptively partitioning tokens into window quadrants. An omnidirectional window shifting scheme is also introduced to capture more intact and informative features across different local regions. To make the discretized quadtree partition end-to-end trainable, we further devise a sequence masking strategy based on Gumbel-Softmax and its straight-through gradient estimator. Extensive experiments demonstrate that QuadMamba achieves state-of-the-art performance in various vision tasks, including image classification, object detection, instance segmentation, and semantic segmentation. The code is in https://github.com/VISION-SJTU/QuadMamba.
Foundation models (FMs) are pre-trained on large-scale datasets and then fine-tuned for a specific downstream task. The most common fine-tuning method is to update pretrained weights via low-rank adaptation (LoRA). Existing initialization strategies for LoRA often rely on singular value decompositions (SVD) of gradients or weight matrices. However, they do not provably maximize the expected gradient signal, which is critical for fast adaptation. To this end, we introduce Explained Variance Adaptation (EVA), an initialization scheme that uses the directions capturing the most activation variance, provably maximizing the expected gradient signal and accelerating fine-tuning. EVA performs incremental SVD on minibatches of activation vectors and selects the right-singular vectors for initialization once they converged. Further, by selecting the directions that capture the most activation-variance for a given rank budget, EVA accommodates adaptive ranks that reduce the number of trainable parameters, while maintaining or improving downstream performance. We apply EVA to a variety of fine-tuning tasks as language generation and understanding, image classification, and reinforcement learning. EVA exhibits faster convergence than competitors and achieves the highest average score across a multitude of tasks per domain while reducing the number of trainable parameters through rank redistribution.
This work describes an approach towards pixel quantization using variable resolution which is made feasible using image transformation in the analog domain. The main aim is to reduce the average bits-per-pixel (BPP) necessary for representing an image while maintaining the classification accuracy of a Convolutional Neural Network (CNN) that is trained for image classification. The proposed algorithm is based on the Hadamard transform that leads to a low-resolution variable quantization by the analog-to-digital converter (ADC) thus reducing the power dissipation in hardware at the sensor node. Despite the trade-offs inherent in image transformation, the proposed algorithm achieves competitive accuracy levels across various image sizes and ADC configurations, highlighting the importance of considering both accuracy and power consumption in edge computing applications. The schematic of a novel 1.5 bit ADC that incorporates the Hadamard transform is also proposed. A hardware implementation of the analog transformation followed by software-based variable quantization is done for the CIFAR-10 test dataset. The digitized data shows that the network can still identify transformed images with a remarkable 90% accuracy for 3-BPP transformed images following the proposed method.
Hyperspectral image (HSI) classification has garnered substantial attention in remote sensing fields. Recent Mamba architectures built upon the Selective State Space Models (S6) have demonstrated enormous potential in long-range sequence modeling. However, the high dimensionality of hyperspectral data and information redundancy pose challenges to the application of Mamba in HSI classification, suffering from suboptimal performance and computational efficiency. In light of this, this paper investigates a lightweight Interval Group Spatial-Spectral Mamba framework (IGroupSS-Mamba) for HSI classification, which allows for multi-directional and multi-scale global spatial-spectral information extraction in a grouping and hierarchical manner. Technically, an Interval Group S6 Mechanism (IGSM) is developed as the core component, which partitions high-dimensional features into multiple non-overlapping groups at intervals, and then integrates a unidirectional S6 for each group with a specific scanning direction to achieve non-redundant sequence modeling. Compared to conventional applying multi-directional scanning to all bands, this grouping strategy leverages the complementary strengths of different scanning directions while decreasing computational costs. To adequately capture the spatial-spectral contextual information, an Interval Group Spatial-Spectral Block (IGSSB) is introduced, in which two IGSM-based spatial and spectral operators are cascaded to characterize the global spatial-spectral relationship along the spatial and spectral dimensions, respectively. IGroupSS-Mamba is constructed as a hierarchical structure stacked by multiple IGSSB blocks, integrating a pixel aggregation-based downsampling strategy for multiscale spatial-spectral semantic learning from shallow to deep stages. Extensive experiments demonstrate that IGroupSS-Mamba outperforms the state-of-the-art methods.
We initiate a study of the geometry of the visual representation space -- the information channel from the vision encoder to the action decoder -- in an image-based control pipeline learned from behavior cloning. Inspired by the phenomenon of neural collapse (NC) in image classification (arXiv:2008.08186), we empirically demonstrate the prevalent emergence of a similar law of clustering in the visual representation space. Specifically, in discrete image-based control (e.g., Lunar Lander), the visual representations cluster according to the natural discrete action labels; in continuous image-based control (e.g., Planar Pushing and Block Stacking), the clustering emerges according to "control-oriented" classes that are based on (a) the relative pose between the object and the target in the input or (b) the relative pose of the object induced by expert actions in the output. Each of the classes corresponds to one relative pose orthant (REPO). Beyond empirical observation, we show such a law of clustering can be leveraged as an algorithmic tool to improve test-time performance when training a policy with limited expert demonstrations. Particularly, we pretrain the vision encoder using NC as a regularization to encourage control-oriented clustering of the visual features. Surprisingly, such an NC-pretrained vision encoder, when finetuned end-to-end with the action decoder, boosts the test-time performance by 10% to 35%. Real-world vision-based planar pushing experiments confirmed the surprising advantage of control-oriented visual representation pretraining.
Data curation is the problem of how to collect and organize samples into a dataset that supports efficient learning. Despite the centrality of the task, little work has been devoted towards a large-scale, systematic comparison of various curation methods. In this work, we take steps towards a formal evaluation of data curation strategies and introduce SELECT, the first large-scale benchmark of curation strategies for image classification. In order to generate baseline methods for the SELECT benchmark, we create a new dataset, ImageNet++, which constitutes the largest superset of ImageNet-1K to date. Our dataset extends ImageNet with 5 new training-data shifts, each approximately the size of ImageNet-1K itself, and each assembled using a distinct curation strategy. We evaluate our data curation baselines in two ways: (i) using each training-data shift to train identical image classification models from scratch (ii) using the data itself to fit a pretrained self-supervised representation. Our findings show interesting trends, particularly pertaining to recent methods for data curation such as synthetic data generation and lookup based on CLIP embeddings. We show that although these strategies are highly competitive for certain tasks, the curation strategy used to assemble the original ImageNet-1K dataset remains the gold standard. We anticipate that our benchmark can illuminate the path for new methods to further reduce the gap. We release our checkpoints, code, documentation, and a link to our dataset at https://github.com/jimmyxu123/SELECT.
Art authentication has historically established itself as a task requiring profound connoisseurship of one particular artist. Nevertheless, famous art forgers such as Wolfgang Beltracchi were able to deceive dozens of art experts. In recent years Artificial Intelligence algorithms have been successfully applied to various image processing tasks. In this work, we leverage the growing improvements in AI to present an art authentication framework for the identification of the forger Wolfgang Beltracchi. Differently from existing literature on AI-aided art authentication, we focus on a specialized model of a forger, rather than an artist, flipping the approach of traditional AI methods. We use a carefully compiled dataset of known artists forged by Beltracchi and a set of known works by the forger to train a multiclass image classification model based on EfficientNet. We compare the results with Kolmogorov Arnold Networks (KAN) which, to the best of our knowledge, have never been tested in the art domain. The results show a general agreement between the different models' predictions on artworks flagged as forgeries, which are then closely studied using visual analysis.
Whole slide image (WSI) classification is a crucial problem for cancer diagnostics in clinics and hospitals. A WSI, acquired at gigapixel size, is commonly tiled into patches and processed by multiple-instance learning (MIL) models. Previous MIL-based models designed for this problem have only been evaluated on individual tasks for specific organs, and the ability to handle multiple tasks within a single model has not been investigated. In this study, we propose MECFormer, a generative Transformer-based model designed to handle multiple tasks within one model. To leverage the power of learning multiple tasks simultaneously and to enhance the model's effectiveness in focusing on each individual task, we introduce an Expert Consultation Network, a projection layer placed at the beginning of the Transformer-based model. Additionally, to enable flexible classification, autoregressive decoding is incorporated by a language decoder for WSI classification. Through extensive experiments on five datasets involving four different organs, one cancer classification task, and four cancer subtyping tasks, MECFormer demonstrates superior performance compared to individual state-of-the-art multiple-instance learning models.
This paper introduces Idempotent Test-Time Training (IT$^3$), a novel approach to addressing the challenge of distribution shift. While supervised-learning methods assume matching train and test distributions, this is rarely the case for machine learning systems deployed in the real world. Test-Time Training (TTT) approaches address this by adapting models during inference, but they are limited by a domain specific auxiliary task. IT$^3$ is based on the universal property of idempotence. An idempotent operator is one that can be applied sequentially without changing the result beyond the initial application, that is $f(f(x))=f(x)$. At training, the model receives an input $x$ along with another signal that can either be the ground truth label $y$ or a neutral "don't know" signal $0$. At test time, the additional signal can only be $0$. When sequentially applying the model, first predicting $y_0 = f(x, 0)$ and then $y_1 = f(x, y_0)$, the distance between $y_0$ and $y_1$ measures certainty and indicates out-of-distribution input $x$ if high. We use this distance, that can be expressed as $||f(x, f(x, 0)) - f(x, 0)||$ as our TTT loss during inference. By carefully optimizing this objective, we effectively train $f(x,\cdot)$ to be idempotent, projecting the internal representation of the input onto the training distribution. We demonstrate the versatility of our approach across various tasks, including corrupted image classification, aerodynamic predictions, tabular data with missing information, age prediction from face, and large-scale aerial photo segmentation. Moreover, these tasks span different architectures such as MLPs, CNNs, and GNNs.