Escaping The Big Data Paradigm in Self-Supervised Representation Learning

The reliance on large-scale datasets and extensive computational resources has become a major barrier to advancing representation learning in vision, especially in data-scarce domains. In this paper, we address the critical question: Can we escape the big data paradigm in self-supervised representation learning from images? We introduce SCOTT (Sparse Convolutional Tokenizer for Transformers), a shallow tokenization architecture that is compatible with Masked Image Modeling (MIM) tasks. SCOTT injects convolutional inductive biases into Vision Transformers (ViTs), enhancing their efficacy in small-scale data regimes. Alongside, we propose to train on a Joint-Embedding Predictive Architecture within a MIM framework (MIM-JEPA), operating in latent representation space to capture more semantic features. Our approach enables ViTs to be trained from scratch on datasets orders of magnitude smaller than traditionally required --without relying on massive external datasets for pretraining. We validate our method on three small-size, standard-resoultion, fine-grained datasets: Oxford Flowers-102, Oxford IIIT Pets-37, and ImageNet-100. Despite the challenges of limited data and high intra-class similarity, frozen SCOTT models pretrained with MIM-JEPA significantly outperform fully supervised methods and achieve competitive results with SOTA approaches that rely on large-scale pretraining, complex image augmentations and bigger model sizes. By demonstrating that robust off-the-shelf representations can be learned with limited data, compute, and model sizes, our work paves the way for computer applications in resource constrained environments such as medical imaging or robotics. Our findings challenge the prevailing notion that vast amounts of data are indispensable for effective representation learning in vision, offering a new pathway toward more accessible and inclusive advancements in the field.

ReViT: Enhancing Vision Transformers with Attention Residual Connections for Visual Recognition

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.

Mixture of Gaussian-distributed Prototypes with Generative Modelling for Interpretable Image Classification

Prototypical-part interpretable methods, e.g., ProtoPNet, enhance interpretability by connecting classification predictions to class-specific training prototypes, thereby offering an intuitive insight into their decision-making. Current methods rely on a discriminative classifier trained with point-based learning techniques that provide specific values for prototypes. Such prototypes have relatively low representation power due to their sparsity and potential redundancy, with each prototype containing no variability measure. In this paper, we present a new generative learning of prototype distributions, named Mixture of Gaussian-distributed Prototypes (MGProto), which are represented by Gaussian mixture models (GMM). Such an approach enables the learning of more powerful prototype representations since each learned prototype will own a measure of variability, which naturally reduces the sparsity given the spread of the distribution around each prototype, and we also integrate a prototype diversity objective function into the GMM optimisation to reduce redundancy. Incidentally, the generative nature of MGProto offers a new and effective way for detecting out-of-distribution samples. To improve the compactness of MGProto, we further propose to prune Gaussian-distributed prototypes with a low prior. Experiments on CUB-200-2011, Stanford Cars, Stanford Dogs, and Oxford-IIIT Pets datasets show that MGProto achieves state-of-the-art classification and OoD detection performances with encouraging interpretability results.

Directly Optimizing IoU for Bounding Box Localization

Object detection has seen remarkable progress in recent years with the introduction of Convolutional Neural Networks (CNN). Object detection is a multi-task learning problem where both the position of the objects in the images as well as their classes needs to be correctly identified. The idea here is to maximize the overlap between the ground-truth bounding boxes and the predictions i.e. the Intersection over Union (IoU). In the scope of work seen currently in this domain, IoU is approximated by using the Huber loss as a proxy but this indirect method does not leverage the IoU information and treats the bounding box as four independent, unrelated terms of regression. This is not true for a bounding box where the four coordinates are highly correlated and hold a semantic meaning when taken together. The direct optimization of the IoU is not possible due to its non-convex and non-differentiable nature. In this paper, we have formulated a novel loss namely, the Smooth IoU, which directly optimizes the IoUs for the bounding boxes. This loss has been evaluated on the Oxford IIIT Pets, Udacity self-driving car, PASCAL VOC, and VWFS Car Damage datasets and has shown performance gains over the standard Huber loss.

Scale Equivariant U-Net

In neural networks, the property of being equivariant to transformations improves generalization when the corresponding symmetry is present in the data. In particular, scale-equivariant networks are suited to computer vision tasks where the same classes of objects appear at different scales, like in most semantic segmentation tasks. Recently, convolutional layers equivariant to a semigroup of scalings and translations have been proposed. However, the equivariance of subsampling and upsampling has never been explicitly studied even though they are necessary building blocks in some segmentation architectures. The U-Net is a representative example of such architectures, which includes the basic elements used for state-of-the-art semantic segmentation. Therefore, this paper introduces the Scale Equivariant U-Net (SEU-Net), a U-Net that is made approximately equivariant to a semigroup of scales and translations through careful application of subsampling and upsampling layers and the use of aforementioned scale-equivariant layers. Moreover, a scale-dropout is proposed in order to improve generalization to different scales in approximately scale-equivariant architectures. The proposed SEU-Net is trained for semantic segmentation of the Oxford Pet IIIT and the DIC-C2DH-HeLa dataset for cell segmentation. The generalization metric to unseen scales is dramatically improved in comparison to the U-Net, even when the U-Net is trained with scale jittering, and to a scale-equivariant architecture that does not perform upsampling operators inside the equivariant pipeline. The scale-dropout induces better generalization on the scale-equivariant models in the Pet experiment, but not on the cell segmentation experiment.

Towards Fine-grained Image Classification with Generative Adversarial Networks and Facial Landmark Detection

Fine-grained classification remains a challenging task because distinguishing categories needs learning complex and local differences. Diversity in the pose, scale, and position of objects in an image makes the problem even more difficult. Although the recent Vision Transformer models achieve high performance, they need an extensive volume of input data. To encounter this problem, we made the best use of GAN-based data augmentation to generate extra dataset instances. Oxford-IIIT Pets was our dataset of choice for this experiment. It consists of 37 breeds of cats and dogs with variations in scale, poses, and lighting, which intensifies the difficulty of the classification task. Furthermore, we enhanced the performance of the recent Generative Adversarial Network (GAN), StyleGAN2-ADA model to generate more realistic images while preventing overfitting to the training set. We did this by training a customized version of MobileNetV2 to predict animal facial landmarks; then, we cropped images accordingly. Lastly, we combined the synthetic images with the original dataset and compared our proposed method with standard GANs augmentation and no augmentation with different subsets of training data. We validated our work by evaluating the accuracy of fine-grained image classification on the recent Vision Transformer (ViT) Model.

Deep Spiking Convolutional Neural Network for Single Object Localization Based On Deep Continuous Local Learning

With the advent of neuromorphic hardware, spiking neural networks can be a good energy-efficient alternative to artificial neural networks. However, the use of spiking neural networks to perform computer vision tasks remains limited, mainly focusing on simple tasks such as digit recognition. It remains hard to deal with more complex tasks (e.g. segmentation, object detection) due to the small number of works on deep spiking neural networks for these tasks. The objective of this paper is to make the first step towards modern computer vision with supervised spiking neural networks. We propose a deep convolutional spiking neural network for the localization of a single object in a grayscale image. We propose a network based on DECOLLE, a spiking model that enables local surrogate gradient-based learning. The encouraging results reported on Oxford-IIIT-Pet validates the exploitation of spiking neural networks with a supervised learning approach for more elaborate vision tasks in the future.

Distance Metric Learned Collaborative Representation Classifier

Any generic deep machine learning algorithm is essentially a function fitting exercise, where the network tunes its weights and parameters to learn discriminatory features by minimizing some cost function. Though the network tries to learn the optimal feature space, it seldom tries to learn an optimal distance metric in the cost function, and hence misses out on an additional layer of abstraction. We present a simple effective way of achieving this by learning a generic Mahalanabis distance in a collaborative loss function in an end-to-end fashion with any standard convolutional network as the feature learner. The proposed method DML-CRC gives state-of-the-art performance on benchmark fine-grained classification datasets CUB Birds, Oxford Flowers and Oxford-IIIT Pets using the VGG-19 deep network. The method is network agnostic and can be used for any similar classification tasks.