With the availability of data, hardware, software ecosystem and relevant skill sets, the machine learning community is undergoing a rapid development with new architectures and approaches appearing at high frequency every year. In this article, we conduct an exemplary image classification study in order to demonstrate how confidence intervals around accuracy measurements can greatly enhance the communication of research results as well as impact the reviewing process. In addition, we explore the hallmarks and limitations of this approximation. We discuss the relevance of this approach reflecting on a spotlight publication of ICLR22. A reproducible workflow is made available as an open-source adjoint to this publication. Based on our discussion, we make suggestions for improving the authoring and reviewing process of machine learning articles.
The application of zero-shot learning in computer vision has been revolutionized by the use of image-text matching models. The most notable example, CLIP, has been widely used for both zero-shot classification and guiding generative models with a text prompt. However, the zero-shot use of CLIP is unstable with respect to the phrasing of the input text, making it necessary to carefully engineer the prompts used. We find that this instability stems from a selective similarity score, which is based only on a subset of the semantically meaningful input tokens. To mitigate it, we present a novel explainability-based approach, which adds a loss term to ensure that CLIP focuses on all relevant semantic parts of the input, in addition to employing the CLIP similarity loss used in previous works. When applied to one-shot classification through prompt engineering, our method yields an improvement in the recognition rate, without additional training or fine-tuning. Additionally, we show that CLIP guidance of generative models using our method significantly improves the generated images. Finally, we demonstrate a novel use of CLIP guidance for text-based image generation with spatial conditioning on object location, by requiring the image explainability heatmap for each object to be confined to a pre-determined bounding box.
Evolutionary algorithms are sensitive to the mutation rate (MR); no single value of this parameter works well across domains. Self-adaptive MR approaches have been proposed but they tend to be brittle: Sometimes they decay the MR to zero, thus halting evolution. To make self-adaptive MR robust, this paper introduces the Group Elite Selection of Mutation Rates (GESMR) algorithm. GESMR co-evolves a population of solutions and a population of MRs, such that each MR is assigned to a group of solutions. The resulting best mutational change in the group, instead of average mutational change, is used for MR selection during evolution, thus avoiding the vanishing MR problem. With the same number of function evaluations and with almost no overhead, GESMR converges faster and to better solutions than previous approaches on a wide range of continuous test optimization problems. GESMR also scales well to high-dimensional neuroevolution for supervised image-classification tasks and for reinforcement learning control tasks. Remarkably, GESMR produces MRs that are optimal in the long-term, as demonstrated through a comprehensive look-ahead grid search. Thus, GESMR and its theoretical and empirical analysis demonstrate how self-adaptation can be harnessed to improve performance in several applications of evolutionary computation.
In this paper, we introduce a novel self-supervised learning (SSL) loss for image representation learning. There is a growing belief that generalization in deep neural networks is linked to their ability to discriminate object shapes. Since object shape is related to the location of its parts, we propose to detect those that have been artificially misplaced. We represent object parts with image tokens and train a ViT to detect which token has been combined with an incorrect positional embedding. We then introduce sparsity in the inputs to make the model more robust to occlusions and to speed up the training. We call our method DILEMMA, which stands for Detection of Incorrect Location EMbeddings with MAsked inputs. We apply DILEMMA to MoCoV3, DINO and SimCLR and show an improvement in their performance of respectively 4.41%, 3.97%, and 0.5% under the same training time and with a linear probing transfer on ImageNet-1K. We also show full fine-tuning improvements of MAE combined with our method on ImageNet-100. We evaluate our method via fine-tuning on common SSL benchmarks. Moreover, we show that when downstream tasks are strongly reliant on shape (such as in the YOGA-82 pose dataset), our pre-trained features yield a significant gain over prior work.
Driver observation models are rarely deployed under perfect conditions. In practice, illumination, camera placement and type differ from the ones present during training and unforeseen behaviours may occur at any time. While observing the human behind the steering wheel leads to more intuitive human-vehicle-interaction and safer driving, it requires recognition algorithms which do not only predict the correct driver state, but also determine their prediction quality through realistic and interpretable confidence measures. Reliable uncertainty estimates are crucial for building trust and are a serious obstacle for deploying activity recognition networks in real driving systems. In this work, we for the first time examine how well the confidence values of modern driver observation models indeed match the probability of the correct outcome and show that raw neural network-based approaches tend to significantly overestimate their prediction quality. To correct this misalignment between the confidence values and the actual uncertainty, we consider two strategies. First, we enhance two activity recognition models often used for driver observation with temperature scaling-an off-the-shelf method for confidence calibration in image classification. Then, we introduce Calibrated Action Recognition with Input Guidance (CARING)-a novel approach leveraging an additional neural network to learn scaling the confidences depending on the video representation. Extensive experiments on the Drive&Act dataset demonstrate that both strategies drastically improve the quality of model confidences, while our CARING model out-performs both, the original architectures and their temperature scaling enhancement, leading to best uncertainty estimates.
Deep neural networks exploiting millions of parameters are nowadays the norm in deep learning applications. This is a potential issue because of the great amount of computational resources needed for training, and of the possible loss of generalization performance of overparametrized networks. We propose in this paper a method for learning sparse neural topologies via a regularization technique which identifies non relevant weights and selectively shrinks their norm, while performing a classic update for relevant ones. This technique, which is an improvement of classical weight decay, is based on the definition of a regularization term which can be added to any loss functional regardless of its form, resulting in a unified general framework exploitable in many different contexts. The actual elimination of parameters identified as irrelevant is handled by an iterative pruning algorithm. We tested the proposed technique on different image classification and Natural language generation tasks, obtaining results on par or better then competitors in terms of sparsity and metrics, while achieving strong models compression.
Diversity conveys advantages in nature, yet homogeneous neurons typically comprise the layers of artificial neural networks. Here we construct neural networks from neurons that learn their own activation functions, quickly diversify, and subsequently outperform their homogeneous counterparts on image classification and nonlinear regression tasks. Sub-networks instantiate the neurons, which meta-learn especially efficient sets of nonlinear responses. Examples include conventional neural networks classifying digits and forecasting a van der Pol oscillator and physics-informed Hamiltonian neural networks learning H\'enon-Heiles stellar orbits and the swing of a video recorded pendulum clock. Such \textit{learned diversity} provides examples of dynamical systems selecting diversity over uniformity and elucidates the role of diversity in natural and artificial systems.
Estimating the preferences of consumers is of utmost importance for the fashion industry as appropriately leveraging this information can be beneficial in terms of profit. Trend detection in fashion is a challenging task due to the fast pace of change in the fashion industry. Moreover, forecasting the visual popularity of new garment designs is even more demanding due to lack of historical data. To this end, we propose MuQAR, a Multimodal Quasi-AutoRegressive deep learning architecture that combines two modules: (1) a multi-modal multi-layer perceptron processing categorical, visual and textual features of the product and (2) a quasi-autoregressive neural network modelling the "target" time series of the product's attributes along with the "exogenous" time series of all other attributes. We utilize computer vision, image classification and image captioning, for automatically extracting visual features and textual descriptions from the images of new products. Product design in fashion is initially expressed visually and these features represent the products' unique characteristics without interfering with the creative process of its designers by requiring additional inputs (e.g manually written texts). We employ the product's target attributes time series as a proxy of temporal popularity patterns, mitigating the lack of historical data, while exogenous time series help capture trends among interrelated attributes. We perform an extensive ablation analysis on two large scale image fashion datasets, Mallzee and SHIFT15m to assess the adequacy of MuQAR and also use the Amazon Reviews: Home and Kitchen dataset to assess generalisability to other domains. A comparative study on the VISUELLE dataset, shows that MuQAR is capable of competing and surpassing the domain's current state of the art by 4.65% and 4.8% in terms of WAPE and MAE respectively.
As clean ImageNet accuracy nears its ceiling, the research community is increasingly more concerned about robust accuracy under distributional shifts. While a variety of methods have been proposed to robustify neural networks, these techniques often target models trained on ImageNet classification. At the same time, it is a common practice to use ImageNet pretrained backbones for downstream tasks such as object detection, semantic segmentation, and image classification from different domains. This raises a question: Can these robust image classifiers transfer robustness to downstream tasks? For object detection and semantic segmentation, we find that a vanilla Swin Transformer, a variant of Vision Transformer tailored for dense prediction tasks, transfers robustness better than Convolutional Neural Networks that are trained to be robust to the corrupted version of ImageNet. For CIFAR10 classification, we find that models that are robustified for ImageNet do not retain robustness when fully fine-tuned. These findings suggest that current robustification techniques tend to emphasize ImageNet evaluations. Moreover, network architecture is a strong source of robustness when we consider transfer learning.
This survey focuses on categorizing and evaluating the methods of supernet optimization in the field of Neural Architecture Search (NAS). Supernet optimization involves training a single, over-parameterized network that encompasses the search space of all possible network architectures. The survey analyses supernet optimization methods based on their approaches to spatial and temporal optimization. Spatial optimization relates to optimizing the architecture and parameters of the supernet and its subnets, while temporal optimization deals with improving the efficiency of selecting architectures from the supernet. The benefits, limitations, and potential applications of these methods in various tasks and settings, including transferability, domain generalization, and Transformer models, are also discussed.
DeepTensor is a computationally efficient framework for low-rank decomposition of matrices and tensors using deep generative networks. We decompose a tensor as the product of low-rank tensor factors (e.g., a matrix as the outer product of two vectors), where each low-rank tensor is generated by a deep network (DN) that is trained in a self-supervised manner to minimize the mean-squared approximation error. Our key observation is that the implicit regularization inherent in DNs enables them to capture nonlinear signal structures (e.g., manifolds) that are out of the reach of classical linear methods like the singular value decomposition (SVD) and principal component analysis (PCA). Furthermore, in contrast to the SVD and PCA, whose performance deteriorates when the tensor's entries deviate from additive white Gaussian noise, we demonstrate that the performance of DeepTensor is robust to a wide range of distributions. We validate that DeepTensor is a robust and computationally efficient drop-in replacement for the SVD, PCA, nonnegative matrix factorization (NMF), and similar decompositions by exploring a range of real-world applications, including hyperspectral image denoising, 3D MRI tomography, and image classification. In particular, DeepTensor offers a 6dB signal-to-noise ratio improvement over standard denoising methods for signals corrupted by Poisson noise and learns to decompose 3D tensors 60 times faster than a single DN equipped with 3D convolutions.
Modern deep learning training procedures rely on model regularization techniques such as data augmentation methods, which generate training samples that increase the diversity of data and richness of label information. A popular recent method, mixup, uses convex combinations of pairs of original samples to generate new samples. However, as we show in our experiments, mixup can produce undesirable synthetic samples, where the data is sampled off the manifold and can contain incorrect labels. We propose $\zeta$-mixup, a generalization of mixup with provably and demonstrably desirable properties that allows convex combinations of $N \geq 2$ samples, leading to more realistic and diverse outputs that incorporate information from $N$ original samples by using a $p$-series interpolant. We show that, compared to mixup, $\zeta$-mixup better preserves the intrinsic dimensionality of the original datasets, which is a desirable property for training generalizable models. Furthermore, we show that our implementation of $\zeta$-mixup is faster than mixup, and extensive evaluation on controlled synthetic and 24 real-world natural and medical image classification datasets shows that $\zeta$-mixup outperforms mixup and traditional data augmentation techniques.
A pretrain-finetune strategy is widely used to reduce the overfitting that can occur when data is insufficient for CNN training. First few layers of a CNN pretrained on a large-scale RGB dataset are capable of acquiring general image characteristics which are remarkably effective in tasks targeted for different RGB datasets. However, when it comes down to hyperspectral domain where each domain has its unique spectral properties, the pretrain-finetune strategy no longer can be deployed in a conventional way while presenting three major issues: 1) inconsistent spectral characteristics among the domains (e.g., frequency range), 2) inconsistent number of data channels among the domains, and 3) absence of large-scale hyperspectral dataset. We seek to train a universal cross-domain model which can later be deployed for various spectral domains. To achieve, we physically furnish multiple inlets to the model while having a universal portion which is designed to handle the inconsistent spectral characteristics among different domains. Note that only the universal portion is used in the finetune process. This approach naturally enables the learning of our model on multiple domains simultaneously which acts as an effective workaround for the issue of the absence of large-scale dataset. We have carried out a study to extensively compare models that were trained using cross-domain approach with ones trained from scratch. Our approach was found to be superior both in accuracy and in training efficiency. In addition, we have verified that our approach effectively reduces the overfitting issue, enabling us to deepen the model up to 13 layers (from 9) without compromising the accuracy.
In this work, we introduce Dual Attention Vision Transformers (DaViT), a simple yet effective vision transformer architecture that is able to capture global context while maintaining computational efficiency. We propose approaching the problem from an orthogonal angle: exploiting self-attention mechanisms with both "spatial tokens" and "channel tokens". With spatial tokens, the spatial dimension defines the token scope, and the channel dimension defines the token feature dimension. With channel tokens, we have the inverse: the channel dimension defines the token scope, and the spatial dimension defines the token feature dimension. We further group tokens along the sequence direction for both spatial and channel tokens to maintain the linear complexity of the entire model. We show that these two self-attentions complement each other: (i) since each channel token contains an abstract representation of the entire image, the channel attention naturally captures global interactions and representations by taking all spatial positions into account when computing attention scores between channels; (ii) the spatial attention refines the local representations by performing fine-grained interactions across spatial locations, which in turn helps the global information modeling in channel attention. Extensive experiments show our DaViT achieves state-of-the-art performance on four different tasks with efficient computations. Without extra data, DaViT-Tiny, DaViT-Small, and DaViT-Base achieve 82.8%, 84.2%, and 84.6% top-1 accuracy on ImageNet-1K with 28.3M, 49.7M, and 87.9M parameters, respectively. When we further scale up DaViT with 1.5B weakly supervised image and text pairs, DaViT-Gaint reaches 90.4% top-1 accuracy on ImageNet-1K. Code is available at https://github.com/dingmyu/davit.
Optimization within a layer of a deep-net has emerged as a new direction for deep-net layer design. However, there are two main challenges when applying these layers to computer vision tasks: (a) which optimization problem within a layer is useful?; (b) how to ensure that computation within a layer remains efficient? To study question (a), in this work, we propose total variation (TV) minimization as a layer for computer vision. Motivated by the success of total variation in image processing, we hypothesize that TV as a layer provides useful inductive bias for deep-nets too. We study this hypothesis on five computer vision tasks: image classification, weakly supervised object localization, edge-preserving smoothing, edge detection, and image denoising, improving over existing baselines. To achieve these results we had to address question (b): we developed a GPU-based projected-Newton method which is $37\times$ faster than existing solutions.
Visual recognition is recently learned via either supervised learning on human-annotated image-label data or language-image contrastive learning with webly-crawled image-text pairs. While supervised learning may result in a more discriminative representation, language-image pretraining shows unprecedented zero-shot recognition capability, largely due to the different properties of data sources and learning objectives. In this work, we introduce a new formulation by combining the two data sources into a common image-text-label space. In this space, we propose a new learning paradigm, called Unified Contrastive Learning (UniCL) with a single learning objective to seamlessly prompt the synergy of two data types. Extensive experiments show that our UniCL is an effective way of learning semantically rich yet discriminative representations, universally for image recognition in zero-shot, linear-probe, fully finetuning and transfer learning scenarios. Particularly, it attains gains up to 9.2% and 14.5% in average on zero-shot recognition benchmarks over the language-image contrastive learning and supervised learning methods, respectively. In linear probe setting, it also boosts the performance over the two methods by 7.3% and 3.4%, respectively. Our study also indicates that UniCL stand-alone is a good learner on pure image-label data, rivaling the supervised learning methods across three image classification datasets and two types of vision backbones, ResNet and Swin Transformer. Code is available at https://github.com/microsoft/UniCL.
ImageNet serves as the primary dataset for evaluating the quality of computer-vision models. The common practice today is training each architecture with a tailor-made scheme, designed and tuned by an expert. In this paper, we present a unified scheme for training any backbone on ImageNet. The scheme, named USI (Unified Scheme for ImageNet), is based on knowledge distillation and modern tricks. It requires no adjustments or hyper-parameters tuning between different models, and is efficient in terms of training times. We test USI on a wide variety of architectures, including CNNs, Transformers, Mobile-oriented and MLP-only. On all models tested, USI outperforms previous state-of-the-art results. Hence, we are able to transform training on ImageNet from an expert-oriented task to an automatic seamless routine. Since USI accepts any backbone and trains it to top results, it also enables to perform methodical comparisons, and identify the most efficient backbones along the speed-accuracy Pareto curve. Implementation is available at:https://github.com/Alibaba-MIIL/Solving_ImageNet
A quantum neural network (QNN) is interpreted today as any quantum circuit with trainable continuous parameters. This work builds on previous works by the authors and addresses QNN for image classification with Novel Enhanced Quantum Representation of (NEQR) processed classical data where Principal component analysis (PCA) and Projected Quantum Kernel features (PQK) were investigated previously by the authors as a path to quantum advantage for the same classical dataset. For each of these cases the Fashion-MNIST dataset was downscaled using PCA to convert into quantum data where the classical NN easily outperformed the QNN. However, we demonstrated quantum advantage by using PQK where quantum models achieved more than ~90% accuracy surpassing their classical counterpart on the same training dataset as in the first case. In this current work, we use the same dataset fed into a QNN and compare that with performance of a classical NN model. We built an NEQR model circuit to pre-process the same data and feed the images into the QNN. Our results showed marginal improvements (only about ~5.0%) where the QNN performance with NEQR exceeded the performance of QNN without NEQR. We conclude that given the computational cost and the massive circuit depth associated with running NEQR, the advantage offered by this specific Quantum Image Processing (QIMP) algorithm is questionable at least for classical image dataset. No actual quantum computing hardware platform exists today that can support the circuit depth needed to run NEQR even for the reduced image sizes of our toy classical dataset.
We propose a unified look at jointly learning multiple vision tasks and visual domains through universal representations, a single deep neural network. Learning multiple problems simultaneously involves minimizing a weighted sum of multiple loss functions with different magnitudes and characteristics and thus results in unbalanced state of one loss dominating the optimization and poor results compared to learning a separate model for each problem. To this end, we propose distilling knowledge of multiple task/domain-specific networks into a single deep neural network after aligning its representations with the task/domain-specific ones through small capacity adapters. We rigorously show that universal representations achieve state-of-the-art performances in learning of multiple dense prediction problems in NYU-v2 and Cityscapes, multiple image classification problems from diverse domains in Visual Decathlon Dataset and cross-domain few-shot learning in MetaDataset. Finally we also conduct multiple analysis through ablation and qualitative studies.
Would you trust physicians if they cannot explain their decisions to you? Medical diagnostics using machine learning gained enormously in importance within the last decade. However, without further enhancements many state-of-the-art machine learning methods are not suitable for medical application. The most important reasons are insufficient data set quality and the black-box behavior of machine learning algorithms such as Deep Learning models. Consequently, end-users cannot correct the model's decisions and the corresponding explanations. The latter is crucial for the trustworthiness of machine learning in the medical domain. The research field explainable interactive machine learning searches for methods that address both shortcomings. This paper extends the explainable and interactive CAIPI algorithm and provides an interface to simplify human-in-the-loop approaches for image classification. The interface enables the end-user (1) to investigate and (2) to correct the model's prediction and explanation, and (3) to influence the data set quality. After CAIPI optimization with only a single counterexample per iteration, the model achieves an accuracy of $97.48\%$ on the Medical MNIST and $95.02\%$ on the Fashion MNIST. This accuracy is approximately equal to state-of-the-art Deep Learning optimization procedures. Besides, CAIPI reduces the labeling effort by approximately $80\%$.