Sparse mixture of expert architectures (MoEs) scale model capacity without significant increases in training or inference costs. Despite their success, MoEs suffer from a number of issues: training instability, token dropping, inability to scale the number of experts, or ineffective finetuning. In this work, we propose Soft MoE, a fully-differentiable sparse Transformer that addresses these challenges, while maintaining the benefits of MoEs. Soft MoE performs an implicit soft assignment by passing different weighted combinations of all input tokens to each expert. As in other MoEs, experts in Soft MoE only process a subset of the (combined) tokens, enabling larger model capacity (and performance) at lower inference cost. In the context of visual recognition, Soft MoE greatly outperforms dense Transformers (ViTs) and popular MoEs (Tokens Choice and Experts Choice). Soft MoE scales well: Soft MoE Huge/14 with 128 experts in 16 MoE layers has over 40x more parameters than ViT Huge/14, with only 2% increased inference time, and substantially better quality.

The scaling of Transformers has driven breakthrough capabilities for language models. At present, the largest large language models (LLMs) contain upwards of 100B parameters. Vision Transformers (ViT) have introduced the same architecture to image and video modelling, but these have not yet been successfully scaled to nearly the same degree; the largest dense ViT contains 4B parameters (Chen et al., 2022). We present a recipe for highly efficient and stable training of a 22B-parameter ViT (ViT-22B) and perform a wide variety of experiments on the resulting model. When evaluated on downstream tasks (often with a lightweight linear model on frozen features), ViT-22B demonstrates increasing performance with scale. We further observe other interesting benefits of scale, including an improved tradeoff between fairness and performance, state-of-the-art alignment to human visual perception in terms of shape/texture bias, and improved robustness. ViT-22B demonstrates the potential for "LLM-like" scaling in vision, and provides key steps towards getting there.

The top-$k$ operator returns a $k$-sparse vector, where the non-zero values correspond to the $k$ largest values of the input. Unfortunately, because it is a discontinuous function, it is difficult to incorporate in neural networks trained end-to-end with backpropagation. Recent works have considered differentiable relaxations, based either on regularization or perturbation techniques. However, to date, no approach is fully differentiable and sparse. In this paper, we propose new differentiable and sparse top-$k$ operators. We view the top-$k$ operator as a linear program over the permutahedron, the convex hull of permutations. We then introduce a $p$-norm regularization term to smooth out the operator, and show that its computation can be reduced to isotonic optimization. Our framework is significantly more general than the existing one and allows for example to express top-$k$ operators that select values in magnitude. On the algorithmic side, in addition to pool adjacent violator (PAV) algorithms, we propose a new GPU/TPU-friendly Dykstra algorithm to solve isotonic optimization problems. We successfully use our operators to prune weights in neural networks, to fine-tune vision transformers, and as a router in sparse mixture of experts.

The ubiquitous and demonstrably suboptimal choice of resizing images to a fixed resolution before processing them with computer vision models has not yet been successfully challenged. However, models such as the Vision Transformer (ViT) offer flexible sequence-based modeling, and hence varying input sequence lengths. We take advantage of this with NaViT (Native Resolution ViT) which uses sequence packing during training to process inputs of arbitrary resolutions and aspect ratios. Alongside flexible model usage, we demonstrate improved training efficiency for large-scale supervised and contrastive image-text pretraining. NaViT can be efficiently transferred to standard tasks such as image and video classification, object detection, and semantic segmentation and leads to improved results on robustness and fairness benchmarks. At inference time, the input resolution flexibility can be used to smoothly navigate the test-time cost-performance trade-off. We believe that NaViTmarks a departure from the standard, CNN-designed, input and modelling pipeline used by most computer vision models, and represents a promising direction for ViTs.

Abstract Keyword Spotting (KWS) is here considered as a basic technology for Probabilistic Indexing (PrIx) of large collections of handwritten text images to allow fast textual access to the contents of these collections. Under this perspective, a probabilistic framework for lexicon-based KWS in text images is presented. The presentation aims at providing formal insights which help understanding classical statements of KWS (from which PrIx borrows fundamental concepts), as well as the relative challenges entailed by these statements. The development of the proposed framework makes it clear that word recognition or classification implicitly or explicitly underlies any formulation of KWS. Moreover, it suggests that the same statistical models and training methods successfully used for handwriting text recognition can advantageously be used also for PrIx, even though PrIx does not generally require or rely on any kind of previously produced image transcripts. Experiments carried out using these approaches support the consistency and the general interest of the proposed framework. Results on three datasets traditionally used for KWS benchmarking are significantly better than those previously published for these datasets. In addition, good results are also reported on two new, larger handwritten text image datasets (B entham and P lantas ), showing the great potential of the methods proposed in this paper for indexing and textual search in large collections of untranscribed handwritten documents. Specifically, we achieved the following Average Precision values: IAMDB: 0.89, G eorge W ashington : 0.91, P arzival : 0.95, B entham : 0.91 and P lantas : 0.92.

Transfer learning has been recently popularized as a data-efficient alternative to training models from scratch, in particular for computer vision tasks where it provides a remarkably solid baseline. The emergence of rich model repositories, such as TensorFlow Hub, enables the practitioners and researchers to unleash the potential of these models across a wide range of downstream tasks. As these repositories keep growing exponentially, efficiently selecting a good model for the task at hand becomes paramount. We provide a formalization of this problem through afamiliar notion of regret and introduce the predominant strategies, namely task-agnostic (e.g. ranking models by their ImageNet performance) and task-aware search strategies (such as linear or kNN evaluation). We conduct a large-scale empirical study and show that both task-agnostic and task-aware methods can yield high regret. We then propose a simple and computationally efficient hybrid search strategy which outperforms the existing approaches. We highlight the practical benefits of the proposed solution on a set of 19 diverse vision tasks.

Large sparsely-activated models have obtained excellent performance in multiple domains.However, such models are typically trained on a single modality at a time.We present the Language-Image MoE, LIMoE, a sparse mixture of experts model capable of multimodal learning.LIMoE accepts both images and text simultaneously, while being trained using a contrastive loss.MoEs are a natural fit for a multimodal backbone, since expert layers can learn an appropriate partitioning of modalities.However, new challenges arise; in particular, training stability and balanced expert utilization, for which we propose an entropy-based regularization scheme.Across multiple scales, we demonstrate performance improvement over dense models of equivalent computational cost.LIMoE-L/16 trained comparably to CLIP-L/14 achieves 77.9% zero-shot ImageNet accuracy (vs. 76.2%), and when further scaled to H/14 (with additional data) it achieves 83.8%, approaching state-of-the-art methods which use custom per-modality backbones and pre-training schemes.We analyse the quantitative and qualitative behavior of LIMoE, and demonstrate phenomena such as differing treatment of the modalities and the emergence of modality-specific experts.

Adversarial robustness is a key desirable property of neural networks. It has been empirically shown to be affected by their sizes, with larger networks being typically more robust. Recently, \citet{bubeck2021universal} proved a lower bound on the Lipschitz constant of functions that fit the training data in terms of their number of parameters. This raises an interesting open question, do---and can---functions with more parameters, but not necessarily more computational cost, have better robustness? We study this question for sparse Mixture of Expert models (MoEs), that make it possible to scale up the model size for a roughly constant computational cost. We theoretically show that under certain conditions on the routing and the structure of the data, MoEs can have significantly smaller Lipschitz constants than their dense counterparts. The robustness of MoEs can suffer when the highest weighted experts for an input implement sufficiently different functions. We next empirically evaluate the robustness of MoEs on ImageNet using adversarial attacks and show they are indeed more robust than dense models with the same computational cost. We make key observations showing the robustness of MoEs to the choice of experts, highlighting the redundancy of experts in models trained in practice.

Modern deep convolutional networks (CNNs) are often criticized for not generalizing under distributional shifts. However, several recent breakthroughs in transfer learning suggest that these networks can cope with severe distribution shifts and successfully adapt to new tasks from a few training examples. In this work we study the interplay between out-of-distribution and transfer performance of modern image classification CNNs for the first time and investigate the impact of the pre-training data size, the model scale, and the data preprocessing pipeline. We find that increasing both the training set and model sizes significantly improve the distributional shift robustness. Furthermore, we show that, perhaps surprisingly, simple changes in the preprocessing such as modifying the image resolution can significantly mitigate robustness issues in some cases. Finally, we outline the shortcomings of existing robustness evaluation datasets and introduce a synthetic dataset SI-SCORE we use for a systematic analysis across factors of variation common in visual data such as object size and position.

Sparsely-gated Mixture of Experts networks (MoEs) have demonstrated excellent scalability in Natural Language Processing. In Computer Vision, however, almost all performant networks are "dense", that is, every input is processed by every parameter. We present a Vision MoE (V-MoE), a sparse version of the Vision Transformer, that is scalable and competitive with the largest dense networks. When applied to image recognition, V-MoE matches the performance of state-of-the-art networks, while requiring as little as half of the compute at inference time. Further, we propose an extension to the routing algorithm that can prioritize subsets of each input across the entire batch, leading to adaptive per-image compute. This allows V-MoE to trade-off performance and compute smoothly at test-time. Finally, we demonstrate the potential of V-MoE to scale vision models, and train a 15B parameter model that attains 90.35% on ImageNet.

We endeavor to perform very large scale indexing of an ancient German collection of manuscript parish records. To this end we will compute "probabilistic indexes" (PIs), which are known to allow for very accurate and efficient implementation of (single-)keyword spotting. PIs may become prohibitively large for vast manuscript collections. Therefore we analyze simple index pruning methods to achieve adequate tradeoffs between memory requirements and search performance. We also study how to adequately deal with the large variety of non-ASCII symbols and handwritten word spelling variations (accents, umlauts, etc.) which appear in this kind of historical collections. Finally, and most importantly, since most of the images of the collection we aim to index are handwritten tables, we explore the use of PIs to support structured queries for information extraction from untranscribed handwritten images containing tabular data. Empirical results on a small, but complex and representative dataset extracted from the collection considered confirm the viability and adequateness of the chosen approaches.

Making large-scale collections of digitized historical documents searchable is being earnestly demanded by many archives and libraries. Probabilistically indexing the text images of these collections by means of keyword spotting techniques is currently seen as perhaps the only feasible approach to meet this demand. A vast medieval manuscript collection, written in both Latin and French, called "Chancery", is currently being considered for indexing at large. In addition to its bilingual nature, one of the major difficulties of this collection is the very high rate of abbreviated words which, on the other hand, are completely expanded in the ground truth transcripts available. In preparation to undertake full indexing of Chancery, experiments have been carried out on a relatively small but fully representative subset of this collection. To this end, a keyword spotting approach has been adopted which computes word relevance probabilities using character lattices produced by a recurrent neural network and a N-gram character language model. Results confirm the viability of the chosen approach for the large-scale indexing aimed at and show the ability of the proposed modeling and training approaches to properly deal with the abbreviation difficulties mentioned.

Current state-of-the-art approaches to offline Handwritten Text Recognition extensively rely on Multidimensional Long Short-Term Memory networks. However, these architectures come with quite an expensive computational cost, and we observe that they extract features visually similar to those of convolutional layers, which are computationally cheaper. This suggests that the two-dimensional long-term dependencies, which are potentially modeled by multidimensional recurrent layers, may not be essential to achieve a good recognition accuracy, at least in the lower layers of the architecture. In this work, an alternative model is explored that relies only on convolutional and one-dimensional recurrent layers that achieves better or equivalent results than those of the current state-of-the-art architecture, and runs significantly faster. In addition, we observe that using random distortions during training as synthetic data augmentation dramatically improves the accuracy of our model. Thus, are multidimensional recurrent layers really necessary for Handwritten Text Recognition? Probably not.

The H-KWS 2016, organized in the context of the ICFHR 2016 conference aims at setting up an evaluation framework for benchmarking handwritten keyword spotting (KWS) examining both the Query by Example (QbE) and the Query by String (QbS) approaches. Both KWS approaches were hosted into two different tracks, which in turn were split into two distinct challenges, namely, a segmentation-based and a segmentation-free to accommodate different perspectives adopted by researchers in the KWS field. In addition, the competition aims to evaluate the submitted training-based methods under different amounts of training data. Four participants submitted at least one solution to one of the challenges, according to the capabilities and/or restrictions of their systems. The data used in the competition consisted of historical German and English documents with their own characteristics and complexities. This paper presents the details of the competition, including the data, evaluation metrics and results of the best run of each participating methods.

Two methods are presented to improve word confidence scores for Line-Level Query-by-String Lexicon-Free Keyword Spotting (KWS) in handwritten text images. The first one approaches true relevance probabilities by means of computations directly carried out on character lattices obtained from the lines images considered. The second method uses the same character lattices, but it obtains relevance scores by first computing frame-level character sequence scores which resemble the word posteriorgrams used in previous approaches for lexicon-based KWS. The first method results from a formal probabilistic derivation, which allow us to better understand and further develop the underlying ideas. The second one is less formal but, according with experiments presented in the paper, it obtains almost identical results with much lower computational cost. Moreover, in contrast with the first method, the second one allows to directly obtain accurate bounding boxes for the spotted words.

Traditionally, the HMM-Filler approach has been widely used in the fields of speech recognition and handwritten text recognition to tackle lexicon-free, query-by-string keyword spotting (KWS). It computes a score to determine whether a given keyword is written in a certain image region. It is conjectured, that this score is related to the confidence of the system, respect to the previous question. However, it is still not clear what this relationship is. In this paper, the HMM-Filler score is derived from a probabilistic formulation of KWS, which gives a better understanding of its behavior and limits. Additionally, the same probabilistic framework is used to present a new algorithm to compute the KWS scores, which results in better average precision (AP), for a keyword spotting task in the widely used IAM database. We show that the new algorithm can improve the HMM-filler results up to 10.4% relative (5.3% absolute) points in AP, in the considered task.

The principal goal of the Competition on Keyword Spotting for Handwritten Documents was to promote different approaches used in the field of Keyword Spotting and to fairly compare them using uniform data and metrics. To accommodate different perspectives adopted by researches in this field, the competition was divided into two distinct tracks, namely, a training-free and a training-based track, and each track entailed two optional assignments. Six participants submitted solutions to one or both assignments, depending on the capabilities and/or restrictions of their systems. The data used in the competition consisted of historical documents in English with different levels of complexity. This paper presents the details of the competition, including the data, evaluation metrics and results of the best participant methods.

The so-called filler or garbage Hidden Markov Models (HMM-Filler) are among the most widely used models for lexicon-free, query by string key word spotting (KWS) in the fields of speech recognition and (lately) handwritten text recognition. However, it has important drawbacks. First, the keyword-specific HMM Viterbi decoding process needed to obtain the confidence scores of each spotted word involves a large computational cost. Second, in its traditional conception, the model does not take into account any context information - and more recent works where simple character bi-gram context is used show that not only the computational cost becomes even larger, but also the required keyword-specific language model becomes quite intricate to build. In a previous work we introduced KWS methods based on character lattices which proved very much simpler and faster than the traditional HMM-Filler, while providing practically identical results. Here we extend our previous work by using context-aware character lattices obtained by means of Viterbi decoding with high-order character N-gram models. Experimental results show that, as compared with a direct 2-gram HMM-filler implementation, the proposed approach requires between one and two orders of magnitude less query computing time. Moreover, for the first time in the field of handwritten text KWS, Filler-based results for N-grams up to N = 6 are reported, clearly showing a great impact of context on precision-recall performance.

Keyword Spotting (KWS) has been traditionally considered under two distinct frameworks: Query-by-Example (QbE) and Query-by-String (QbS). In both cases the user of the system wished to find occurrences of a particular keyword in a collection of document images. The difference is that, in QbE, the keyword is given as an exemplar image while, in QbS the keyword is given as a text string. In several works, the QbS scenario has been approached using QbE techniques; but the converse has not been studied in depth yet, despite of the fact that QbS systems typically achieve higher accuracy. In the present work, we present a very effective probabilistic approach to QbE KWS, based on highly accurate QbS KWS techniques which rely on models which need to be trained from annotated data. To assess the effectiveness of this approach, we tackle the segmentation-free QbE task of the ICFHR-2014 Competition on Handwritten KWS. Our approach achieves a mean average precision (mAP) as high as 0.715, which improves by more than 70% the best mAP achieved in this competition (0.419 under the same experimental conditions).

We present a handwritten text Keyword Spotting (KWS) approach based on the combination of KWS methods using word-graphs (WGs) and character-lattices (CLs). It aims to solve the problem that WG-based models present for out of vocabulary (OOV) keywords: since there is no available information about them in the lexicon or the language model, null scores are assigned. OOV keywords may have a significant impact on the global performance of KWS systems, as we show. By using a CL approach, which does not suffer from the previous problem, to estimate the OOV scores, we take advantage of both models, using the speed and accuracy that WGs provide for in-vocabulary keywords and the flexibility of the CL approach. This combination improves significantly both average precision and mean average precision over the two methods.

Thanks to the use of lexical and syntactic information, Word Graphs (WG) have shown to provide a competitive Precision-Recall performance, along with fast lookup times, in comparison to other techniques used for Key-Word Spotting (KWS) in handwritten text images. However, a problem of WG approaches is that they assign a null score to any keyword that was not part of the training data, i.e. Out-of-Vocabulary (OOV) keywords, whereas other techniques are able to estimate a reasonable score even for these kind of keywords. We present a smoothing technique which estimates the score of an OOV keyword based on the scores of similar keywords. This makes the WG-based KWS as flexible as other techniques with the benefit of having much faster lookup times.