In this work, we investigate a method for simulation-free training of Neural Ordinary Differential Equations (NODEs) for learning deterministic mappings between paired data. Despite the analogy of NODEs as continuous-depth residual networks, their application in typical supervised learning tasks has not been popular, mainly due to the large number of function evaluations required by ODE solvers and numerical instability in gradient estimation. To alleviate this problem, we employ the flow matching framework for simulation-free training of NODEs, which directly regresses the parameterized dynamics function to a predefined target velocity field. Contrary to generative tasks, however, we show that applying flow matching directly between paired data can often lead to an ill-defined flow that breaks the coupling of the data pairs (e.g., due to crossing trajectories). We propose a simple extension that applies flow matching in the embedding space of data pairs, where the embeddings are learned jointly with the dynamic function to ensure the validity of the flow which is also easier to learn. We demonstrate the effectiveness of our method on both regression and classification tasks, where our method outperforms existing NODEs with a significantly lower number of function evaluations. The code is available at https://github.com/seminkim/simulation-free-node.

We propose a text-to-image generation algorithm based on deep neural networks when text captions for images are unavailable during training. In this work, instead of simply generating pseudo-ground-truth sentences of training images using existing image captioning methods, we employ a pretrained CLIP model, which is capable of properly aligning embeddings of images and corresponding texts in a joint space and, consequently, works well on zero-shot recognition tasks. We optimize a text-to-image generation model by maximizing the data log-likelihood conditioned on pairs of image-text CLIP embeddings. To better align data in the two domains, we employ a principled way based on a variational inference, which efficiently estimates an approximate posterior of the hidden text embedding given an image and its CLIP feature. Experimental results validate that the proposed framework outperforms existing approaches by large margins under unsupervised and semi-supervised text-to-image generation settings.

Despite recent advances in implicit neural representations (INRs), it remains challenging for a coordinate-based multi-layer perceptron (MLP) of INRs to learn a common representation across data instances and generalize it for unseen instances. In this work, we introduce a simple yet effective framework for generalizable INRs that enables a coordinate-based MLP to represent complex data instances by modulating only a small set of weights in an early MLP layer as an instance pattern composer; the remaining MLP weights learn pattern composition rules for common representations across instances. Our generalizable INR frame-work is fully compatible with existing meta-learning and hypernetworks in learning to predict the modulated weight for unseen instances. Extensive experiments demonstrate that our method achieves high performance on a wide range of domains such as an audio, image, and 3D object, while the ablation study validates our weight modulation.

Learning visual and textual representations in the shared space from web-scale image-text pairs improves the performance of diverse vision-and-language tasks, as well as modality-specific tasks. Many attempts in this framework have been made to connect English-only texts and images, and only a few works have been proposed to extend this framework in multilingual settings with the help of many translation pairs. In this multilingual approach, a typical setup is to use pairs of (image and English-text) and translation pairs. The major limitation of this approach is that the learning signal of aligning visual representation with under-resourced language representation is not strong, achieving a sub-optimal performance of vision-and-language tasks. In this work, we propose a simple yet effective enhancement scheme for previous multilingual multi-modal representation methods by using a limited number of pairs of images and non-English texts. In specific, our scheme fine-tunes a pre-trained multilingual model by minimizing a triplet contrastive loss on triplets of image and two different language texts with the same meaning, improving the connection between images and non-English texts. Experiments confirm that our enhancement strategy achieves performance gains in image-text retrieval, zero-shot image classification, and sentence embedding tasks.

For autoregressive (AR) modeling of high-resolution images, vector quantization (VQ) represents an image as a sequence of discrete codes. A short sequence length is important for an AR model to reduce its computational costs to consider long-range interactions of codes. However, we postulate that previous VQ cannot shorten the code sequence and generate high-fidelity images together in terms of the rate-distortion trade-off. In this study, we propose the two-stage framework, which consists of Residual-Quantized VAE (RQ-VAE) and RQ-Transformer, to effectively generate high-resolution images. Given a fixed codebook size, RQ-VAE can precisely approximate a feature map of an image and represent the image as a stacked map of discrete codes. Then, RQ-Transformer learns to predict the quantized feature vector at the next position by predicting the next stack of codes. Thanks to the precise approximation of RQ-VAE, we can represent a$256\times 256$image as$8\times 8$resolution of the feature map, and RQ-Transformer can efficiently reduce the computational costs. Consequently, our framework out-performs the existing AR models on various benchmarks of unconditional and conditional image generation. Our approach also has a significantly faster sampling speed than previous AR models to generate high-quality images.

Although autoregressive models have achieved promising results on image generation, their unidirectional generation process prevents the resultant images from fully reflecting global contexts. To address the issue, we propose an effective image generation framework of \emph{Draft-and-Revise} with \emph{Contextual RQ-transformer} to consider global contexts during the generation process. As a generalized VQ-VAE, RQ-VAE first represents a high-resolution image as a sequence of discrete code stacks. After code stacks in the sequence are randomly masked, Contextual RQ-Transformer is trained to infill the masked code stacks based on the unmasked contexts of the image. Then, we propose the two-phase decoding, Draft-and-Revise, for Contextual RQ-Transformer to generates an image, while fully exploiting the global contexts of the image during the generation process. Specifically. in the \emph{draft} phase, our model first focuses on generating diverse images despite rather low quality. Then, in the \emph{revise} phase, the model iteratively improves the quality of images, while preserving the global contexts of generated images. In experiments, our method achieves state-of-the-art results on conditional image generation. We also validate that the Draft-and-Revise decoding can achieve high performance by effectively controlling the quality-diversity trade-off in image generation.

We propose a locally hierarchical auto-regressive model with multiple resolutions of discrete codes. In the first stage of our algorithm, we represent an image with a pyramid of codes using Hierarchically Quantized Variational AutoEncoder (HQ-VAE), which disentangles the information contained in the multi-level codes. For an example of two-level codes, we create two separate pathways to carry high-level coarse structures of input images using top codes while compensating for missing fine details by constructing a residual connection for bottom codes. An appropriate selection of resizing operations for code embedding maps enables top codes to capture maximal information within images and the first stage algorithm achieves better performance on both vector quantization and image generation. The second stage adopts Hierarchically Quantized Transformer (HQ-Transformer) to process a sequence of local pyramids, which consist of a single top code and its corresponding bottom codes. Contrary to other hierarchical models, we sample bottom codes in parallel by exploiting the conditional independence assumption on the bottom codes. This assumption is naturally harvested from our first-stage model, HQ-VAE, where the bottom code learns to describe local details. On class-conditional and text-conditional generation benchmarks, our model shows competitive performance to previous AR models in terms of fidelity of generated images while enjoying lighter computational budgets.

We present a personalized and reliable prediction model for healthcare, which can provide individually tailored medical services such as diagnosis, disease treatment, and prevention. Our proposed framework targets at making personalized and reliable predictions from time-series data, such as Electronic Health Records (EHR), by modeling two complementary components: i) a shared component that captures global trend across diverse patients and ii) a patient-specific component that models idiosyncratic variability for each patient. To this end, we propose a composite model of a deep neural network to learn complex global trends from the large number of patients, and Gaussian Processes (GP) to probabilistically model individual time-series given relatively small number of visits per patient. We evaluate our model on diverse and heterogeneous tasks from EHR datasets and show practical advantages over standard time-series deep models such as pure Recurrent Neural Network (RNN).

This paper proposes a simple nearest neighbor search algorithm, which provides the exact solution in terms of the Euclidean distance efficiently. Especially, we present an interesting approach to improve the speed of nearest neighbor search by proper translations of data and query although the task is inherently invariant to the Euclidean transformations. The proposed algorithm aims to eliminate nearest neighbor candidates effectively using their distance lower bounds in nonlinear embedded spaces, and further improves the lower bounds by transforming data and query through product quantized translations. Although our framework is composed of simple operations only, it achieves the state-of-the-art performance compared to existing nearest neighbor search techniques, which is illustrated quantitatively using various large-scale benchmark datasets in different sizes and dimensions.

Attention mechanism is effective in both focusing the deep learning models on relevant features and interpreting them. However, attentions may be unreliable since the networks that generate them are often trained in a weakly-supervised manner. To overcome this limitation, we introduce the notion of input-dependent uncertainty to the attention mechanism, such that it generates attention for each feature with varying degrees of noise based on the given input, to learn larger variance on instances it is uncertain about. We learn this Uncertainty-aware Attention (UA) mechanism using variational inference, and validate it on various risk prediction tasks from electronic health records on which our model significantly outperforms existing attention models. The analysis of the learned attentions shows that our model generates attentions that comply with clinicians' interpretation, and provide richer interpretation via learned variance. Further evaluation of both the accuracy of the uncertainty calibration and the prediction performance with "I don't know'' decision show that UA yields networks with high reliability as well.

We propose DropMax, a stochastic version of softmax classifier which at each iteration drops non-target classes according to dropout probabilities adaptively decided for each instance. Specifically, we overlay binary masking variables over class output probabilities, which are input-adaptively learned via variational inference. This stochastic regularization has an effect of building an ensemble classifier out of exponentially many classifiers with different decision boundaries. Moreover, the learning of dropout rates for non-target classes on each instance allows the classifier to focus more on classification against the most confusing classes. We validate our model on multiple public datasets for classification, on which it obtains significantly improved accuracy over the regular softmax classifier and other baselines. Further analysis of the learned dropout probabilities shows that our model indeed selects confusing classes more often when it performs classification.

Binary embedding refers to methods for embedding points in ℝdinto vertices in the Hamming cube of dimension O(d), such that the normalized Hamming distance between the codes preserves a prespecified distance between vectors in the original space. A common approach to binary embedding is to use random projection, followed by one-bit quantization to produce binary codes. Of particular interest, in this letter, is sparse circulant binary embedding (SCBE), where a sparse random circulant matrix with random sampling at a rate of 1 - s for s E (0, 1) is used for random projection. The SCBE has the space complexity O((1 - s)d), while unstructured random projection has the space complexity O(d2). We present an asymptotic analysis of SCBE, when d approaches ∞, showing that the performance of SCBE is comparable to that of binary embedding with unstructured random projection whilethe former has the space complexity O((1 - s)d) and the time complexity O(d log d) but the latter has both space and time complexities O(d2).

A commercial web search engine shards its index among many servers, and therefore the response time of a search query is dominated by the slowest server that processes the query. Prior approaches target improving responsiveness by reducing the tail latency , or high-percentile response time, of an individual search server. They predict query execution time, and if a query is predicted to be long-running, it runs in parallel; otherwise, it runs sequentially. These approaches are, however, not accurate enough for reducing a high tail latency when responses are aggregated from many servers because this requires each server to reduce a substantially higher tail latency (e.g., the 99.99th percentile), which we call extreme tail latency. To address tighter requirements of extreme tail latency, we propose a new design space for the problem, subsuming existing work and also proposing a new solution space. Existing work makes a prediction using features available at indexing time and focuses on optimizing prediction features for accelerating tail queries. In contrast, we identify “when to predict?” as another key optimization question. This opens up a new solution of delaying a prediction by a short duration to allow many short-running queries to complete without parallelization and, at the same time, to allow the predictor to collect a set of dynamic features using runtime information. This new question expands a solution space in two meaningful ways. First, we see a significant reduction of tail latency by leveraging “dynamic” features collected at runtime that estimate query execution time with higher accuracy. Second, we can ask whether to override prediction when the “predictability” is low. We show that considering predictability accelerates the query by achieving a higher recall. With this prediction, we propose to accelerate the queries that are predicted to be long-running. In our preliminary work, we focused on parallelization as an acceleration scenario. We extend to consider heterogeneous multicore hardware for acceleration. This hardware combines processor cores with different microarchitectures such as energy-efficient little cores and high-performance big cores, and accelerating web search using this hardware has remained an open problem. We evaluate the proposed prediction framework in two scenarios: (1) query parallelization on a multicore processor and (2) query scheduling on a heterogeneous processor. Our extensive evaluation results show that, for both scenarios of query acceleration using parallelization and heterogeneous cores, the proposed framework is effective in reducing the extreme tail latency compared to a start-of-the-art predictor because of its higher recall, and it improves server throughput by more than 70% because of its improved precision.

Locality-sensitive hashing (LSH) is a popular data-independent indexing method for approximate similarity search, where random projections followed by quantization hash the points from the database so as to ensure that the probability of collision is much higher for objects that are close to each other than for those that are far apart. Most of high-dimensional visual descriptors for images exhibit a natural matrix structure. When visual descriptors are represented by high-dimensional feature vectors and long binary codes are assigned, a random projection matrix requires expensive complexities in both space and time. In this paper we analyze a bilinear random projection method where feature matrices are transformed to binary codes by two smaller random projection matrices. We base our theoretical analysis on extending Raginsky and Lazebnik's result where random Fourier features are composed with random binary quantizers to form locality sensitive binary codes. To this end, we answer the following two questions: (1) whether a bilinear random projection also yields similarity-preserving binary codes; (2) whether a bilinear random projection yields performance gain or loss, compared to a large linear projection. Regarding the first question, we present upper and lower bounds on the expected Hamming distance between binary codes produced by bilinear random projections. In regards to the second question, we analyze the upper and lower bounds on covariance between two bits of binary codes, showing that the correlation between two bits is small. Numerical experiments on MNIST and Flickr45K datasets confirm the validity of our method.

Near-duplicate image discovery is the task of detecting all clusters of images which duplicate at a significant region. Previous work generally take divide and conquer approaches composed of two steps: generating cluster seeds using min-hashing, and growing the seeds by searching the entire image space with the seeds as queries. Since the computational complexity of the seed growing step is generally O (NL) where N and L are the number of images and seeds respectively, existing work can hardly be scaled up to a billion-scale dataset because L is typically millions. In this paper, we study a feasible solution of near-duplicate image discovery on one billion images, which is easily implemented on MapReduce framework. The major contribution of this work is to introduce the seed growing step designed to efficiently reduce the number of false positives among cluster seeds with O (cNL) time complexity, where c is small enough for a billion-scale dataset. The basis component of the seed growing step is a bottom-k min-hash, which generates different signatures in a sketch to remove all candidate images that share only one common visual word with a cluster seed. Our evaluations suggest that the proposed method can discover near-duplicate clusters with high precision and recall, and represent some interesting properties of our 1 billion dataset.

A commercial web search engine shards its index among many servers, and therefore the response time of a search query is dominated by the slowest server that processes the query. Prior approaches target improving responsiveness by reducing the tail latency of an individual search server. They predict query execution time, and if a query is predicted to be long-running, it runs in parallel, otherwise it runs sequentially. These approaches are, however, not accurate enough for reducing a high tail latency when responses are aggregated from many servers because this requires each server to reduce a substantially higher tail latency (e.g., the 99.99th-percentile), which we call extreme tail latency.

Web search engines are optimized to reduce the high-percentile response time to consistently provide fast responses to almost all user queries. This is a challenging task because the query workload exhibits large variability, consisting of many short-running queries and a few long-running queries that significantly impact the high-percentile response time. With modern multicore servers, parallelizing the processing of an individual query is a promising solution to reduce query execution time, but it gives limited benefits compared to sequential execution since most queries see little or no speedup when parallelized. The root of this problem is that short-running queries, which dominate the workload, do not benefit from parallelization. They incur a large parallelization overhead, taking scarce resources from long-running queries. On the other hand, parallelization substantially reduces the execution time of long-running queries with low overhead and high parallelization efficiency. Motivated by these observations, we propose a predictive parallelization framework with two parts: (1) predicting long-running queries, and (2) selectively parallelizing them. For the first part, prediction should be accurate and efficient. For accuracy, we study a comprehensive feature set covering both term features (reflecting dynamic pruning efficiency) and query features (reflecting query complexity). For efficiency, to keep overhead low, we avoid expensive features that have excessive requirements such as large memory footprints. For the second part, we use the predicted query execution time to parallelize long-running queries and process short-running queries sequentially. We implement and evaluate the predictive parallelization framework in Microsoft Bing search. Our measurements show that under moderate to heavy load, the predictive strategy reduces the 99th-percentile response time by 50% (from 200 ms to 100 ms) compared with prior approaches that parallelize all queries.

Multi-view hashing seeks compact integrated binary codes which preserve similarities averaged over multiple representations of objects. Most of existing multi-view hashing methods resort to linear hash functions where data manifold is not considered. In this paper we present multi-view anchor graph hashing (MVAGH), where nonlinear integrated binary codes are efficiently determined by a subset of eigenvectors of an averaged similarity matrix. The efficiency behind MVAGH is due to a low-rank form of the averaged similarity matrix induced by multi-view anchor graph, where the similarity between two points is measured by two-step transition probability through view-specific anchor (i.e. landmark) points. In addition, we observe that MVAGH suffers from the performance degradation when the high recall is required. To overcome this drawback, we propose a simple heuristic to combine MVAGH with locality sensitive hashing (LSH). Numerical experiments on CIFAR-10 dataset confirms that MVAGH(+LSH) outperforms the existing multi- and single-view hashing methods.

Hashing seeks an embedding of high-dimensional objects into a similarity-preserving low-dimensional Hamming space such that similar objects are indexed by binary codes with small Hamming distances. A variety of hashing methods have been developed, but most of them resort to a single view (representation) of data. However, objects are often described by multiple representations. For instance, images are described by a few different visual descriptors (such as SIFT, GIST, and HOG), so it is desirable to incorporate multiple representations into hashing, leading to multi-view hashing. In this paper we present a deep network for multi-view hashing, referred to as deep multi-view hashing, where each layer of hidden nodes is composed of view-specific and shared hidden nodes, in order to learn individual and shared hidden spaces from multiple views of data. Numerical experiments on image datasets demonstrate the useful behavior of our deep multi-view hashing (DMVH), compared to recently-proposed multi-modal deep network as well as existing shallow models of hashing.

Hashing, which involves learning binary codes to embed high-dimensional data into a similarity-preserving low-dimensional Hamming space, is often formulated as linear dimensionality reduction followed by binary quantization. Linear dimensionality reduction, based on maximum variance formulation, requires leading eigenvectors of data covariance or graph Laplacian matrix. Computing leading singular vectors or eigenvectors in the case of high-dimension and large sample size, is a main bottleneck in most of data-driven hashing methods. In this paper we address the use of generalized Nystrom method where a subset of rows and columns are used to approximately compute leading singular vectors of the data matrix, in order to improve the scalability of hashing methods in the case of high-dimensional data with large sample size. Especially we validate the useful behavior of generalized Nystrom approximation with uniform sampling, in the case of a recently-developed hashing method based on principal component analysis (PCA) followed by an iterative quantization, referred to as PCA+ITQ, developed by Gong and Lazebnik. We compare the performance of generalized Nystrom approximation with uniform and non-uniform sampling, to the full singular value decomposition (SVD) method, confirming that the uniform sampling improves the computational and space complexities dramatically, while the performance is not much sacrificed. In addition we present low-rank approximation error bounds for generalized Nystrom approximation with uniform sampling, which is not a trivial extension of available results on the non-uniform sampling case.

Hashing refers to methods for embedding high dimensional data into a similarity-preserving low-dimensional Hamming space such that similar objects are indexed by binary codes whose Hamming distances are small. Learning hash functions from data has recently been recognized as a promising approach to approximate nearest neighbor search for high dimensional data. Most of 'learning to hash' methods resort to either unsupervised or supervised learning to determine hash functions. Recently semi-supervised learning approach was introduced in hashing where pair wise constraints (must link and cannot-link) using labeled data are leveraged while unlabeled data are used for regularization to avoid over-fitting. In this paper we base our semi-supervised hashing on linear discriminant analysis, where hash functions are learned such that labeled data are used to maximize the separability between binary codes associated with different classes while unlabeled data are used for regularization as well as for balancing condition and pair wise decor relation of bits. The resulting method is referred to as semi-supervised discriminant hashing (SSDH). Numerical experiments on MNIST and CIFAR-10 datasets demonstrate that our method outperforms existing methods, especially in the case of short binary codes.