Image copy detection is the task of detecting edited copies of any image within a reference database. While previous approaches have shown remarkable progress, the large size of their networks and descriptors remains a disadvantage, complicating their practical application. In this paper, we propose a novel method that achieves competitive performance by using a lightweight network and compact descriptors. By utilizing relational self-supervised distillation to transfer knowledge from a large network to a small network, we enable the training of lightweight networks with smaller descriptor sizes. We introduce relational self-supervised distillation for flexible representation in a smaller feature space and apply contrastive learning with a hard negative loss to prevent dimensional collapse. For the DISC2021 benchmark, ResNet-50 and EfficientNet-B0 are used as the teacher and student models, respectively, with micro average precision improving by$5.0 \% / 4.9 \% / 5.9 \%$for$64 / 128 / 256$descriptor sizes compared to the baseline method. The code is available at https://github.com/juntae9926/RDCD.

For automatic speaker verification systems, spoof speech detection (SSD) is an essential countermeasure. Although SSD with magnitude features in the frequency domain has shown promising results, phase information can also be useful in capturing the artefacts of certain spoofing attacks. Thus, both magnitude and phase features must be considered to ensure the ability to generalize diverse types of spoofing attacks. In this study, we discovered that the randomness difference between magnitude and phase features is large, which can interrupt the feature-level fusion via backend neural network. In this regard, we propose a phase network to reduce that difference, which makes the Res2Net-based feature-level fusion feasible. To validate our SSD system for practical environment, both known- and unknown-type SSD scenarios are considered. As a result, our SSD system delivers competitive results compared to other state-of-the-art SSD systems in all scenarios.

The liquidity of the exchange is the most important factor when crytocurrency traders choose an exchange. However, the amount of liquidity provided by the liquidity providers in decentralized exchanges is insufficient when compared to centralized exchanges. This is because the liquidity providers in decentralized exchanges suffer from the risk of divergence loss inherent to the automated market making system. To this end, we introduce a new concept called margin liquidity and leverage this concept to propose a highly profitable margin liquidity-providing position. Then, we extend this margin liquidity-providing position to a virtual margin liquidity-providing position to alleviate the risk of divergence loss for the liquidity providers and encourage them to provide more liquidity to the pool. We show that our proposed margin liquidity is 8K times more capital efficient than the concentrated liquidity proposed in Uniswap V3.

Existing deep interactive colorization models have focused on ways to utilize various types of interactions, such as point-wise color hints, scribbles, or natural-language texts, as methods to reflect a user’s intent at runtime. However, another approach, which actively informs the user of the most effective regions to give hints for sketch image colorization, has been under-explored. This paper proposes a novel model-guided deep interactive colorization framework that reduces the required amount of user interactions, by prioritizing the regions in a colorization model. Our method, called GuidingPainter, prioritizes these regions where the model most needs a color hint, rather than just relying on the user’s manual decision on where to give a color hint. In our extensive experiments, we show that our approach outperforms existing interactive colorization methods in terms of the conventional metrics, such as PSNR and FID, and reduces required amount of interactions.

Recurrent neural network transducer (RNN-T) is an endto-end speech recognition framework converting input acoustic frames into a character sequence.The state-of-the-art encoder network for RNN-T is the Conformer, which can effectively model the local-global context information via its convolution and self-attention layers.Although Conformer RNN-T has shown outstanding performance, most studies have been verified in the setting where the train and test data are drawn from the same domain.The domain mismatch problem for Conformer RNN-T has not been intensively investigated yet, which is an important issue for the product-level speech recognition system.In this study, we identified that fully connected self-attention layers in the Conformer caused high deletion errors, specifically in the long-form out-domain utterances.To address this problem, we introduce sparse self-attention layers for Conformer-based encoder networks, which can exploit local and generalized global information by pruning most of the indomain fitted global connections.Also, we propose a state reset method for the generalization of the prediction network to cope with long-form utterances.Applying proposed methods to an out-domain test, we obtained 27.6% relative character error rate (CER) reduction compared to the fully connected self-attention layer-based Conformers.

When expressing personal data on the displays of personal IoT devices, it is important to be intuitively aware of privacy settings and perform ready-to-hand interactions to respond appropriately to various situations occurring in shared spaces. In this paper, we developed Posit, an interactive calendar in which the disclosure level of schedule content can be changed in three stages according to the object's placement by the user. The results of our three-week in-field study with six participants revealed that Posit's interaction was considered to be a simple way of hiding personal schedules quickly, and we could identify the roles of the positional messages in determining the others’ gazes on displays. Additionally, we confirmed that social relationships and trust between colleagues affect the use of Posit. Our findings imply new opportunities in designing interactions for the management of personal privacy by applying physical state-changing interaction and understanding social factors in shared spaces.

Most neural vocoders employ band-limited mel-spectrograms to generate waveforms.If full-band spectral features are used as the input, the vocoder can be provided with as much acoustic information as possible.However, in some models employing full-band mel-spectrograms, an over-smoothing problem occurs as part of which non-sharp spectrograms are generated.To address this problem, we propose UnivNet, a neural vocoder that synthesizes high-fidelity waveforms in real time.Inspired by works in the field of voice activity detection, we added a multiresolution spectrogram discriminator that employs multiple linear spectrogram magnitudes computed using various parameter sets.Using full-band mel-spectrograms as input, we expect to generate high-resolution signals by adding a discriminator that employs spectrograms of multiple resolutions as the input.In an evaluation on a dataset containing information on hundreds of speakers, UnivNet obtained the best objective and subjective results among competing models for both seen and unseen speakers.These results, including the best subjective score for text-to-speech, demonstrate the potential for fast adaptation to new speakers without a need for training from scratch.

Abstract Adopting a re-materialization approach, we designed and implemented an everyday interactive artifact that enables an individual to monitor and establish the reconfirmation time of email data. This device represents a new means of handling and interacting with email. To investigate the value of the materialization of email data through a daily interactive object, Maili, we conducted a 1-month field study with five participants in their work environments. The results showed that applying physicality to email handling helped to increase accessibility to and interest in the data, as well as in the reconfirmation function of email. Results also indicated the value of combining non-digital (the tray) with digital functions. By presenting the process of using the dematerialized data, our findings offer new insights into how we can materialize digital information in everyday tangible artifacts.

End-to-end automatic speech recognition (E2E-ASR) prefers the common words during training rather than rare ones related to contextual information such as song names. Thus, recognizing contextual information correctly is a hurdle for E2E-ASR to reach the production-level. To overcome the limitations of E2E-ASR in recognizing contextual information, this work presents a post-processing followed by E2E-ASR in an algorithmic way, referred to as a word-matching algorithm with backward search (WMA-BS). At first, we allow E2E-ASR to roughly detect the position of target words that has similar pronunciation with desired contextual phrases. After that, given the hypothesis from E2E-ASR with the rough position of target words, WMA-BS estimates the correct target words and decides whether to replace the target words with the contextual phrase or not, according to their phonetic and literal similarity. Applying the proposed method to E2E-ASR achieved relative improvement up to 52.7% in word error rate across several harsh conditions.

Abstract Currently, large data streams are constantly being generated in diverse environments, and continuous storage of the data and periodic batch‐type principal component analysis (PCA) are becoming increasingly difficult. Various online PCA algorithms have been proposed to solve this problem. In this study, we propose an online PCA methodology based on online eigenvector transformation with the moving average of the data stream that can reflect concept drift. We compared the network intrusion detection performance based on online transformation of eigenvectors with that of offline methods by applying three machine learning algorithms. Both online and offline methods demonstrated excellent performance in terms of precision. However, in terms of the recall ratio, the performance of the proposed methodology with integrated online eigenvector transformation was better; thus, the F1‐measure also indicated better performance. The visualization of the principal component score shows the effectiveness of our method.

Recurrent neural network transducers (RNN-T) are a promising end-to-end speech recognition framework that transduce input acoustic frames to a character sequence. Best- and breadth-first searches have been used as decoding strategies for RNN-T. However, best-first search follows a sequential process for its expansion search, which slows down the decoding process. Although breadth-first search replaces the sequential process of best-first search with a parallel one, it unnecessarily conducts an expansion search for all decoding steps. As most of the decoding frames correspond to a blank symbol because the length of the character sequence is much shorter than that of the decoding frames, this induces computational overhead. To address these limitations, we introduce an adaptive expansion search (AES) to accelerate RNN-T inference. AES overcomes the aforementioned limitations by batching the hypotheses and adopting a decision-making process that decides whether to continue the expansion search; thus, AES can avoid unnecessary expansion search. Furthermore, pruning is applied to AES for further acceleration. We achieved significant speedup and a lower word error rate compared with other baselines.

A novel neural network architecture, called two-stage network (TSN), with a multi-objective learning (MOL) method for an efficient boosting strategy (BS) is proposed for speech enhancement. BS is an ensemble method using multiple base predictions (MBPs) for better final prediction. Because of the necessity for MBPs, the computational cost and model size of BS-based methods are greater than those of a single model. In overcoming this, TSN first obtains MBPs from a single deep neural network. Then, to obtain better final prediction, the convolution layers of TSN aggregate not only MBPs but also some auxiliary information such as contextual information, while adaptively filtering out some unnecessary information, e.g., poor base predictions. At the training phase, the MOL enables all stages of TSN to learn jointly, whereas allowing the TSN framework to embed a BS. Our experimental results confirm that the embedded BS leads TSN to outperform other baseline methods with a reasonably low computational cost and model size.

We introduce Traffico, a tangible timetable representing dematerialized schedule and transportation information. It delivers a user's schedules in chronological order along with transportation information between schedules. Placed on the user's desk, Traffico suggests required transportation times using four options-walking, bicycling, bussing, and driving a car-and through an e-ink display. To investigate the advantages that Traffico provides to users, we conducted an in-field study of 10 participants over five days. The results revealed that Traffico supports the planning of moving times in a day through displaying transportation options on each schedule. We also found that Traffico provides better schedule reminders with event notifications in a sequential order, along with rotating interaction for checking events. Through this type of tangible interaction, Traffico provides possibilities to reflect dematerialized digital information into a physical form and to adopt a new way of scheduling and handling time.

Voice activity detection (VAD) classifies incoming signal segments into speech or background noise; its performance is crucial in various speech-related applications. Although speech-signal context is a relevant VAD asset, its usefulness varies in unpredictable noise environments. Therefore, its usage should be adaptively adjustable to the noise type. This letter improves the use of context information by using an adaptive context attention model (ACAM) with a novel training strategy for effective attention, which weights the most crucial parts of the context for proper classification. Experiments in real-world scenarios demonstrate that the proposed ACAM-based VAD outperforms the other baseline VAD methods.

Recommendation systems analyze user preferences and recommend items to a user by predicting the user's preference for those items. Among various kinds of recommendation methods, collaborative filtering (CF) has been widely used and successfully applied to practical applications. However, collaborative filtering has two inherent problems: data sparseness and the cold-start problems. In this paper, we propose a method of integrating additional feature information of users and items into CF to overcome the difficulties caused by sparseness and improve the accuracy of recommendation. Several experimental results that show the effectiveness of the proposed method are also presented.

This paper presents the SERF (System for Electronic Recommendation Filtering) which is a collaborative filtering system that recommends context-sensitive, high-quality information sources for document search. Collaborative filtering systems remove the limitation of traditional content-based search by using individual's ratings to evaluate and recommend information sources. SERF uses collaborative filtering algorithms to predict the relevance and quality of each document with respect to each particular user and their specific information need. In our system, users specify their need in the form of a natural language query, and are provided with recommended documents based on ratings by other users with similar questions. Preliminary experiments show that the collaborative filtering recommendations increase the efficiency of the document search process. We also discuss some key challenges of designing a collaborative filtering system for document search.

Recommendation systems predict user's preference to suggest items. Collaborative filtering is the most popular method in implementing a recommendation system. The collaborative filtering method computes similarities between users based on each user's known preference, and recommends the items preferred by similar users. Although the collaborative filtering method generally shows good performance, it suffers from two major problems - data sparseness and scalability. We present a model-based recommendation algorithm that uses multilevel association rules to alleviate those problems. In this algorithm, we build a model for preference prediction by using association rule mining. Multilevel association rules are used to compute preferences for items. The experimental results show that applying multilevel association rules is effective, and performance of the algorithm is improved compared with the collaborative filtering method in terms of the recall and the computation time.