Thanh Tran

Amazon AGI, Aaron Langford, Aayush Shah et al. · amazon-science

We present Amazon Nova, a new generation of state-of-the-art foundation models that deliver frontier intelligence and industry-leading price performance. Amazon Nova Pro is a highly-capable multimodal model with the best combination of accuracy, speed, and cost for a wide range of tasks. Amazon Nova Lite is a low-cost multimodal model that is lightning fast for processing images, video, documents and text. Amazon Nova Micro is a text-only model that delivers our lowest-latency responses at very low cost. Amazon Nova Canvas is an image generation model that creates professional grade images with rich customization controls. Amazon Nova Reel is a video generation model offering high-quality outputs, customization, and motion control. Our models were built responsibly and with a commitment to customer trust, security, and reliability. We report benchmarking results for core capabilities, agentic performance, long context, functional adaptation, runtime performance, and human evaluation.

Kai Wei, Dillon Knox, Martin Radfar et al.

Dialogue act classification (DAC) is a critical task for spoken language understanding in dialogue systems. Prosodic features such as energy and pitch have been shown to be useful for DAC. Despite their importance, little research has explored neural approaches to integrate prosodic features into end-to-end (E2E) DAC models which infer dialogue acts directly from audio signals. In this work, we propose an E2E neural architecture that takes into account the need for characterizing prosodic phenomena co-occurring at different levels inside an utterance. A novel part of this architecture is a learnable gating mechanism that assesses the importance of prosodic features and selectively retains core information necessary for E2E DAC. Our proposed model improves DAC accuracy by 1.07% absolute across three publicly available benchmark datasets.

Norbert Heuer, Thanh Tran

We present a simple discretization by radial basis functions for the Poisson equation with Dirichlet boundary condition. A Lagrangian multiplier using piecewise polynomials is used to accommodate the boundary condition. This simplifies previous attempts to use radial basis functions in the interior domain to approximate the solution and on the boundary to approximate the multiplier, which technically requires that the mesh norm in the interior domain is significantly smaller than that on the boundary. Numerical experiments confirm theoretical results.

Norbert Heuer, Thanh Tran

We analyze the approximation by radial basis functions of a hypersingular integral equation on an open surface. In order to accommodate the homogeneous essential boundary condition along the surface boundary, scaled radial basis functions on an extended surface and Lagrangian multipliers on the extension are used. We prove that our method converges quasi-optimally. Approximation results for scaled radial basis functions indicate that, for highly regular radial basis functions, the achieved convergence rates are close to the one of low-order conforming boundary element schemes. Numerical experiments confirm our conclusions.

Thanh Tran, Sebastian Bader, Jan Lundgren

Denoising is the process of removing noise from sound signals while improving the quality and adequacy of the sound signals. Denoising sound has many applications in speech processing, sound events classification, and machine failure detection systems. This paper describes a method for creating an autoencoder to map noisy machine sounds to clean sounds for denoising purposes. There are several types of noise in sounds, for example, environmental noise and generated frequency-dependent noise from signal processing methods. Noise generated by environmental activities is environmental noise. In the factory, environmental noise can be created by vehicles, drilling, people working or talking in the survey area, wind, and flowing water. Those noises appear as spikes in the sound record. In the scope of this paper, we demonstrate the removal of generated noise with Gaussian distribution and the environmental noise with a specific example of the water sink faucet noise from the induction motor sounds. The proposed method was trained and verified on 49 normal function sounds and 197 horizontal misalignment fault sounds from the Machinery Fault Database (MAFAULDA). The mean square error (MSE) was used as the assessment criteria to evaluate the similarity between denoised sounds using the proposed autoencoder and the original sounds in the test set. The MSE is below or equal to 0.14 when denoise both types of noises on 15 testing sounds of the normal function category. The MSE is below or equal to 0.15 when denoising 60 testing sounds on the horizontal misalignment fault category. The low MSE shows that both the generated Gaussian noise and the environmental noise were almost removed from the original sounds with the proposed trained autoencoder.

Thanh Tran, Sebastian Bader, Jan Lundgren

In an effort to advocate the research for a deep learning-based machine failure detection system, we present a case study of our proposed system based on a tiny sound dataset. Our case study investigates a variational autoencoder (VAE) for augmenting a small drill sound dataset from Valmet AB. A Valmet dataset contains 134 sounds that have been divided into two categories: "Anomaly" and "Normal" recorded from a drilling machine in Valmet AB, a company in Sundsvall, Sweden that supplies equipment and processes for the production of biofuels. Using deep learning models to detect failure drills on such a small sound dataset is typically unsuccessful. We employed a VAE to increase the number of sounds in the tiny dataset by synthesizing new sounds from original sounds. The augmented dataset was created by combining these synthesized sounds with the original sounds. We used a high-pass filter with a passband frequency of 1000 Hz and a low-pass filter with a passband frequency of 22\kern 0.16667em000 Hz to pre-process sounds in the augmented dataset before transforming them to Mel spectrograms. The pre-trained 2D-CNN Alexnet was then trained using these Mel spectrograms. When compared to using the original tiny sound dataset to train pre-trained Alexnet, using the augmented sound dataset enhanced the CNN model's classification results by 6.62\%(94.12\% when trained on the augmented dataset versus 87.5\% when trained on the original dataset).

Chau Nguyen, Phuong Nguyen, Thanh Tran et al.

The Competition on Legal Information Extraction/Entailment (COLIEE) is held annually to encourage advancements in the automatic processing of legal texts. Processing legal documents is challenging due to the intricate structure and meaning of legal language. In this paper, we outline our strategies for tackling Task 2, Task 3, and Task 4 in the COLIEE 2023 competition. Our approach involved utilizing appropriate state-of-the-art deep learning methods, designing methods based on domain characteristics observation, and applying meticulous engineering practices and methodologies to the competition. As a result, our performance in these tasks has been outstanding, with first places in Task 2 and Task 3, and promising results in Task 4. Our source code is available at https://github.com/Nguyen2015/CAPTAIN-COLIEE2023/tree/coliee2023.

Thanh Tran, Kyumin Lee, Yiming Liao et al.

Following recent successes in exploiting both latent factor and word embedding models in recommendation, we propose a novel Regularized Multi-Embedding (RME) based recommendation model that simultaneously encapsulates the following ideas via decomposition: (1) which items a user likes, (2) which two users co-like the same items, (3) which two items users often co-liked, and (4) which two items users often co-disliked. In experimental validation, the RME outperforms competing state-of-the-art models in both explicit and implicit feedback datasets, significantly improving Recall@5 by 5.9~7.0%, NDCG@20 by 4.3~5.6%, and MAP@10 by 7.9~8.9%. In addition, under the cold-start scenario for users with the lowest number of interactions, against the competing models, the RME outperforms NDCG@5 by 20.2% and 29.4% in MovieLens-10M and MovieLens-20M datasets, respectively. Our datasets and source code are available at: https://github.com/thanhdtran/RME.git.

Beniamin Goldys, Agus L. Soenjaya, Thanh Tran

The incompressible Hall-magnetohydrodynamics (Hall--MHD) system presents substantial analytical and computational challenges due to its stiff, highly nonlinear Hall term and the strict requirement that the magnetic field remains solenoidal. In this paper, we study a Voigt-regularised Hall--MHD system, which is of independent analytical interest and provides a physically consistent, well-posed regularisation of the original model. We propose, analyse, and implement a structure-preserving, linear, fully discrete finite element method for this regularised problem. Using finite element exterior calculus and a mixed formulation, the spatial discretisation enforces the divergence-free condition on the magnetic field exactly, while a skew-symmetric, linearly implicit time discretisation yields unconditional energy stability. We establish optimal convergence rates for the Voigt-regularised problem and, additionally, derive error estimates for the unregularised Hall--MHD system, with the Voigt regularisation playing a crucial role in the non-resistive regime. Finally, numerical simulations in both 2.5D and 3D corroborate the theoretical results and demonstrate the physical fidelity of the scheme.

Jooyoung Lee, Fan Yang, Thanh Tran et al.

We introduce a novel framework, LM-Guided CoT, that leverages a lightweight (i.e., <1B) language model (LM) for guiding a black-box large (i.e., >10B) LM in reasoning tasks. Specifically, the lightweight LM first generates a rationale for each input instance. The Frozen large LM is then prompted to predict a task output based on the rationale generated by the lightweight LM. Our approach is resource-efficient in the sense that it only requires training the lightweight LM. We optimize the model through 1) knowledge distillation and 2) reinforcement learning from rationale-oriented and task-oriented reward signals. We assess our method with multi-hop extractive question answering (QA) benchmarks, HotpotQA, and 2WikiMultiHopQA. Experimental results show that our approach outperforms all baselines regarding answer prediction accuracy. We also find that reinforcement learning helps the model to produce higher-quality rationales with improved QA performance.

Thanh Tran, Viet-Hoang Tran, Thanh Chu et al.

Tree-Sliced methods have recently emerged as an alternative to the traditional Sliced Wasserstein (SW) distance, replacing one-dimensional lines with tree-based metric spaces and incorporating a splitting mechanism for projecting measures. This approach enhances the ability to capture the topological structures of integration domains in Sliced Optimal Transport while maintaining low computational costs. Building on this foundation, we propose a novel nonlinear projectional framework for the Tree-Sliced Wasserstein (TSW) distance, substituting the linear projections in earlier versions with general projections, while ensuring the injectivity of the associated Radon Transform and preserving the well-definedness of the resulting metric. By designing appropriate projections, we construct efficient metrics for measures on both Euclidean spaces and spheres. Finally, we validate our proposed metric through extensive numerical experiments for Euclidean and spherical datasets. Applications include gradient flows, self-supervised learning, and generative models, where our methods demonstrate significant improvements over recent SW and TSW variants.

Cheng-Fu Yang, Thanh Tran, Christos Christodoulopoulos et al.

A multi-modal guardrail must effectively filter image content based on user-defined policies, identifying material that may be hateful, reinforce harmful stereotypes, contain explicit material, or spread misinformation. Deploying such guardrails in real-world applications, however, poses significant challenges. Users often require varied and highly customizable policies and typically cannot provide abundant examples for each custom policy. Consequently, an ideal guardrail should be scalable to the multiple policies and adaptable to evolving user standards with minimal retraining. Existing fine-tuning methods typically condition predictions on pre-defined policies, restricting their generalizability to new policies or necessitating extensive retraining to adapt. Conversely, training-free methods struggle with limited context lengths, making it difficult to incorporate all the policies comprehensively. To overcome these limitations, we propose to condition model's judgment on "precedents", which are the reasoning processes of prior data points similar to the given input. By leveraging precedents instead of fixed policies, our approach greatly enhances the flexibility and adaptability of the guardrail. In this paper, we introduce a critique-revise mechanism for collecting high-quality precedents and two strategies that utilize precedents for robust prediction. Experimental results demonstrate that our approach outperforms previous methods across both few-shot and full-dataset scenarios and exhibits superior generalization to novel policies.

Thanh Tran, Son T. Luu, Quan Bui et al.

This paper proposes a method for automatic GUI component detection for the IBM i system (formerly and still more commonly known as AS/400). We introduce a human-annotated dataset consisting of 1,050 system screen images, in which 381 images are screenshots of IBM i system screens in Japanese. Each image contains multiple components, including text labels, text boxes, options, tables, instructions, keyboards, and command lines. We then develop a detection system based on state-of-the-art deep learning models and evaluate different approaches using our dataset. The experimental results demonstrate the effectiveness of our dataset in constructing a system for component detection from GUI screens. By automatically detecting GUI components from the screen, AS400-DET has the potential to perform automated testing on systems that operate via GUI screens.

Michiharu Yamashita, Jia Tracy Shen, Thanh Tran et al.

In online job marketplaces, it is important to establish a well-defined job title taxonomy for various downstream tasks (e.g., job recommendation, users' career analysis, and turnover prediction). Job Title Normalization (JTN) is such a cleaning step to classify user-created non-standard job titles into normalized ones. However, solving the JTN problem is non-trivial with challenges: (1) semantic similarity of different job titles, (2) non-normalized user-created job titles, and (3) large-scale and long-tailed job titles in real-world applications. To this end, we propose a novel solution, named JAMES, that constructs three unique embeddings (i.e., graph, contextual, and syntactic) of a target job title to effectively capture its various traits. We further propose a multi-aspect co-attention mechanism to attentively combine these embeddings, and employ neural logical reasoning representations to collaboratively estimate similarities between messy job titles and normalized job titles in a reasoning space. To evaluate JAMES, we conduct comprehensive experiments against ten competing models on a large-scale real-world dataset with over 350,000 job titles. Our experimental results show that JAMES significantly outperforms the best baseline by 10.06% in Precision@10 and by 17.52% in NDCG@10, respectively.

Kai Wei, Thanh Tran, Feng-Ju Chang et al.

Recent years have seen significant advances in end-to-end (E2E) spoken language understanding (SLU) systems, which directly predict intents and slots from spoken audio. While dialogue history has been exploited to improve conventional text-based natural language understanding systems, current E2E SLU approaches have not yet incorporated such critical contextual signals in multi-turn and task-oriented dialogues. In this work, we propose a contextual E2E SLU model architecture that uses a multi-head attention mechanism over encoded previous utterances and dialogue acts (actions taken by the voice assistant) of a multi-turn dialogue. We detail alternative methods to integrate these contexts into the state-ofthe-art recurrent and transformer-based models. When applied to a large de-identified dataset of utterances collected by a voice assistant, our method reduces average word and semantic error rates by 10.8% and 12.6%, respectively. We also present results on a publicly available dataset and show that our method significantly improves performance over a noncontextual baseline

Thanh Tran, Nhat Truong Pham, Jan Lundgren

Monitoring the conditions of machines is vital in the manufacturing industry. Early detection of faulty components in machines for stopping and repairing the failed components can minimize the downtime of the machine. This article presents an approach to detect the failure occurring in drill machines based on drill sounds from Valmet AB. The drill dataset includes three classes: anomalous sounds, normal sounds, and irrelevant sounds, which are also labeled as "Broken", "Normal", and "Other", respectively. Detecting drill failure effectively remains a challenge due to the following reasons. The waveform of drill sound is complex and short for detection. Additionally, in realistic soundscapes, there are sounds and noise in the context at the same time. Moreover, the balanced dataset is small to apply state-of-the-art deep learning techniques. To overcome these aforementioned difficulties, we augmented sounds to increase the number of sounds in the dataset. We then proposed a convolutional neural network (CNN) combined with a long short-term memory (LSTM) to extract features from log-Mel spectrograms and learn global high-level feature representation for the classification of three classes. A leaky rectified linear unit (Leaky ReLU) was utilized as the activation function for our proposed CNN instead of the rectified linear unit (ReLU). Moreover, we deployed an attention mechanism at the frame level after the LSTM layer to learn long-term global feature representations. As a result, the proposed method reached an overall accuracy of 92.35% for the drill failure detection system.

Yifan Hu, Changwei Hu, Thanh Tran et al.

Gender information is no longer a mandatory input when registering for an account at many leading Internet companies. However, prediction of demographic information such as gender and age remains an important task, especially in intervention of unintentional gender/age bias in recommender systems. Therefore it is necessary to infer the gender of those users who did not to provide this information during registration. We consider the problem of predicting the gender of registered users based on their declared name. By analyzing the first names of 100M+ users, we found that genders can be very effectively classified using the composition of the name strings. We propose a number of character based machine learning models, and demonstrate that our models are able to infer the gender of users with much higher accuracy than baseline models. Moreover, we show that using the last names in addition to the first names improves classification performance further.

Thanh Tran, Yifan Hu, Changwei Hu et al.

We present our HABERTOR model for detecting hatespeech in large scale user-generated content. Inspired by the recent success of the BERT model, we propose several modifications to BERT to enhance the performance on the downstream hatespeech classification task. HABERTOR inherits BERT's architecture, but is different in four aspects: (i) it generates its own vocabularies and is pre-trained from the scratch using the largest scale hatespeech dataset; (ii) it consists of Quaternion-based factorized components, resulting in a much smaller number of parameters, faster training and inferencing, as well as less memory usage; (iii) it uses our proposed multi-source ensemble heads with a pooling layer for separate input sources, to further enhance its effectiveness; and (iv) it uses a regularized adversarial training with our proposed fine-grained and adaptive noise magnitude to enhance its robustness. Through experiments on the large-scale real-world hatespeech dataset with 1.4M annotated comments, we show that HABERTOR works better than 15 state-of-the-art hatespeech detection methods, including fine-tuning Language Models. In particular, comparing with BERT, our HABERTOR is 4~5 times faster in the training/inferencing phase, uses less than 1/3 of the memory, and has better performance, even though we pre-train it by using less than 1% of the number of words. Our generalizability analysis shows that HABERTOR transfers well to other unseen hatespeech datasets and is a more efficient and effective alternative to BERT for the hatespeech classification.

Thanh Tran, Di You, Kyumin Lee

Quaternion space has brought several benefits over the traditional Euclidean space: Quaternions (i) consist of a real and three imaginary components, encouraging richer representations; (ii) utilize Hamilton product which better encodes the inter-latent interactions across multiple Quaternion components; and (iii) result in a model with smaller degrees of freedom and less prone to overfitting. Unfortunately, most of the current recommender systems rely on real-valued representations in Euclidean space to model either user's long-term or short-term interests. In this paper, we fully utilize Quaternion space to model both user's long-term and short-term preferences. We first propose a QUaternion-based self-Attentive Long term user Encoding (QUALE) to study the user's long-term intents. Then, we propose a QUaternion-based self-Attentive Short term user Encoding (QUASE) to learn the user's short-term interests. To enhance our models' capability, we propose to fuse QUALE and QUASE into one model, namely QUALSE, by using a Quaternion-based gating mechanism. We further develop Quaternion-based Adversarial learning along with the Bayesian Personalized Ranking (QABPR) to improve our model's robustness. Extensive experiments on six real-world datasets show that our fused QUALSE model outperformed 11 state-of-the-art baselines, improving 8.43% at HIT@1 and 10.27% at NDCG@1 on average compared with the best baseline.

Thanh Tran, Renee Sweeney, Kyumin Lee

In this paper, we aim to solve the automatic playlist continuation (APC) problem by modeling complex interactions among users, playlists, and songs using only their interaction data. Prior methods mainly rely on dot product to account for similarities, which is not ideal as dot product is not metric learning, so it does not convey the important inequality property. Based on this observation, we propose three novel deep learning approaches that utilize Mahalanobis distance. Our first approach uses user-playlist-song interactions, and combines Mahalanobis distance scores between (i) a target user and a target song, and (ii) between a target playlist and the target song to account for both the user's preference and the playlist's theme. Our second approach measures song-song similarities by considering Mahalanobis distance scores between the target song and each member song (i.e., existing song) in the target playlist. The contribution of each distance score is measured by our proposed memory metric-based attention mechanism. In the third approach, we fuse the two previous models into a unified model to further enhance their performance. In addition, we adopt and customize Adversarial Personalized Ranking (APR) for our three approaches to further improve their robustness and predictive capabilities. Through extensive experiments, we show that our proposed models outperform eight state-of-the-art models in two large-scale real-world datasets.

Thanh Tran, Xinyue Liu, Kyumin Lee et al.

Personalized recommendation algorithms learn a user's preference for an item by measuring a distance/similarity between them. However, some of the existing recommendation models (e.g., matrix factorization) assume a linear relationship between the user and item. This approach limits the capacity of recommender systems, since the interactions between users and items in real-world applications are much more complex than the linear relationship. To overcome this limitation, in this paper, we design and propose a deep learning framework called Signed Distance-based Deep Memory Recommender, which captures non-linear relationships between users and items explicitly and implicitly, and work well in both general recommendation task and shopping basket-based recommendation task. Through an extensive empirical study on six real-world datasets in the two recommendation tasks, our proposed approach achieved significant improvement over ten state-of-the-art recommendation models.

Jacopo Acquarelli, Elena Marchiori, Lutgarde M. C. Buydens et al.

Spectral-spatial classification of hyperspectral images has been the subject of many studies in recent years. In the presence of only very few labeled pixels, this task becomes challenging. In this paper we address the following two research questions: 1) Can a simple neural network with just a single hidden layer achieve state of the art performance in the presence of few labeled pixels? 2) How is the performance of hyperspectral image classification methods affected when using disjoint train and test sets? We give a positive answer to the first question by using three tricks within a very basic shallow Convolutional Neural Network (CNN) architecture: a tailored loss function, and smooth- and label-based data augmentation. The tailored loss function enforces that neighborhood wavelengths have similar contributions to the features generated during training. A new label-based technique here proposed favors selection of pixels in smaller classes, which is beneficial in the presence of very few labeled pixels and skewed class distributions. To address the second question, we introduce a new sampling procedure to generate disjoint train and test set. Then the train set is used to obtain the CNN model, which is then applied to pixels in the test set to estimate their labels. We assess the efficacy of the simple neural network method on five publicly available hyperspectral images. On these images our method significantly outperforms considered baselines. Notably, with just 1% of labeled pixels per class, on these datasets our method achieves an accuracy that goes from 86.42% (challenging dataset) to 99.52% (easy dataset). Furthermore we show that the simple neural network method improves over other baselines in the new challenging supervised setting. Our analysis substantiates the highly beneficial effect of using the entire image (so train and test data) for constructing a model.

Thanh Tran, Kyumin Lee, Nguyen Vo et al.

In Crowdfunding platforms, people turn their prototype ideas into real products by raising money from the crowd, or invest in someone else's projects. In reward-based crowdfunding platforms such as Kickstarter and Indiegogo, selecting accurate reward delivery duration becomes crucial for creators, backers, and platform providers to keep the trust between the creators and the backers, and the trust between the platform providers and users. According to Kickstarter, 35% backers did not receive rewards on time. Unfortunately, little is known about on-time and late reward delivery projects, and there is no prior work to estimate reward delivery duration. To fill the gap, in this paper, we (i) extract novel features that reveal latent difficulty levels of project rewards; (ii) build predictive models to identify whether a creator will deliver all rewards in a project on time or not; and (iii) build a regression model to estimate accurate reward delivery duration (i.e., how long it will take to produce and deliver all the rewards). Experimental results show that our models achieve good performance -- 82.5% accuracy, 78.1 RMSE, and 0.108 NRMSE at the first 5% of the longest reward delivery duration.