Search

Current time-synchronous sequence-to-sequence automatic speech recognition (ASR) models are trained by using sequence level cross-entropy that sums over all alignments. Due to the discriminative formulation, incorporating the right label context into the training criterion's gradient causes normalization problems and is not mathematically well-defined. The classic hybrid neural network hidden Markov model (NN-HMM) with its inherent generative formulation enables conditioning on the right label context. However, due to the HMM state-tying the identity of the right label context is never modeled explicitly. In this work, we propose a factored loss with auxiliary left and right label contexts that sums over all alignments. We show that the inclusion of the right label context is particularly beneficial when training data resources are limited. Moreover, we also show that it is possible to build a factored hybrid HMM system by relying exclusively on the full-sum criterion. Experiments were conducted on Switchboard 300h and LibriSpeech 960h., Comment: Accepted for presentation at ICASSP 2025

Self-supervised learning (SSL) has revolutionized learning from large-scale unlabeled datasets, yet the intrinsic relationship between pretraining data and the learned representations remains poorly understood. Traditional supervised learning benefits from gradient-based data attribution tools like influence functions that measure the contribution of an individual data point to model predictions. However, existing definitions of influence rely on labels, making them unsuitable for SSL settings. We address this gap by introducing Influence-SSL, a novel and label-free approach for defining influence functions tailored to SSL. Our method harnesses the stability of learned representations against data augmentations to identify training examples that help explain model predictions. We provide both theoretical foundations and empirical evidence to show the utility of Influence-SSL in analyzing pre-trained SSL models. Our analysis reveals notable differences in how SSL models respond to influential data compared to supervised models. Finally, we validate the effectiveness of Influence-SSL through applications in duplicate detection, outlier identification and fairness analysis. Code is available at: \url{https://github.com/cryptonymous9/Influence-SSL}.

In this paper, we review the new method Physics-Informed Neural Networks (PINNs) that has become the main pillar in scientific machine learning, we present recent practical extensions, and provide a specific example in data-driven discovery of governing differential equations., Comment: Frontiers of Science Awards 2024

Graph states are a class of multi-partite entangled quantum states, where colorability, a property rooted in their mathematical foundation, has significant implications for quantum information processing. In this study, we investigate the colorability of graph states in qudit systems to simplify their representation and enhance their practical applications. We present closed-form expressions for all two-colorable graph states. Our findings show that the closed-form expression of these states is tightly linked to the graph structure and the distribution of particles in red ($n_R$) and blue ($n_B$). Additionally, we explore a broad family of three-colorable graph states constructed from two two-colorable graph states. The closed-form expression for these states is in the form of one two-colorable state tensor product with the graph basis formed from another two-colorable state. Our approach systematically reduces the number of terms required to represent these states. Furthermore, we demonstrate that many well-known mathematical graphs, including friendship graphs, fit within our formalism. Finally, we discuss the LU/SLOCC (Local Unitary/Stochastic Local Operation and Classical Communication) equivalence between two- and three-colorable graph states. Our findings have broad implications for characterizing the LU/SLOCC equivalence of graph state classes and pave the way for future research., Comment: 13 pages + appendix and references, 22 figures. Comments are welcome!

Non-symmetric GHZ states ($n$-GHZ$_\alpha$), characterized by unequal superpositions of $|00...0>$ and $|11...1>$, represent a significant yet underexplored class of multipartite entangled states with potential applications in quantum information. Despite their importance, the lack of a well-defined stabilizer formalism and corresponding graph representation has hindered their comprehensive study. In this paper, we address this gap by introducing two novel graph formalisms and stabilizers for non-symmetric GHZ states. First, we provide a weighted hypergraph representation and demonstrate that non-symmetric GHZ states are local unitary (LU) equivalent to fully connected weighted hypergraphs. Although these weighted hypergraphs are not stabilizer states, we show that they can be stabilized using local operations, and an ancilla. We further extend this framework to qudits, offering a specific form for non-symmetric qudit GHZ states and their LU equivalent weighted qudit hypergraphs. Second, we propose a graph formalism using controlled-unitary (CU) operations, showing that non-symmetric qudit GHZ states can be described using star-shaped CU graphs. Our findings enhance the understanding of non-symmetric GHZ states and their potential applications in quantum information science., Comment: 13 pages(main text + Appendix), 4 figures. Comments and suggestions are very welcome!

The ongoing research scenario for automatic speech recognition (ASR) envisions a clear division between end-to-end approaches and classic modular systems. Even though a high-level comparison between the two approaches in terms of their requirements and (dis)advantages is commonly addressed, a closer comparison under similar conditions is not readily available in the literature. In this work, we present a comparison focused on the label topology and training criterion. We compare two discriminative alignment models with hidden Markov model (HMM) and connectionist temporal classification topology, and two first-order label context ASR models utilizing factored HMM and strictly monotonic recurrent neural network transducer, respectively. We use different measurements for the evaluation of the alignment quality, and compare word error rate and real time factor of our best systems. Experiments are conducted on the LibriSpeech 960h and Switchboard 300h tasks., Comment: Accepted for presentation at Interspeech 2024

Compared to physics-based computational manufacturing, data-driven models such as machine learning (ML) are alternative approaches to achieve smart manufacturing. However, the data-driven ML's "black box" nature has presented a challenge to interpreting its outcomes. On the other hand, governing physical laws are not effectively utilized to develop data-efficient ML algorithms. To leverage the advantages of ML and physical laws of advanced manufacturing, this paper focuses on the development of a physics-informed machine learning (PIML) model by integrating neural networks and physical laws to improve model accuracy, transparency, and generalization with case studies in laser metal deposition (LMD)., Comment: 6 pages, 7 figures, 18th CIRP Conference on Intelligent Computation in Manufacturing Engineering

This paper introduces MixDiff, a new self-supervised learning (SSL) pre-training framework that combines real and synthetic images. Unlike traditional SSL methods that predominantly use real images, MixDiff uses a variant of Stable Diffusion to replace an augmented instance of a real image, facilitating the learning of cross real-synthetic image representations. Our key insight is that while models trained solely on synthetic images underperform, combining real and synthetic data leads to more robust and adaptable representations. Experiments show MixDiff enhances SimCLR, BarlowTwins, and DINO across various robustness datasets and domain transfer tasks, boosting SimCLR's ImageNet-1K accuracy by 4.56%. Our framework also demonstrates comparable performance without needing any augmentations, a surprising finding in SSL where augmentations are typically crucial., Comment: Accepted in WACV 2025

The Log-Periodic Power Law Singularity (LPPLS) model offers a general framework for capturing dynamics and predicting transition points in diverse natural and social systems. In this work, we present two calibration techniques for the LPPLS model using deep learning. First, we introduce the Mono-LPPLS-NN (M-LNN) model; for any given empirical time series, a unique M-LNN model is trained and shown to outperform state-of-the-art techniques in estimating the nonlinear parameters $(t_c, m, \omega)$ of the LPPLS model as evidenced by the comprehensive distribution of parameter errors. Second, we extend the M-LNN model to a more general model architecture, the Poly-LPPLS-NN (P-LNN), which is able to quickly estimate the nonlinear parameters of the LPPLS model for any given time-series of a fixed length, including previously unseen time-series during training. The Poly class of models train on many synthetic LPPLS time-series augmented with various noise structures in a supervised manner. Given enough training examples, the P-LNN models also outperform state-of-the-art techniques for estimating the parameters of the LPPLS model as evidenced by the comprehensive distribution of parameter errors. Additionally, this class of models is shown to substantially reduce the time to obtain parameter estimates. Finally, we present applications to the diagnostic and prediction of two financial bubble peaks (followed by their crash) and of a famous rockslide. These contributions provide a bridge between deep learning and the study of the prediction of transition times in complex time series., Comment: 10 pages, 6 figures

Artificial neural networks often suffer from catastrophic forgetting, where learning new concepts leads to a complete loss of previously acquired knowledge. We observe that this issue is particularly magnified in vision transformers (ViTs), where post-pre-training and fine-tuning on new tasks can significantly degrade the model's original general abilities. For instance, a DINO ViT-Base/16 pre-trained on ImageNet-1k loses over 70% accuracy on ImageNet-1k after just 10 iterations of fine-tuning on CIFAR-100. Overcoming this stability-plasticity dilemma is crucial for enabling ViTs to continuously learn and adapt to new domains while preserving their initial knowledge. In this work, we study two new parameter-efficient fine-tuning strategies: (1)~Block Expansion, and (2) Low-rank adaptation (LoRA). Our experiments reveal that using either Block Expansion or LoRA on self-supervised pre-trained ViTs surpass fully fine-tuned ViTs in new domains while offering significantly greater parameter efficiency. Notably, we find that Block Expansion experiences only a minimal performance drop in the pre-training domain, thereby effectively mitigating catastrophic forgetting in pre-trained ViTs., Comment: Accepted at eLVM Workshop, CVPR, 2024

3D modeling holds significant importance in the realms of AR/VR and gaming, allowing for both artistic creativity and practical applications. However, the process is often time-consuming and demands a high level of skill. In this paper, we present a novel approach to create volumetric representations of 3D characters from consistent turnaround concept art, which serves as the standard input in the 3D modeling industry. While Neural Radiance Field (NeRF) has been a game-changer in image-based 3D reconstruction, to the best of our knowledge, there is no known research that optimizes the pipeline for concept art. To harness the potential of concept art, with its defined body poses and specific view angles, we propose encoding it as priors for our model. We train the network to make use of these priors for various 3D points through a learnable view-direction-attended multi-head self-attention layer. Additionally, we demonstrate that a combination of ray sampling and surface sampling enhances the inference capabilities of our network. Our model is able to generate high-quality 360-degree views of characters. Subsequently, we provide a simple guideline to better leverage our model to extract the 3D mesh. It is important to note that our model's inferencing capabilities are influenced by the training data's characteristics, primarily focusing on characters with a single head, two arms, and two legs. Nevertheless, our methodology remains versatile and adaptable to concept art from diverse subject matters, without imposing any specific assumptions on the data.

Physics-informed neural networks (PINNs) have gained prominence for their capability to tackle supervised learning tasks that conform to physical laws, notably nonlinear partial differential equations (PDEs). This paper presents "PINNs-TF2", a Python package built on the TensorFlow V2 framework. It not only accelerates PINNs implementation but also simplifies user interactions by abstracting complex PDE challenges. We underscore the pivotal role of compilers in PINNs, highlighting their ability to boost performance by up to 119x. Across eight diverse examples, our package, integrated with XLA compilers, demonstrated its flexibility and achieved an average speed-up of 18.12 times over TensorFlow V1. Moreover, a real-world case study is implemented to underscore the compilers' potential to handle many trainable parameters and large batch sizes. For community engagement and future enhancements, our package's source code is openly available at: https://github.com/rezaakb/pinns-tf2., Comment: Accepted at Machine Learning and the Physical Sciences Workshop, NeurIPS 2023

Automatic crash bucketing is a crucial phase in the software development process for efficiently triaging bug reports. It generally consists in grouping similar reports through clustering techniques. However, with real-time streaming bug collection, systems are needed to quickly answer the question: What are the most similar bugs to a new one?, that is, efficiently find near-duplicates. It is thus natural to consider nearest neighbors search to tackle this problem and especially the well-known locality-sensitive hashing (LSH) to deal with large datasets due to its sublinear performance and theoretical guarantees on the similarity search accuracy. Surprisingly, LSH has not been considered in the crash bucketing literature. It is indeed not trivial to derive hash functions that satisfy the so-called locality-sensitive property for the most advanced crash bucketing metrics. Consequently, we study in this paper how to leverage LSH for this task. To be able to consider the most relevant metrics used in the literature, we introduce DeepLSH, a Siamese DNN architecture with an original loss function, that perfectly approximates the locality-sensitivity property even for Jaccard and Cosine metrics for which exact LSH solutions exist. We support this claim with a series of experiments on an original dataset, which we make available.

We investigate a novel modeling approach for end-to-end neural network training using hidden Markov models (HMM) where the transition probabilities between hidden states are modeled and learned explicitly. Most contemporary sequence-to-sequence models allow for from-scratch training by summing over all possible label segmentations in a given topology. In our approach there are explicit, learnable probabilities for transitions between segments as opposed to a blank label that implicitly encodes duration statistics. We implement a GPU-based forward-backward algorithm that enables the simultaneous training of label and transition probabilities. We investigate recognition results and additionally Viterbi alignments of our models. We find that while the transition model training does not improve recognition performance, it has a positive impact on the alignment quality. The generated alignments are shown to be viable targets in state-of-the-art Viterbi trainings., Comment: Accepted for Presentation at ASRU2023

Control and state estimation procedures need to be robust against imprecisely known parameters, uncertainty in initial conditions, and external disturbances. Interval methods and other set-based techniques form the basis for the implementation of powerful approaches that can be used to identify parameters of dynamic system models in the presence of the aforementioned types of uncertainty. Moreover, they are applicable to a verified feasibility and stability analysis of controllers and state estimators. In addition to these approaches which are typically used offline for analysis of system models designed with classical floating point procedures, interval and set-based methods have also been developed in recent years, which allow to directly solve the associated design tasks and to implement reliable techniques that are applicable online, i.e., during system operation. The latter approaches include set-based model predictive control, online parameter adaptation techniques for nonlinear variable-structure and backstepping controllers, interval observers, and fault diagnosis techniques. This paper provides an overview of the methodological background and reviews numerous practical applications for which interval and other set-valued approaches have been employed successfully.

The inclusion of physical information in machine learning frameworks has revolutionized many application areas. This involves enhancing the learning process by incorporating physical constraints and adhering to physical laws. In this work we explore their utility for reinforcement learning applications. We present a thorough review of the literature on incorporating physics information, as known as physics priors, in reinforcement learning approaches, commonly referred to as physics-informed reinforcement learning (PIRL). We introduce a novel taxonomy with the reinforcement learning pipeline as the backbone to classify existing works, compare and contrast them, and derive crucial insights. Existing works are analyzed with regard to the representation/ form of the governing physics modeled for integration, their specific contribution to the typical reinforcement learning architecture, and their connection to the underlying reinforcement learning pipeline stages. We also identify core learning architectures and physics incorporation biases (i.e., observational, inductive and learning) of existing PIRL approaches and use them to further categorize the works for better understanding and adaptation. By providing a comprehensive perspective on the implementation of the physics-informed capability, the taxonomy presents a cohesive approach to PIRL. It identifies the areas where this approach has been applied, as well as the gaps and opportunities that exist. Additionally, the taxonomy sheds light on unresolved issues and challenges, which can guide future research. This nascent field holds great potential for enhancing reinforcement learning algorithms by increasing their physical plausibility, precision, data efficiency, and applicability in real-world scenarios.

Melt pool dynamics in metal additive manufacturing (AM) is critical to process stability, microstructure formation, and final properties of the printed materials. Physics-based simulation including computational fluid dynamics (CFD) is the dominant approach to predict melt pool dynamics. However, the physics-based simulation approaches suffer from the inherent issue of very high computational cost. This paper provides a physics-informed machine learning (PIML) method by integrating neural networks with the governing physical laws to predict the melt pool dynamics such as temperature, velocity, and pressure without using any training data on velocity. This approach avoids solving the highly non-linear Navier-Stokes equation numerically, which significantly reduces the computational cost. The difficult-to-determine model constants of the governing equations of the melt pool can also be inferred through data-driven discovery. In addition, the physics-informed neural network (PINN) architecture has been optimized for efficient model training. The data-efficient PINN model is attributed to the soft penalty by incorporating governing partial differential equations (PDEs), initial conditions, and boundary conditions in the PINN model.

The origin of high-energy cosmic rays, atomic nuclei that continuously impact Earth's atmosphere, has been a mystery for over a century. Due to deflection in interstellar magnetic fields, cosmic rays from the Milky Way arrive at Earth from random directions. However, near their sources and during propagation, cosmic rays interact with matter and produce high-energy neutrinos. We search for neutrino emission using machine learning techniques applied to ten years of data from the IceCube Neutrino Observatory. We identify neutrino emission from the Galactic plane at the 4.5$\sigma$ level of significance, by comparing diffuse emission models to a background-only hypothesis. The signal is consistent with modeled diffuse emission from the Galactic plane, but could also arise from a population of unresolved point sources., Comment: Submitted on May 12th, 2022; Accepted on May 4th, 2023

Building competitive hybrid hidden Markov model~(HMM) systems for automatic speech recognition~(ASR) requires a complex multi-stage pipeline consisting of several training criteria. The recent sequence-to-sequence models offer the advantage of having simpler pipelines that can start from-scratch. We propose a purely neural based single-stage from-scratch pipeline for a context-dependent hybrid HMM that offers similar simplicity. We use an alignment from a full-sum trained zero-order posterior HMM with a BLSTM encoder. We show that with this alignment we can build a Conformer factored hybrid that performs even better than both a state-of-the-art classic hybrid and a factored hybrid trained with alignments taken from more complex Gaussian mixture based systems. Our finding is confirmed on Switchboard 300h and LibriSpeech 960h tasks with comparable results to other approaches in the literature, and by additionally relying on a responsible choice of available computational resources., Comment: Accepted for presentation at InterSpeech 2023