The rapid scaling of language models is motivating research using low-bitwidth quantization. In this work, we propose a novel binarization technique for Transformers applied to machine translation (BMT), the first of its kind. We identify and address the problem of inflated dot-product variance when using one-bit weights and activations. Specifically, BMT leverages additional LayerNorms and residual connections to improve binarization quality. Experiments on the WMT dataset show that a one-bit weight-only Transformer can achieve the same quality as a float one, while being 16x smaller in size. One-bit activations incur varying degrees of quality drop, but mitigated by the proposed architectural changes. We further conduct a scaling law study using production-scale translation datasets, which shows that one-bit weight Transformers scale and generalize well in both in-domain and out-of-domain settings. Implementation in JAX/Flax will be open sourced.
Recent advances of data-driven machine learning have revolutionized fields like computer vision, reinforcement learning, and many scientific and engineering domains. In many real-world and scientific problems, systems that generate data are governed by physical laws. Recent work shows that it provides potential benefits for machine learning models by incorporating the physical prior and collected data, which makes the intersection of machine learning and physics become a prevailing paradigm. In this survey, we present this learning paradigm called Physics-Informed Machine Learning (PIML) which is to build a model that leverages empirical data and available physical prior knowledge to improve performance on a set of tasks that involve a physical mechanism. We systematically review the recent development of physics-informed machine learning from three perspectives of machine learning tasks, representation of physical prior, and methods for incorporating physical prior. We also propose several important open research problems based on the current trends in the field. We argue that encoding different forms of physical prior into model architectures, optimizers, inference algorithms, and significant domain-specific applications like inverse engineering design and robotic control is far from fully being explored in the field of physics-informed machine learning. We believe that this study will encourage researchers in the machine learning community to actively participate in the interdisciplinary research of physics-informed machine learning.
Recent years have seen an increasing amount of work on embodied AI agents that can perform tasks by following human language instructions. However, most of these agents are reactive, meaning that they simply learn and imitate behaviors encountered in the training data. These reactive agents are insufficient for long-horizon complex tasks. To address this limitation, we propose a neuro-symbolic deliberative agent that, while following language instructions, proactively applies reasoning and planning based on its neural and symbolic representations acquired from past experience (e.g., natural language and egocentric vision). We show that our deliberative agent achieves greater than 70% improvement over reactive baselines on the challenging TEACh benchmark. Moreover, the underlying reasoning and planning processes, together with our modular framework, offer impressive transparency and explainability to the behaviors of the agent. This enables an in-depth understanding of the agent's capabilities, which shed light on challenges and opportunities for future embodied agents for instruction following. The code is available at https://github.com/sled-group/DANLI.
In this work, we share our experience on tele-knowledge pre-training for fault analysis. Fault analysis is a vital task for tele-application, which should be timely and properly handled. Fault analysis is also a complex task, that has many sub-tasks. Solving each task requires diverse tele-knowledge. Machine log data and product documents contain part of the tele-knowledge. We create a Tele-KG to organize other tele-knowledge from experts uniformly. With these valuable tele-knowledge data, in this work, we propose a tele-domain pre-training model KTeleBERT and its knowledge-enhanced version KTeleBERT, which includes effective prompt hints, adaptive numerical data encoding, and two knowledge injection paradigms. We train our model in two stages: pre-training TeleBERT on 20 million telecommunication corpora and re-training TeleBERT on 1 million causal and machine corpora to get the KTeleBERT. Then, we apply our models for three tasks of fault analysis, including root-cause analysis, event association prediction, and fault chain tracing. The results show that with KTeleBERT, the performance of task models has been boosted, demonstrating the effectiveness of pre-trained KTeleBERT as a model containing diverse tele-knowledge.
Knowledge graphs (KGs) that modelings the world knowledge as structural triples are inevitably incomplete. Such problems still exist for multimodal knowledge graphs (MMKGs). Thus, knowledge graph completion (KGC) is of great importance to predict the missing triples in the existing KGs. As for the existing KGC methods, embedding-based methods rely on manual design to leverage multimodal information while finetune-based approaches are not superior to embedding-based methods in link prediction. To address these problems, we propose a VisualBERT-enhanced Knowledge Graph Completion model (VBKGC for short). VBKGC could capture deeply fused multimodal information for entities and integrate them into the KGC model. Besides, we achieve the co-design of the KGC model and negative sampling by designing a new negative sampling strategy called twins negative sampling. Twins negative sampling is suitable for multimodal scenarios and could align different embeddings for entities. We conduct extensive experiments to show the outstanding performance of VBKGC on the link prediction task and make further exploration of VBKGC.
Medical image segmentation is a fundamental and critical step in many image-guided clinical approaches. Recent success of deep learning-based segmentation methods usually relies on a large amount of labeled data, which is particularly difficult and costly to obtain especially in the medical imaging domain where only experts can provide reliable and accurate annotations. Semi-supervised learning has emerged as an appealing strategy and been widely applied to medical image segmentation tasks to train deep models with limited annotations. In this paper, we present a comprehensive review of recently proposed semi-supervised learning methods for medical image segmentation and summarized both the technical novelties and empirical results. Furthermore, we analyze and discuss the limitations and several unsolved problems of existing approaches. We hope this review could inspire the research community to explore solutions for this challenge and further promote the developments in medical image segmentation field.
Randomized singular value decomposition (RSVD) is a class of computationally efficient algorithms for computing the truncated SVD of large data matrices. Given a $n \times n$ symmetric matrix $\mathbf{M}$, the prototypical RSVD algorithm outputs an approximation of the $k$ leading singular vectors of $\mathbf{M}$ by computing the SVD of $\mathbf{M}^{g} \mathbf{G}$; here $g \geq 1$ is an integer and $\mathbf{G} \in \mathbb{R}^{n \times k}$ is a random Gaussian sketching matrix. In this paper we study the statistical properties of RSVD under a general "signal-plus-noise" framework, i.e., the observed matrix $\hat{\mathbf{M}}$ is assumed to be an additive perturbation of some true but unknown signal matrix $\mathbf{M}$. We first derive upper bounds for the $\ell_2$ (spectral norm) and $\ell_{2\to\infty}$ (maximum row-wise $\ell_2$ norm) distances between the approximate singular vectors of $\hat{\mathbf{M}}$ and the true singular vectors of the signal matrix $\mathbf{M}$. These upper bounds depend on the signal-to-noise ratio (SNR) and the number of power iterations $g$. A phase transition phenomenon is observed in which a smaller SNR requires larger values of $g$ to guarantee convergence of the $\ell_2$ and $\ell_{2\to\infty}$ distances. We also show that the thresholds for $g$ where these phase transitions occur are sharp whenever the noise matrices satisfy a certain trace growth condition. Finally, we derive normal approximations for the row-wise fluctuations of the approximate singular vectors and the entrywise fluctuations of the approximate matrix. We illustrate our theoretical results by deriving nearly-optimal performance guarantees for RSVD when applied to three statistical inference problems, namely, community detection, matrix completion, and principal component analysis with missing data.
Hyperdimensional computing (HDC) is an emerging learning paradigm that computes with high dimensional binary vectors. It is attractive because of its energy efficiency and low latency, especially on emerging hardware -- but HDC suffers from low model accuracy, with little theoretical understanding of what limits its performance. We propose a new theoretical analysis of the limits of HDC via a consideration of what similarity matrices can be "expressed" by binary vectors, and we show how the limits of HDC can be approached using random Fourier features (RFF). We extend our analysis to the more general class of vector symbolic architectures (VSA), which compute with high-dimensional vectors (hypervectors) that are not necessarily binary. We propose a new class of VSAs, finite group VSAs, which surpass the limits of HDC. Using representation theory, we characterize which similarity matrices can be "expressed" by finite group VSA hypervectors, and we show how these VSAs can be constructed. Experimental results show that our RFF method and group VSA can both outperform the state-of-the-art HDC model by up to 7.6\% while maintaining hardware efficiency.
Medical image segmentation is a fundamental and critical step in many clinical approaches. Semi-supervised learning has been widely applied to medical image segmentation tasks since it alleviates the heavy burden of acquiring expert-examined annotations and takes the advantage of unlabeled data which is much easier to acquire. Although consistency learning has been proven to be an effective approach by enforcing an invariance of predictions under different distributions, existing approaches cannot make full use of region-level shape constraint and boundary-level distance information from unlabeled data. In this paper, we propose a novel uncertainty-guided mutual consistency learning framework to effectively exploit unlabeled data by integrating intra-task consistency learning from up-to-date predictions for self-ensembling and cross-task consistency learning from task-level regularization to exploit geometric shape information. The framework is guided by the estimated segmentation uncertainty of models to select out relatively certain predictions for consistency learning, so as to effectively exploit more reliable information from unlabeled data. We extensively validate our proposed method on two publicly available benchmark datasets: Left Atrium Segmentation (LA) dataset and Brain Tumor Segmentation (BraTS) dataset. Experimental results demonstrate that our method achieves performance gains by leveraging unlabeled data and outperforms existing semi-supervised segmentation methods.