How language models process complex input that requires multiple steps of inference is not well understood. Previous research has shown that information about intermediate values of these inputs can be extracted from the activations of the models, but it is unclear where that information is encoded and whether that information is indeed used during inference. We introduce a method for analyzing how a Transformer model processes these inputs by focusing on simple arithmetic problems and their intermediate values. To trace where information about intermediate values is encoded, we measure the correlation between intermediate values and the activations of the model using principal component analysis (PCA). Then, we perform a causal intervention by manipulating model weights. This intervention shows that the weights identified via tracing are not merely correlated with intermediate values, but causally related to model predictions. Our findings show that the model has a locality to certain intermediate values, and this is useful for enhancing the interpretability of the models.

The Stochastic Gradient Langevin Dynamics (SGLD), first introduced by Welling and Teh [25], has attracted a lot of attention in various areas [18, 26, 4]. The SGLD algorithm and its variants have fantastic performance when dealing with many practical sampling or optimization tasks. Recent decades have witnessed great development of theoretical research for SGLD, where most researchers focus on its discretization error, namely, the "distance" between the SGLD algorithm and the corresponding Langevin diffusion in terms of the time step (or learning rate) η [12, 18, 26, 16]. The SGLD itself can be regarded as a stochastic process and the ergodicity is also of great importance. So far, the justification of the geometric ergodicity of SGLD mostly relies on the strong convexity conditions, namely, the strong log-concaveness of the target distribution. In [4], under strong convexity settings, the authors considered the Synchronous coupling and established the geometric ergodicity of SGLD and some other numerical schemes in terms of Wasserstein-2 distance. However, the strong convexity assumption seems to limit the applicability of the result, and the ergodicity of the SGLD algorithm in a general setting and the existence of an invariant measure are still unclear. In our work, we aim to study the geometric ergodicity under locally nonconvex setting in this paper. The main technique we apply is reflection coupling [8], which was originally designed earlier to study the contraction property of many continuous SDEs.

Prescription medications often impose temporal constraints on regular health behaviors (RHBs) of patients, e.g., eating before taking medication. Violations of such medical temporal constraints (MTCs) can result in adverse effects. Detecting and predicting such violations before they occur can help alert the patient. We formulate the problem of modeling MTCs and develop a proof-of-concept solution, ActSafe, to predict violations of MTCs well ahead of time. ActSafe utilizes a context-free grammar based approach for extracting and mapping MTCs from patient education materials. It also addresses the challenges of accurately predicting RHBs central to MTCs (e.g., medication intake). Our novel behavior prediction model, HERBERT , utilizes a basis vectorization of time series that is generalizable across temporal scale and duration of behaviors, explicitly capturing the dependency between temporally collocated behaviors. Based on evaluation using a real-world RHB dataset collected from 28 patients in uncontrolled environments, HERBERT outperforms baseline models with an average of 51% reduction in root mean square error. Based on an evaluation involving patients with chronic conditions, ActSafe can predict MTC violations a day ahead of time with an average F1 score of 0.86.

Web-based interactions can be frequently represented by an attributed graph, and node clustering in such graphs has received much attention lately. Multiple efforts have successfully applied Graph Convolutional Networks (GCN), though with some limits on accuracy as GCNs have been shown to suffer from over-smoothing issues. Though other methods (particularly those based on Laplacian Smoothing) have reported better accuracy, a fundamental limitation of all the work is a lack of scalability. This paper addresses this open problem by relating the Laplacian smoothing to the Generalized PageRank and applying a random-walk based algorithm as a scalable graph filter. This forms the basis for our scalable deep clustering algorithm, RwSL, where through a self-supervised mini-batch training mechanism, we simultaneously optimize a deep neural network for sample-cluster assignment distribution and an autoencoder for a clustering-oriented embedding. Using 6 real-world datasets and 6 clustering metrics, we show that RwSL achieved improved results over several recent baselines. Most notably, we show that RwSL, unlike all other deep clustering frameworks, can continue to scale beyond graphs with more than one million nodes, i.e., handle web-scale. We also demonstrate how RwSL could perform node clustering on a graph with 1.8 billion edges using only a single GPU.

Automatic term extraction (ATE) is a Natural Language Processing (NLP) task that eases the effort of manually identifying terms from domain-specific corpora by providing a list of candidate terms. As units of knowledge in a specific field of expertise, extracted terms are not only beneficial for several terminographical tasks, but also support and improve several complex downstream tasks, e.g., information retrieval, machine translation, topic detection, and sentiment analysis. ATE systems, along with annotated datasets, have been studied and developed widely for decades, but recently we observed a surge in novel neural systems for the task at hand. Despite a large amount of new research on ATE, systematic survey studies covering novel neural approaches are lacking. We present a comprehensive survey of deep learning-based approaches to ATE, with a focus on Transformer-based neural models. The study also offers a comparison between these systems and previous ATE approaches, which were based on feature engineering and non-neural supervised learning algorithms.

Every day we are exposed to many images, only a few of which are remembered, while most of them we tend to forget. Though the human cognitive system has an enormous storage capacity [1,2], it may only be able to store some images as detailed as they are. Few images are remembered in great detail, even fewer in minor details, and the remainder is quickly forgotten [3]. Natural scenery photos, for example, are less likely to be remembered than images of animals, vehicles, and people [4]. According to previous research, images are consistently memorable to different viewers [5] and some images have better memorability than others. They also showed that memorability is an intrinsic and measurable property of an image. When we discuss memorability as a measurable property, the question of an artificial system successfully predicting the image memorability score comes along. Previous works done in the domain of image memorability can be grouped into three categories - understanding features that affect image memorability, Prediction of images' memorability scores, and modifying images' memorability.

Data processing tasks over graphs couple the data residing over the nodes with the topology through graph signal processing tools. Graph filters are one such prominent tool, having been used in applications such as denoising, interpolation, and classification. However, they are mainly used on fixed graphs although many networks grow in practice, with nodes continually attaching to the topology. Re-training the filter every time a new node attaches is computationally demanding; hence an online learning solution that adapts to the evolving graph is needed. We propose an online update of the filter, based on the principles of online machine learning. To update the filter, we perform online gradient descent, which has a provable regret bound with respect to the filter computed offline. We show the performance of our method for signal interpolation at the incoming nodes. Numerical results on synthetic and graph-based recommender systems show that the proposed approach compares well to the offline baseline filter while outperforming competitive approaches. These findings lay the foundation for efficient filtering over expanding graphs.

In this work, we consider the problem of minimizing the sum of Moreau envelopes of given functions, which has previously appeared in the context of meta-learning and personalized federated learning. In contrast to the existing theory that requires running subsolvers until a certain precision is reached, we only assume that a finite number of gradient steps is taken at each iteration. As a special case, our theory allows us to show the convergence of First-Order Model-Agnostic Meta-Learning (FO-MAML) to the vicinity of a solution of Moreau objective. We also study a more general family of first-order algorithms that can be viewed as a generalization of FO-MAML. Our main theoretical achievement is a theoretical improvement upon the inexact SGD framework. In particular, our perturbed-iterate analysis allows for tighter guarantees that improve the dependency on the problem's conditioning. In contrast to the related work on meta-learning, ours does not require any assumptions on the Hessian smoothness, and can leverage smoothness and convexity of the reformulation based on Moreau envelopes. Furthermore, to fill the gaps in the comparison of FO-MAML to the Implicit MAML (iMAML), we show that the objective of iMAML is neither smooth nor convex, implying that it has no convergence guarantees based on the existing theory.

We describe the system used by our team for the VoxCeleb Speaker Recognition Challenge 2022 (VoxSRC 2022) in the speaker diarization track. Our solution was designed around a new combination of voice activity detection algorithms that uses the strengths of several systems. We introduce a novel multi stream approach with a decision protocol based on classifiers entropy. We called this method a multi-stream voice activity detection and used it with standard baseline diarization embeddings, clustering and resegmentation. With this work, we successfully demonstrated that using a strong baseline and working only on voice activity detection, one can achieved close to state-of-theart results.

In autonomous driving, 3D object detection based on multi-modal data has become an indispensable approach when facing complex environments around the vehicle. During multi-modal detection, LiDAR and camera are simultaneously applied for capturing and modeling. However, due to the intrinsic discrepancies between the LiDAR point and camera image, the fusion of the data for object detection encounters a series of problems. Most multi-modal detection methods perform even worse than LiDAR-only methods. In this investigation, we propose a method named PTA-Det to improve the performance of multi-modal detection. Accompanied by PTA-Det, a Pseudo Point Cloud Generation Network is proposed, which can convert image information including texture and semantic features by pseudo points. Thereafter, through a transformer-based Point Fusion Transition (PFT) module, the features of LiDAR points and pseudo points from image can be deeply fused under a unified point-based representation. The combination of these modules can conquer the major obstacle in feature fusion across modalities and realizes a complementary and discriminative representation for proposal generation. Extensive experiments on the KITTI dataset show the PTA-Det achieves a competitive result and support its effectiveness.