Search

Monolithic large language models (LLMs) like GPT-4 have paved the way for modern generative AI applications. Training, serving, and maintaining monolithic LLMs at scale, however, remains prohibitively expensive and challenging. The disproportionate increase in compute-to-memory ratio of modern AI accelerators have created a memory wall, necessitating new methods to deploy AI. Composition of Experts (CoE) is an alternative modular approach that lowers the cost and complexity of training and serving. However, this approach presents two key challenges when using conventional hardware: (1) without fused operations, smaller models have lower operational intensity, which makes high utilization more challenging to achieve; and (2) hosting a large number of models can be either prohibitively expensive or slow when dynamically switching between them. In this paper, we describe how combining CoE, streaming dataflow, and a three-tier memory system scales the AI memory wall. We describe Samba-CoE, a CoE system with 150 experts and a trillion total parameters. We deploy Samba-CoE on the SambaNova SN40L Reconfigurable Dataflow Unit (RDU) - a commercial dataflow accelerator architecture that has been co-designed for enterprise inference and training applications. The chip introduces a new three-tier memory system with on-chip distributed SRAM, on-package HBM, and off-package DDR DRAM. A dedicated inter-RDU network enables scaling up and out over multiple sockets. We demonstrate speedups ranging from 2x to 13x on various benchmarks running on eight RDU sockets compared with an unfused baseline. We show that for CoE inference deployments, the 8-socket RDU Node reduces machine footprint by up to 19x, speeds up model switching time by 15x to 31x, and achieves an overall speedup of 3.7x over a DGX H100 and 6.6x over a DGX A100.

Multiple object detection and pose estimation are vital computer vision tasks. The latter relates to the former as a downstream problem in applications such as robotics and autonomous driving. However, due to the high complexity of both tasks, existing methods generally treat them independently, which is sub-optimal. We propose simultaneous neural modeling of both using monocular vision and 3D model infusion. Our Simultaneous Multiple Object detection and Pose Estimation network (SMOPE-Net) is an end-to-end trainable multitasking network with a composite loss that also provides the advantages of anchor-free detections for efficient downstream pose estimation. To enable the annotation of training data for our learning objective, we develop a Twin-Space object labeling method and demonstrate its correctness analytically and empirically. Using the labeling method, we provide the KITTI-6DoF dataset with $\sim7.5$K annotated frames. Extensive experiments on KITTI-6DoF and the popular LineMod datasets show a consistent performance gain with SMOPE-Net over existing pose estimation methods. Here are links to our proposed SMOPE-Net, KITTI-6DoF dataset, and LabelImg3D labeling tool.

It is a well known phenomenon in optics that spectroscopic resolution of a diffraction grating is much better compared to an interference device having just two slits, as in the Young's famous double-slit experiment. On the other hand, it is well known that a classical superconducting quantum interference device (SQUID) is analogous to the optical double-slit experiment. Here we report experiments and present a model describing a superconducting analogue to the diffraction grating, namely an array of superconducting islands positioned on a topological insulator (TI) film Bi0.8Sb1.2Te3. In the limit of extremely weak field, of the order of one vortex per the entire array, such devices exhibit a critical current peak that is much sharper than the analogous peak of an ordinary SQUID. Because of this, such arrays can be used as sensitive absolute magnetic field sensors. An important finding is that, due to the inherent asymmetry of such arrays, the device also acts as a superconducting diode., Comment: 8 pages, 6 figures

Influence Maximization (IM), which aims to select a set of users from a social network to maximize the expected number of influenced users, is an evergreen hot research topic. Its research outcomes significantly impact real-world applications such as business marketing. The booming location-based network platforms of the last decade appeal to the researchers embedding the location information into traditional IM research. In this survey, we provide a comprehensive review of the existing location-driven IM studies from the perspective of the following key aspects: (1) a review of the application scenarios of these works, (2) the diffusion models to evaluate the influence propagation, and (3) a comprehensive study of the approaches to deal with the location-driven IM problems together with a particular focus on the accelerating techniques. In the end, we draw prospects into the research directions in future IM research.

The goal of Knowledge Tracing (KT) is to estimate how well students have mastered a concept based on their historical learning of related exercises. The benefit of knowledge tracing is that students' learning plans can be better organised and adjusted, and interventions can be made when necessary. With the recent rise of deep learning, Deep Knowledge Tracing (DKT) has utilised Recurrent Neural Networks (RNNs) to accomplish this task with some success. Other works have attempted to introduce Graph Neural Networks (GNNs) and redefine the task accordingly to achieve significant improvements. However, these efforts suffer from at least one of the following drawbacks: 1) they pay too much attention to details of the nodes rather than to high-level semantic information; 2) they struggle to effectively establish spatial associations and complex structures of the nodes; and 3) they represent either concepts or exercises only, without integrating them. Inspired by recent advances in self-supervised learning, we propose a Bi-Graph Contrastive Learning based Knowledge Tracing (Bi-CLKT) to address these limitations. Specifically, we design a two-layer contrastive learning scheme based on an "exercise-to-exercise" (E2E) relational subgraph. It involves node-level contrastive learning of subgraphs to obtain discriminative representations of exercises, and graph-level contrastive learning to obtain discriminative representations of concepts. Moreover, we designed a joint contrastive loss to obtain better representations and hence better prediction performance. Also, we explored two different variants, using RNN and memory-augmented neural networks as the prediction layer for comparison to obtain better representations of exercises and concepts respectively. Extensive experiments on four real-world datasets show that the proposed Bi-CLKT and its variants outperform other baseline models., Comment: 12pages, 2 figures

With the rapid development of 5th Generation Mobile Communication Technology (5G), the diverse forms of collaboration and extensive data in academic social networks constructed by 5G papers make the management and analysis of academic social networks increasingly challenging. Despite the particular success achieved by representation learning in analyzing academic and social networks, most present presentation learning models focus on maintaining the first-order and second-order similarity of nodes. They rarely possess similar structural characteristics of spatial independence in the network. This paper proposes a Low-order Network representation Learning Model (LNLM) based on Non-negative Matrix Factorization (NMF) to solve these problems. The model uses the random walk method to extract low-order features of nodes and map multiple components to a low-dimensional space, effectively maintaining the internal correlation between members. This paper verifies the performance of this model, conducts comparative experiments on four test datasets and four real network datasets through downstream tasks such as multi-label classification, clustering, and link prediction. Comparing eight mainstream network representation learning models shows that the proposed model can significantly improve the detection efficiency and learning methods and effectively extract local and low-order features of the network., Comment: 38 pages, 9 figures

The outbreak of the novel Coronavirus Disease 2019 (COVID-19) has lasted for nearly two years and caused unprecedented impacts on people's daily life around the world. Even worse, the emergence of the COVID-19 Delta variant once again puts the world in danger. Fortunately, many countries and companies have started to develop coronavirus vaccines since the beginning of this disaster. Till now, more than 20 vaccines have been approved by the World Health Organization (WHO), bringing light to people besieged by the pandemic. The promotion of COVID-19 vaccination around the world also brings a lot of discussions on social media about different aspects of vaccines, such as efficacy and security. However, there does not exist much research work to systematically analyze public opinion towards COVID-19 vaccines. In this study, we conduct an in-depth analysis of tweets related to the coronavirus vaccine on Twitter to understand the trending topics and their corresponding sentimental polarities regarding the country and vaccine levels. The results show that a majority of people are confident in the effectiveness of vaccines and are willing to get vaccinated. In contrast, the negative tweets are often associated with the complaints of vaccine shortages, side effects after injections and possible death after being vaccinated. Overall, this study exploits popular NLP and topic modeling methods to mine people's opinions on the COVID-19 vaccines on social media and to analyse and visualise them objectively. Our findings can improve the readability of the noisy information on social media and provide effective data support for the government and policy makers.

In recent years, Graph Convolutional Networks (GCNs) have achieved great success in learning from graph-structured data. With the growing tendency of graph nodes and edges, GCN training by single processor cannot meet the demand for time and memory, which led to a boom into distributed GCN training frameworks research. However, existing distributed GCN training frameworks require enormous communication costs between processors since multitudes of dependent nodes and edges information need to be collected and transmitted for GCN training from other processors. To address this issue, we propose a Graph Augmentation based Distributed GCN framework(GAD). In particular, GAD has two main components, GAD-Partition and GAD-Optimizer. We first propose a graph augmentation-based partition (GAD-Partition) that can divide original graph into augmented subgraphs to reduce communication by selecting and storing as few significant nodes of other processors as possible while guaranteeing the accuracy of the training. In addition, we further design a subgraph variance-based importance calculation formula and propose a novel weighted global consensus method, collectively referred to as GAD-Optimizer. This optimizer adaptively reduces the importance of subgraphs with large variances for the purpose of reducing the effect of extra variance introduced by GAD-Partition on distributed GCN training. Extensive experiments on four large-scale real-world datasets demonstrate that our framework significantly reduces the communication overhead (50%), improves the convergence speed (2X) of distributed GCN training, and slight gain in accuracy (0.45%) based on minimal redundancy compared to the state-of-the-art methods.

Effective representation learning is critical for short text clustering due to the sparse, high-dimensional and noise attributes of short text corpus. Existing pre-trained models (e.g., Word2vec and BERT) have greatly improved the expressiveness for short text representations with more condensed, low-dimensional and continuous features compared to the traditional Bag-of-Words (BoW) model. However, these models are trained for general purposes and thus are suboptimal for the short text clustering task. In this paper, we propose two methods to exploit the unsupervised autoencoder (AE) framework to further tune the short text representations based on these pre-trained text models for optimal clustering performance. In our first method Structural Text Network Graph Autoencoder (STN-GAE), we exploit the structural text information among the corpus by constructing a text network, and then adopt graph convolutional network as encoder to fuse the structural features with the pre-trained text features for text representation learning. In our second method Soft Cluster Assignment Autoencoder (SCA-AE), we adopt an extra soft cluster assignment constraint on the latent space of autoencoder to encourage the learned text representations to be more clustering-friendly. We tested two methods on seven popular short text datasets, and the experimental results show that when only using the pre-trained model for short text clustering, BERT performs better than BoW and Word2vec. However, as long as we further tune the pre-trained representations, the proposed method like SCA-AE can greatly increase the clustering performance, and the accuracy improvement compared to use BERT alone could reach as much as 14\%., Comment: To be published in 22nd International Conference on Web Information Systems Engineering (WISE2021)

We report superconducting kinetic inductance memory (SKIM) element, which can be controlled exclusively by the bias supercurrent, without involving magnetic fields and heating elements. The SKIM is non-volatile memory. The device is made of Nb and it can operate reliable up to 2.8 K. The achieved error rate is as low as one in 100000 operations., Comment: 9 pages, 6 figures

We study one dimensional chains of superconducting islands with a particular emphasis on the regime in which every second island is switched into its normal state, thus forming a superconductor-insulator-normal metal (S-I-N) repetition pattern. As is known since Giaever tunneling experiments, tunneling charge transport between a superconductor and a normal metal becomes exponentially suppressed, and zero-bias resistance diverges, as the temperature is reduced and the energy gap of the superconductor grows larger than the thermal energy. Here we demonstrate that this physical phenomenon strongly impacts transport properties of inhomogeneous superconductors made of weakly coupled islands with fluctuating values of the critical temperature. We observe a non-monotonous dependence of the chain resistance on both temperature and magnetic field, with a pronounced resistance peak at temperatures at which some but not all islands are superconducting. We explain this phenomenon by the inhomogeneity of the chains, in which neighboring superconducting islands have slightly different critical temperatures. We argue that the Giaever's resistance divergence can also occur in the zero-temperature limit. Such quantum transition can occur if the magnetic field is tuned such that it suppresses superconductivity in the islands with the weaker critical field, while the islands with stronger energy gap remain superconducting. In such a field, the system acts as a chain of S-I-N junctions., Comment: 12 pages, 8 figures

Deep learning-based Multiple Object Tracking (MOT) currently relies on off-the-shelf detectors for tracking-by-detection.This results in deep models that are detector biased and evaluations that are detector influenced. To resolve this issue, we introduce Deep Motion Modeling Network (DMM-Net) that can estimate multiple objects' motion parameters to perform joint detection and association in an end-to-end manner. DMM-Net models object features over multiple frames and simultaneously infers object classes, visibility, and their motion parameters. These outputs are readily used to update the tracklets for efficient MOT. DMM-Net achieves PR-MOTA score of 12.80 @ 120+ fps for the popular UA-DETRAC challenge, which is better performance and orders of magnitude faster. We also contribute a synthetic large-scale public dataset Omni-MOT for vehicle tracking that provides precise ground-truth annotations to eliminate the detector influence in MOT evaluation. This 14M+ frames dataset is extendable with our public script (Code at Dataset , Dataset Recorder , Omni-MOT Source ). We demonstrate the suitability of Omni-MOT for deep learning with DMMNet and also make the source code of our network public., Comment: 25 pages, 9 figures, has been accepted by the ECCCV 2020