Receive a weekly summary and discussion of the top papers of the week by leading researchers in the field.

General General

Vehicle and Person Re-Identification With Support Neighbor Loss.

In IEEE transactions on neural networks and learning systems

One of the key tasks for an intelligent visual surveillance system is to automatically re-identify objects of interest, e.g., persons or vehicles, from nonoverlapping camera views. This demand incurs the vast investigation of person re-identification (re-ID) and vehicle re-ID techniques, especially those deep learning-based ones. While most recent algorithms focus on designing new convolutional neural networks, less attention is paid to the loss functions, which are of vital roles as well. Triplet loss and softmax loss are the two losses that are extensively used, both of which, however, have limitations. Triplet loss optimizes the model to produce features with which samples from the same class have higher similarity than those from different classes. The problem of triplet loss is that the number of triplets to be constructed grows cubically with training samples, which causes scalability issue, unstable performance, and slow convergence. Softmax loss has favorable scalable property and is widely used for large-scale classification problems. However, since Softmax loss only aims to separate well training classes, its performance for re-ID tasks is not desirable because the model is tested to measure the similarity of samples from unseen classes. We propose the support neighbor (SN) loss, which avoids the limitations of the abovementioned two losses. Unlike triplet loss that is calculated based on triplets, SN loss is derived from K-nearest neighbors (SNs) of anchor samples. The SNs of an anchor are unique, containing more valuable contextual information and neighborhood structure of the anchor, and thus contribute to more stable performance and reliable embedding from image space to feature space. Based on the SNs, a softmax-like separation term and a squeeze term are proposed, which encourage interclass separation and intraclass compactness, respectively. Experiments show that SN loss surpasses triplet and softmax losses with the same backbone network and reaches the state-of-the-art performance for both person and vehicle re-ID using a ResNet50 backbone when combined with training tricks.

Li Kai, Ding Zhengming, Li Kunpeng, Zhang Yulun, Fu Yun


General General

A Review of Single-Source Deep Unsupervised Visual Domain Adaptation.

In IEEE transactions on neural networks and learning systems

Large-scale labeled training datasets have enabled deep neural networks to excel across a wide range of benchmark vision tasks. However, in many applications, it is prohibitively expensive and time-consuming to obtain large quantities of labeled data. To cope with limited labeled training data, many have attempted to directly apply models trained on a large-scale labeled source domain to another sparsely labeled or unlabeled target domain. Unfortunately, direct transfer across domains often performs poorly due to the presence of domain shift or dataset bias. Domain adaptation (DA) is a machine learning paradigm that aims to learn a model from a source domain that can perform well on a different (but related) target domain. In this article, we review the latest single-source deep unsupervised DA methods focused on visual tasks and discuss new perspectives for future research. We begin with the definitions of different DA strategies and the descriptions of existing benchmark datasets. We then summarize and compare different categories of single-source unsupervised DA methods, including discrepancy-based methods, adversarial discriminative methods, adversarial generative methods, and self-supervision-based methods. Finally, we discuss future research directions with challenges and possible solutions.

Zhao Sicheng, Yue Xiangyu, Zhang Shanghang, Li Bo, Zhao Han, Wu Bichen, Krishna Ravi, Gonzalez Joseph E, Sangiovanni-Vincentelli Alberto L, Seshia Sanjit A, Keutzer Kurt


General General

Challenges in Evaluating Interactive Visual Machine Learning Systems.

In IEEE computer graphics and applications

In interactive visual machine learning (IVML), humans and machine learning algorithms collaborate to achieve tasks mediated by interactive visual interfaces. This human-in-the-loop approach to machine learning brings forth not only numerous intelligibility, trust, and usability issues, but also many open questions with respect to the evaluation of the IVML system, both as separate components, and as a holistic entity that includes both human and machine intelligence. This article describes the challenges and research gaps identified in an IEEE VIS workshop on the evaluation of IVML systems.

Boukhelifa N, Bezerianos A, Chang R, Collins C, Drucker S, Endert A, Hullman J, North C, Sedlmair M, Rhyne Theresa-Marie

Radiology Radiology

Influence of a novel deep-learning based reconstruction software on the objective and subjective image quality in low-dose abdominal computed tomography.

In The British journal of radiology

OBJECTIVES : Modern reconstruction and post-processing software aims at reducing image noise in CT images, potentially allowing for a reduction of the employed radiation exposure. This study aimed at assessing the influence of a novel deep-learning based software on the subjective and objective image quality compared to two traditional methods [filtered back-projection (FBP), iterative reconstruction (IR)].

METHODS : In this institutional review board-approved retrospective study, abdominal low-dose CT images of 27 patients (mean age 38 ± 12 years, volumetric CT dose index 2.9 ± 1.8 mGy) were reconstructed with IR, FBP and, furthermore, post-processed using a novel software. For the three reconstructions, qualitative and quantitative image quality was evaluated by means of CT numbers, noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) in six different ROIs. Additionally, the reconstructions were compared using SNR, peak SNR, root mean square error and mean absolute error to assess structural differences.

RESULTS : On average, CT numbers varied within 1 Hounsfield unit (HU) for the three assessed methods in the assessed ROIs. In soft tissue, image noise was up to 42% lower compared to FBP and up to 27% lower to IR when applying the novel software. Consequently, SNR and CNR were highest with the novel software. For both IR and the novel software, subjective image quality was equal but higher than the image quality of FBP-images.

CONCLUSION : The assessed software reduces image noise while maintaining image information, even in comparison to IR, allowing for a potential dose reduction of approximately 20% in abdominal CT imaging.

ADVANCES IN KNOWLEDGE : The assessed software reduces image noise by up to 27% compared to IR and 48% compared to FBP while maintaining the image information.The reduced image noise allows for a potential dose reduction of approximately 20% in abdominal imaging.

Steuwe Andrea, Weber Marie, Bethge Oliver Thomas, Rademacher Christin, Boschheidgen Matthias, Sawicki Lino Morris, Antoch Gerald, Aissa Joel


General General

Dust-Sized High-Power-Density Photovoltaic Cells on Si and SOI Substrates for Wafer-Level-Packaged Small Edge Computers.

In Advanced materials (Deerfield Beach, Fla.)

Advancement in microelectronics technology enables autonomous edge computing platforms in the size of a dust mote (<1 mm), bringing efficient and low-cost artificial intelligence close to the end user and Internet-of-Things (IoT) applications. The key challenge for these compact high-performance edge computers is the integration of a power source that satisfies the high-power-density requirement and does not increase the complexity and cost of the packaging. Here, it is shown that dust-sized III-V photovoltaic (PV) cells grown on Si and silicon-on-insulator (SOI) substrates can be integrated using a wafer-level-packaging process and achieve higher power density than all prior micro-PVs on Si and SOI substrates. The high-throughput heterogeneous integration unlocks the potential of large-scale manufacturing of these integrated systems with low cost for IoT applications. The negative effect of crystallographic defects in the heteroepitaxial materials on PV performance diminishes at high power density. Simultaneous power delivery and data transmission to the dust mote with heteroepitaxially grown PV are also demonstrated using hand-held illumination sources.

Li Ning, Han Kevin, Spratt William, Bedell Stephen, Ren Jinhan, Gunawan Oki, Ott John, Hopstaken Marinus, Cabral Cyril, Libsch Frank, Subramanian Chitra, Shahidi Ghavam, Sadana Devendra


compound semiconductors, edge computing, heterogeneous integration, photovoltaics, wafer-level packaging

General General

Privacy-preserving Collaborative Training for Medical Image Analysis Based on Multi-Blockchain.

In Combinatorial chemistry & high throughput screening

BACKGROUND : As artificial intelligence and big data analysis develop rapidly, data privacy, especially patient medical data privacy, is getting more and more attention.

OBJECTIVE : To strengthen the protection of private data while ensuring the model training process, this article introduces a multi-Blockchain-based decentralized collaborative machine learning training method for medical image analysis. In this way, researchers from different medical institutions are able to collaborate to train models without exchanging sensitive patient data.

METHOD : Partial parameter update method is applied to prevent indirect privacy leakage during model propagation. With the peer-to-peer communication in the multi-Blockchain system, a machine learning task can leverage auxiliary information from another similar task in another Blockchain. In addition, after the collaborative training process, personalized models of different medical institutions will be trained.

RESULTS : The experimental results show that our method achieves similar performance with the centralized model-training method by collecting data sets of all participants and prevents private data leakage at the same time. Transferring auxiliary information from similar task on another Blockchain has also been proven to effectively accelerate model convergence and improve model accuracy, especially in the scenario of absence of data. Personalization training process further improves model performance.

CONCLUSION : Our approach can effectively help researchers from different organizations to achieve collaborative training without disclosing their private data.

Zhang Wanlu, Wang Qigang, Li Mei


Blockchain, deep learning, distributed training, medical image analysis., personalized learning, transfer learning