About

Tong Zhang is a professor of Computer Science at the University of Illinois Urbana-Champaign. His research interests include machine learning algorithms and theory, statistical methods for big data and their applications. He has been a professor at the Hong Kong University of Science and Technology, Rutgers University, and has worked at IBM, Yahoo, Google, Baidu, and Tencent.

Research topics

• Political Science
• Social Science
• Sociology
• Economics
• Chemistry
• Business
• Public economics
• Environmental science
• Economic growth
• Positive economics
• Monetary economics
• Financial economics
• Environmental chemistry
• Environmental engineering
• Pulp and paper industry

Selected publications

• Methods for Studying Union Effects: A Review and Comparative Analysis of Empirical Industrial Relations Literature

  Industrial Relations Journal · 2026-01-20

  articleOpen accessCorresponding

  ABSTRACT This paper reviews methodological developments in Industrial Relations (IR) research on union effects from 1990 to 2023, based on 511 studies in six leading IR journals in the United States, the United Kingdom, Canada, and Australia. We find that institutional contexts shape methodological choices over time and note a general shift from descriptive analyses to advanced quantitative approaches, such as fixed effects, instrumental variables, and quasi‐experimental designs. At the same time, however, qualitative and mixed methods remain central to the field. The paper further shows that research agendas have expanded from focusing on wages, collective bargaining, and workplace HR policies to include political and societal outcomes. Finally, we situate IR studies of union effects relative to adjacent disciplines: economics, sociology, political science, psychology, and management. Bibliometric analysis reveals close ties between IR and economics, as well as shared research interests with sociology and political science. The findings suggest IR has increased its methodological sophistication and maintained a pluralist identity – with both features informed by changing research priorities, national institutions, and ongoing dialogue with adjacent disciplines.

  PublisherOA PDFDOI

• Between diversity and meritocracy: employer and skilled immigrant perspectives from the Canadian context

  Equality Diversity and Inclusion An International Journal · 2025-01-22 · 4 citations

  article

  Purpose This study aims to empirically investigate and extend the diversity-meritocracy paradox outlined by Konrad et al. (2021) using skilled immigrants in Canada as a case study. Despite their significance in knowledge-based economies, immigrant voices are often marginalized in diversity, equity and inclusion (DEI) literature and management research. By focusing on skilled immigrants, who embody both diversity and meritocratic principles, this research addresses this gap. Through semi-structured interviews, we examine whether diversity and meritocracy are perceived as contradictory or complementary for skilled immigrants. Our findings not only contribute to theoretical understanding but also offer practical insights into the complexities of diversity and meritocracy in contemporary organizations. Design/methodology/approach This study utilizes qualitative, semi-structured, interviews and focus groups to gather data from both employers and skilled immigrants. Thematic analysis, guided by Braun and Clarke (2006), is employed to analyze the data. Participants include skilled immigrants and human resource (HR) professionals/managers. Data are collected through interviews and focus groups conducted between December 2018 and February 2020 in person and via video-conferencing. Findings This study unveils a discrepancy in perceptions between employers and skilled immigrants on DEI in Canada’s labor market. While employers prioritize meritocracy, emphasizing Canadian qualifications and experience, immigrants feel undervalued, encountering barriers due to cultural differences. Employers focus on past work experience over credentials, using behavioral interviews and proficiency tests for assessment. However, immigrants often perceive the selection process as opaque, and encounter explicit preferences for Canadian education and experience, which they view as discriminatory. Challenges in onboarding, training and workplace culture further exacerbate their experiences. These findings highlight the nuanced dynamics between meritocracy and diversity, underscoring the need for systemic change. Originality/value Despite employers’ claims of valuing diversity, our findings reveal a preference for “Canadian-ness” over immigrants' international expertise, perpetuating systemic barriers. Employers prioritize meritocracy but often conflate it with cultural conformity, hindering immigrant integration. Our analysis underscores the disconnection between organizational rhetoric and practices, urging a reconceptualization of diversity and inclusion policies. To foster truly inclusive workplaces, both surface-level and deep-level diversity must be considered. Policy interventions and enhanced intercultural competence are essential for leveraging the talents of skilled immigrants and promoting equitable employment practices.

  PublisherDOI

• Voice without Representation: Worker Voice in China’s Networked Public Sphere

  Industrial and Labor Relations Review · 2025-06-13 · 4 citations

  articleSenior author

  Does worker voice on social media empower individuals to advocate for better working conditions when traditional voice mechanisms are absent? This study examined Chinese tech workers’ use of social media to resist overwork culture. The analysis of social media data, interviews, and news published by state-affiliated outlets shows worker voice on social media raised public awareness of overtime issues and increased state-run media coverage of overtime issues, culminating in a landmark ruling in China’s Supreme People’s Court against exploitative practices. However, online debates on the legitimacy of these overtime practices failed to build a lasting consensus in support of workers. Inconsistent enforcement of labor laws and administrative directives further weakened the protection of tech workers. Ultimately, while social media amplified worker voice, it failed to drive meaningful workplace improvements in a context in which workers lacked associational and institutional power.

  PublisherDOI

• Adaptive Optimization of Multi-scene Image Feature Extraction for Metering Asset Fragmentation Operations and Maintenance

  2025-08-15

  article

  In the fragmented operation and maintenance scenarios of electric power metering assets, the construction of metering boxes, the rotation of old equipment, marketing census, on-site inspections and inspections, etc., put forward very high requirements for image feature extraction technology. These scenarios not only have significant differences in work objectives, archiving needs to accurately record the details of the appearance of the equipment, inspection needs to quickly identify potential safety hazards, marketing census needs to balance the accuracy and efficiency, but also face the complexity of the field environment, uneven data distribution and other issues, resulting in the accuracy, efficiency and stability of image feature extraction is difficult to achieve synergistic optimization. Traditional fixed models such as Mask R-CNN can ensure a certain degree of accuracy, but the delay is too high in real-time scenarios; MobileNet-SSD is lightweight but cannot meet the demand for fine-grained feature extraction, and both of them have a significant decrease in stability under the interference of backlighting, motion blur, and so on. Aiming at this situation, this paper proposes an adaptive feature extraction framework of “Scene Sensing Dynamic Adaptation - Joint Optimization”. The framework realizes automatic scene identification through the “demand interference” two-dimensional quantitative model, adjusts the extraction strategy on demand based on the two-branch dynamic network, and combines meta-learning and transfer learning to complete cross-scene knowledge reuse and anti-interference optimization. The experimental results show that in the measurement box archiving scenario, the feature extraction accuracy of “one box with multiple signatures” reaches 95.3 %, which is 12.7 % higher than that of traditional Mask R-CNN; in the on-site inspection scenario, the inference time of a single image is reduced to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 9 m s}$</tex>, which is <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4 6. 2 \%}$</tex> shorter than that of MobileNetSSD; and under different interference scenarios such as backlight and motion blur, the inference time of a single image is reduced to 89 ms, which is 46.2 % shorter than that of MobileNet-SSD. In different interference scenarios such as backlight and motion blur, the accuracy fluctuation is controlled within 4.8 %, which is significantly better than the fixed model. Meanwhile, in small sample scenarios (e.g., only 1,000 training images in the archiving scenario), the feature extraction accuracy of the framework reaches 91.2 %, which is close to the training effect of the traditional model with 5,000 samples, and significantly reduces the cost of data annotation. This study not only provides a dynamically adapted image feature extraction scheme for the fragmented operation and maintenance of power metering assets, but also provides a referable technical path for the intelligent optimization of visual feature extraction in multiple scenarios.

  PublisherDOI

• Long-Term Earnings Differentials Between Standard and Alternative Credential Recognition Pathways

  Academy of Management Proceedings · 2025-07-01

  article1st authorCorresponding

  This study links a representative immigration survey with longitudinal administrative records to analyze the long-term earnings of Canadian immigrants across varied credential recognition pathways: no credentials (baseline), Canadian credentials, recognized foreign credentials, and unrecognized foreign credentials. Growth curve modeling and propensity score matching reveal that immigrants with Canadian credentials earn significantly more than the baseline, those with recognized foreign credentials earn less, and those with unrecognized credentials have comparable earnings. These disparities persist over time. The results are discussed in terms of traditional human capital and monopoly theories of licensing as well as home-biases and taste-based discrimination.

  PublisherDOI

• Biomimetic Multiscale‐Structured Biomass Graphene/Polyurethane Sponge Composite for Flexible Pressure Sensors and Intelligent Cushioning Materials

  Advanced Science · 2025-11-25 · 2 citations

  articleOpen access

  Abstract The development of furniture cushioning materials that combine excellent mechanical properties with sensing capabilities is essential for non‐intrusive, long‐term health monitoring. This study presents a multifunctional conductive sponge (MAPU) that synergistically integrates the macroscale mechanical support of a polyurethane (PU) sponge with the microscale sensing characteristics of aerogels through bionic multiscale structural design. Biomass‐derived graphite nanoflakes serve as the conductive units to in situ construct a 3D interpenetrating aerogel network on the PU sponge skeleton. This unique heterogeneous structure retains the flexibility and elasticity of PU sponge while providing exceptional piezoresistive sensing performance, including high sensitivity (0.821 kPa −1 in the 0–53 kPa range), a wide response range (up to 242 kPa), fast response time (≤ 50 ms), and outstanding cycling stability (&gt; 30 000 cycles). Equally important, MAPU also demonstrates washability, flame retardancy, breathability, and sound absorption, making it practical for household applications. An intelligent mattress composed of a MAPU sensor array enables real‐time monitoring and precise recognition of sleep postures, along with bedsore risk alerts. This work offers a high‐performance, multifunctional, and high‐safety core material solution for advanced smart home technologies and continuous health monitoring systems.

  PublisherOA PDFDOI

• Irisin improves acute kidney injury induced by ischemia-reperfusion through targeting energy metabolism reprogramming

  International Journal of Biological Macromolecules · 2025-08-25 · 6 citations

  articleOpen access

  The kidney possesses a highly complex reabsorption function, with tubular epithelial cells (TECs) being the primary functional units. The TECs mainly rely on fatty acid oxidation (FAO) for energy supply. FAO is hindered during ischemia-reperfusion (I/R)-induced acute kidney injury (AKI), leading to impairments in oxidative phosphorylation. The energy supply to renal TECs then shifts from FAO to glycolysis. Impaired fatty acid uptake and oxidation, coupled with sustained increases in glycolysis, can lead to irreversible damage or failure of TEC regeneration. This causes TEC atrophy, inadequate adaptive repair, progression to chronic kidney disease (CKD), and fibrosis. Irisin, a protein secreted by skeletal muscle, is a factor in regulating energy homeostasis. The beneficial effects of exercise on diabetic kidney disease are primarily mediated through its actions. In vitro and in vivo models of our experiments show that irisin can correct the energy metabolic disorders of renal TECs following I/R, increase FAO, reduce glycolysis, and improve I/R-induced AKI. Target points of irisin were explored to elucidate the specific mechanisms of action. We found that the integrin αVβ5 receptor, a target of irisin, is also expressed in TECs and is upregulated following hypoxia/reoxygenation (H/R). Bioinformatics showed significant differences in the AMPK (PRKAA2) and mTOR genes between AKI and normal mice and identified protein interaction among FNDC5/irisin, AMPK, and mTOR. Examination of the effects of irisin on AMPK, mTOR, and the mTOR downstream target HIF-1α showed that pre-treatment with irisin increased the phosphorylation of AMPK in H/R-treated HK-2 cells while simultaneously suppressing mTOR phosphorylation and reducing HIF-1α expression. However, pretreatment with an anti-integrin αVβ5 antibody inhibited the regulatory effects of irisin on AMPK/mTOR phosphorylation and HIF-1α protein levels in H/R-treated HK-2 cells. Given their recognized regulatory roles in energy metabolism, particularly in FAO and glycolysis, our experiments investigating the effects of AMPK and mTOR modulation demonstrated that suppression of AMPK or stimulation of mTOR impeded Irisin's ability to modulate these processes in response to I/R, as well as its ameliorative impact on AKI. Therefore, we concluded that irisin exerts its beneficial effects by binding to the integrin αVβ5 receptor on the membrane of renal TECs, thereby modulating the AMPK/mTOR pathway to improve FAO and reduce glycolysis, ultimately ameliorating I/R-induced AKI, with HIF-1α also participating in this regulatory process. ( Graphical Abstract ). • This study demonstrates, for the first time, that irisin increases fatty acid oxidation and reduces glycolysis in renal tubular epithelial cells induced by ischemia-reperfusion. • The study reveals that the target receptor of irisin, integrin αVβ5, is expressed in renal tubular epithelial cells and is upregulated following hypoxia/reoxygenation (H/R). • This study finds that irisin can regulate energy metabolism by modulating the AMPK/mTOR pathway through interaction with the integrin αVβ5 receptor on renal tubular epithelial cells, thereby improving ischemia-reperfusion-induced acute kidney injury.

  PublisherDOI

• Privilege and Precarity: Migration Journeys of Former International Students in Canada Through the Lens of the Aspirations and Capabilities Framework

  Journal of Immigrant & Refugee Studies · 2025-11-27

  article
  PublisherDOI

• Enhanced Simulated Bifurcation for MIMO Detection

  2025-06-05

  article

  Massive multiple-input, multiple-output (MIMO) systems enhance spectral efficiency, coverage, and reliability, making them essential for next-generation wireless networks. However, maximum likelihood (ML) detection in MIMO remains an NP-hard problem. This work explores CMOS-friendly Ising solvers, particularly simulated bifurcation, as alternative MIMO detection methods. We introduce a novel temperature schedule for ballistic simulated bifurcation (bSB), developing an enhanced variant, bSBG, to improve bit error rate (BER) performance. bSBG achieves faster convergence and mitigates the error floor issues present in bSB. Experimental results show that both bSB and bSBG outperform linear minimum mean squared error (LMMSE) and K-best detection by up to 12 dB and 4 dB, respectively, at a BER of 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−3</sup> for a 128×128 antenna configuration, while bSBG eliminates the error floor for 16×16 and 32×32 configurations. Additionally, bSBG and bSB achieve speedups of 7.2× and 9.2× over K-best, respectively, while running 7.5× slower than LMMSE but providing a 13 dB gain at a BER of 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−4</sup>. These results highlight bSBG as a promising candidate for efficient MIMO detection in large-scale systems.

  PublisherDOI

• Astrocytes expressing mutant hnRNPA1 induce non-cell-autonomous motor neuron death

  Brain Research Bulletin · 2025-08-24

  articleOpen access

  Pathogenic mutation of heterogeneous nuclear ribonucleoprotein A1 ( hnRNPA1) is causative to amyotrophic lateral sclerosis (ALS). Neuron death resulting from pathogenic hnRNPA1 may not require its presence across all pertinent cells types, including neurons, glia, and muscles. Rather, the exclusive presence of pathogenic hnRNPA1 in a specific cell type, such as astrocytes, may suffice to substantially alter cellular functions. Consequently, this alteration initiates abnormal interaction within intricate neuron-glia networks, culminating in non-cell-autonomous motor neuron death. To investigate the pivotal role of non-cell-autonomous neuron death in hnRNPA1-associated ALS, we developed transgenic rats overexpressing mutant hnRNPA1 in specifically astrocytes. The confined overexpression of pathogenic hnRNPA1 in astrocytes instigated a sequence of events resulting in motor neuron death and subsequent muscle atrophy. These findings underscore the critical, non-cell-autonomous contribution of astrocytes to hnRNPA1-induced neurodegeneration in ALS, and point toward astrocytic pathways as potential therapeutic targets. • Pathogenic mutation of hnRNPA1 is linked to ALS, but neuron death may not require its presence in all cell types • Transgenic rats overexpressing mutant hnRNPA1 selectively in astrocytes developed motor neuron death and muscle atrophy • Astrocytes are integral to hnRNPA1-induced neurodegeneration in ALS

  PublisherDOI

Frequent coauthors

• Yong Guo

  7 shared

• Rafael Gómez

  7 shared

• Lorenzo Frangi

  Université du Québec à Montréal

  6 shared

• Alex Bryson

  6 shared

• Wenju Lu

  Guangzhou Medical University

  6 shared

• Zhifeng Liu

  Hefei University of Technology

  6 shared

• Yuxia Chen

  Anhui Agricultural University

  6 shared

• Rupa Banerjee

  Toronto Metropolitan University

  6 shared

Labs

• Tong Zhang's LabPI

  Develops principled methods for machine learning, including algorithms with theoretical guarantees, mathematical understanding of modern learning systems, and rigorous foundations for emerging applications in generative AI, agents, and robotics.

Education

• B.A., Computer Science and Mathematics

  Cornell University

  1994

• M.S., Computer Science

  Stanford University

  1996

• Ph.D., Computer Science

  Stanford University

  1999

Awards & honors

• Fellow of ASA
• Fellow of IEEE
• Fellow of IMS