International Conference on Machine Learning
The International Conference on Machine Learning (ICML) is the premier annual academic conference dedicated to advancing the field of machine learning, a core subdiscipline of artificial intelligence that focuses on algorithms and statistical models enabling computers to perform tasks without explicit instructions. Held every summer, typically in July, it convenes thousands of researchers, academics, and industry professionals from around the world to share innovative research through peer-reviewed papers, invited talks, oral presentations, poster sessions, and specialized workshops covering theoretical foundations, practical applications, and emerging challenges in machine learning.[1][2] ICML originated as the First International Workshop on Machine Learning in July 1980 at Carnegie Mellon University in Pittsburgh, Pennsylvania, organized by pioneers Jaime G. Carbonell, Ryszard S. Michalski, and Tom M. Mitchell, marking the beginning of a series that fostered early developments in the field. Subsequent workshops occurred irregularly in 1983, 1985, 1987, 1988, 1989, 1990, 1991, and 1992, before transitioning to an annual conference format starting in 1993, with continuous editions thereafter; the 2025 event was the 42nd, held at the Vancouver Convention Centre from July 13 to 19, 2025.[3][4][1] Renowned for its rigorous double-blind peer review process, ICML maintains a selective acceptance rate, such as 27.5% for the 2024 edition (2,610 papers accepted out of 9,473 submissions), ensuring high-quality contributions that often influence subsequent research and industry innovations. Along with NeurIPS and ICLR, it forms one of the three flagship venues in machine learning, with proceedings published openly in the Proceedings of Machine Learning Research (PMLR) to promote widespread accessibility and impact. The conference also emphasizes diversity, ethics in AI, and interdisciplinary collaboration, reflecting the field's rapid evolution amid growing applications in areas like healthcare, robotics, and natural language processing.[5][6]Overview
Description
The International Conference on Machine Learning (ICML) is the premier international academic conference dedicated to the advancement of machine learning, a core branch of artificial intelligence. Held annually since 1993, following initial workshops starting in 1980, it serves as a leading forum for presenting and discussing cutting-edge research in areas such as statistical learning theory, reinforcement learning, robotics, optimization, and related AI subfields, including applications in machine vision, computational biology, speech recognition, and data science.[7] ICML's core purpose is to convene a diverse community of professionals committed to pushing the boundaries of machine learning, encompassing researchers, practitioners, engineers, entrepreneurs, graduate students, and postdocs from both academia and industry. Organized by the International Machine Learning Society (IMLS), the conference fosters collaboration and innovation through the dissemination of original, rigorous research.[7][8] Typically spanning 5 to 7 days in July, ICML has adopted hybrid or fully in-person formats since 2022 following virtual events in 2020 and 2021 due to the COVID-19 pandemic, attracting thousands of attendees globally each year.[9][10][11]Significance
The International Conference on Machine Learning (ICML) is widely recognized as one of the premier venues in the field, consistently ranking among the top three machine learning conferences globally alongside NeurIPS and ICLR, based on metrics such as Google Scholar's h5-index, where ICML holds a score of 272 (as of 2025 metrics for 2020-2024 publications), reflecting its high citation impact.[12] Its selectivity is evident in acceptance rates that have hovered around 25% or lower in recent years, with 26.9% for ICML 2025 (3,260 accepted out of 12,107 submissions) and 27.5% for ICML 2024, underscoring the rigorous peer-review process that ensures only the most innovative contributions are featured.[13] This prestige positions ICML as a benchmark for academic and industry researchers seeking to disseminate cutting-edge work. ICML plays a pivotal role in advancing artificial intelligence by serving as a primary platform for breakthrough research that influences practical applications across domains. The conference has been instrumental in shaping developments in deep learning architectures and generative models, where seminal presentations have driven innovations adopted in industry tools for tasks like image synthesis and natural language processing.[14] For instance, ICML proceedings frequently introduce scalable techniques that bridge theoretical advancements with real-world deployment, contributing to the evolution of foundation models and reinforcement learning paradigms that power modern AI systems. Beyond research dissemination, ICML fosters a vibrant community ecosystem that promotes collaboration, inclusivity, and responsible practices in machine learning. Through workshops, tutorials, and ethics guidelines, the conference encourages interdisciplinary dialogue on topics like diversity in authorship and geographical representation, with studies showing ICML's role in improving gender and institutional balance compared to earlier years.[15][16] It also integrates ethical considerations into submissions, requiring authors to address potential societal impacts, thereby advancing discussions on fairness, bias mitigation, and sustainable AI development within the global research community. Attendance has grown dramatically, from hundreds in its early iterations to over 15,000 participants in 2025, amplifying these collaborative opportunities.[17]History
Founding
The International Workshop on Machine Learning, which laid the foundation for the International Conference on Machine Learning (ICML), was established in July 1980 at Carnegie Mellon University (CMU) in Pittsburgh, USA.[3] It was organized by Jaime Carbonell and Tom M. Mitchell, faculty members at CMU, and Ryszard S. Michalski from the University of Illinois at Urbana-Champaign, who recognized the emerging field of machine learning as a distinct subdiscipline within artificial intelligence requiring its own venue for discussion and advancement.[18] The workshop was supported by the Office of Naval Research, reflecting early institutional backing for machine learning research.[19] The founders were motivated by a surge in interest in machine learning around the silver anniversary of artificial intelligence in the early 1980s, driven by the need to model human-like learning processes and develop practical AI systems capable of adaptation through experience.[18] This period saw renewed enthusiasm for AI after initial optimism in the 1950s and 1960s, prompting the creation of a dedicated forum to foster collaboration among researchers exploring learning algorithms and methodologies.[19] The workshop thus served as a pivotal gathering to consolidate and propel these efforts, sparking subsequent publications and events in the field.[3] The inaugural event emphasized symbolic approaches to machine learning, such as inductive learning from examples and rule-based discovery systems, alongside nascent statistical methods for pattern recognition and generalization.[18] Proceedings from the workshop were published in limited formats, including three consecutive 1980 issues (Nos. 2, 3, and 4) of the International Journal of Policy Analysis and Information Systems, which featured selected papers and discussions from the event.[19] A special issue of the SIGART Newsletter (No. 76) in 1981 further documented ongoing research projects highlighted at the workshop, providing an early archival record of the field's directions.[18] In 1993, the event transitioned to a full conference under the ICML name.[3]Evolution
Following its origins as a series of workshops beginning in 1980, the International Conference on Machine Learning (ICML) transitioned to a full-fledged annual conference in 1993, marking a pivotal expansion in its structure and influence within the field. Subsequent workshops were held irregularly in 1983, 1985, 1987, 1988, 1989, 1990, 1991, and 1992 before becoming annual. This shift allowed ICML to accommodate a broader range of submissions and foster deeper discussions, reflecting the growing maturity of machine learning as a discipline. As the field evolved in the 1990s toward greater integration of statistical and data-driven approaches, ICML's scope similarly broadened to emphasize these methods alongside traditional symbolic and neural techniques, enabling the conference to serve as a central venue for interdisciplinary advancements.[20] Several key milestones highlight ICML's adaptation to technological and societal changes. In the early 2010s, the conference adopted open-access proceedings through the Proceedings of Machine Learning Research (PMLR), starting with ICML 2016, which enhanced global accessibility and accelerated the dissemination of research; PMLR itself was formally renamed in 2015 from its predecessor under the Journal of Machine Learning Research. The COVID-19 pandemic prompted further innovation, with ICML 2020 and 2021 held entirely virtually to ensure safety and inclusivity amid travel restrictions, incorporating online platforms for presentations, workshops, and networking that reached over 10,000 participants in 2020 alone. Submission volumes have also surged dramatically, from dozens in the 1980s workshops to more than 10,000 annually by the mid-2020s, underscoring the conference's expanding prominence and the field's rapid growth.[21][22][23][24][25] Institutionally, ICML's affiliation with the International Machine Learning Society (IMLS), established in 2002, provided stable governance and promoted professional standards, including ethical guidelines and award programs. This partnership, building on earlier efforts in the late 1990s to formalize oversight, has driven increased internationalization by hosting events across continents and enhancing diversity through initiatives like targeted travel support for underrepresented researchers. These changes have solidified ICML's role as a global benchmark for machine learning innovation.[26][27]Organization
Governance
The International Conference on Machine Learning (ICML) is primarily governed by the International Machine Learning Society (IMLS), a nonprofit organization established in 2001 to promote and support ICML as its flagship annual event, along with related workshops and initiatives in machine learning research and education.[8][27] The IMLS board, comprising officers such as the president, president-elect, secretary, treasurer, and legal advisor, along with elected members and recent ICML chairs, oversees the conference's administrative structure.[28] The board appoints general chairs annually through a majority vote process, selecting individuals with demonstrated technical and organizational expertise to lead each edition of ICML.[29] Program committees, coordinated by appointed program chairs, handle peer review to maintain rigorous academic standards, ensuring diverse expertise across machine learning subfields.[30] IMLS establishes key policies to uphold academic integrity, including ethics guidelines that require authors to address potential risks such as bias, privacy concerns, and societal impacts in submissions, with compliance to frameworks like the EU AI Act.[16] For inclusivity, ICML enforces a code of conduct prohibiting harassment, discrimination, and retaliation, promoting a respectful environment through dedicated diversity and inclusion co-chairs who investigate reports and impose sanctions as needed.[31] Open science principles are integrated via requirements for reproducible experiments, encouraging transparency in methods and data practices without mandating full code or data release.[32][33] Decision-making on operational aspects, such as venue selection and theme setting, is managed by the IMLS board to align with strategic goals like global accessibility and topical relevance, while conflict-of-interest guidelines—covering reviewer assignments, authorship disclosures, and ethical review conduct—are enforced to prevent bias and ensure fair peer evaluation.[33][29] These mechanisms collectively safeguard the conference's integrity and foster equitable participation in the machine learning community.[16]Sponsors
The International Conference on Machine Learning (ICML) is supported by a range of corporate and institutional sponsors, with major contributors including Google, Microsoft, Amazon, Meta (including DeepMind), Apple, Citadel Securities, Jane Street Capital, and D.E. Shaw & Co. These organizations provide substantial financial backing, often at the highest sponsorship tiers, to facilitate the conference's operations and outreach efforts.[34][35][36] Sponsorship plays a critical role in ICML by covering essential costs such as venue logistics, student scholarships, travel grants for researchers from underrepresented groups, and the development of industry-focused tracks that bridge academic research with practical applications. Funds from sponsors are allocated to need-based programs, supporting graduate student attendance (including hotel, food, and registration) and diversity, equity, and inclusion (DEI) initiatives to broaden participation in machine learning.[37] Sponsorship levels—such as Diamond ($100,000 USD), Platinum ($80,000 USD), Gold ($45,000 USD), and Silver ($30,000 USD)—offer tiered benefits, including booth space in the EXPO hall, multiple conference registrations, logo visibility on the ICML website and materials, acknowledgment at events, and access to recruitment tools like job postings and interview rooms, which enhance sponsors' exposure to emerging talent.[37] Corporate involvement in ICML sponsorship has grown significantly since the 2000s, mirroring the increasing commercial relevance of machine learning technologies in sectors like technology, finance, and healthcare. This expansion has enabled larger-scale events and greater accessibility, while the International Machine Learning Society (IMLS), which organizes ICML, maintains oversight to ensure sponsorships align with the conference's academic priorities and ethical standards.[38][39]Format
Submission and Review
The submission process for the International Conference on Machine Learning (ICML) operates on a fixed annual timeline to accommodate the July conference schedule. Typically, abstract submissions are due in late January, followed by full paper submissions in early February, both under Anywhere on Earth (AoE) deadlines; for ICML 2025, these were January 23 and January 30, respectively. Author notifications are issued in early May, with camera-ready deadlines set for early June; in 2025, notification occurred on May 1, and camera-ready papers were due June 5.[32][33][40] ICML employs a rigorous double-blind peer review process managed by a program committee of thousands of experts, including reviewers, area chairs, and senior area chairs; for instance, ICML 2024 involved 7,474 reviewers and 492 area chairs.[33][5] Each submission receives multiple reviews, typically at least three, evaluating key criteria such as novelty (e.g., innovative applications or combinations of methods), technical soundness (e.g., validity of claims, proofs, and experiments), and reproducibility (e.g., clarity of methodology and availability of supplementary materials).[33][41] The acceptance rate hovers around 25-28%, reflecting the conference's selectivity; in 2024, 2,609 papers were accepted from 9,473 submissions, and in 2025, 3,260 papers were accepted from 12,107 submissions (26.9%).[5][42] Several policies guide the submission and review to ensure fairness and quality. A rebuttal phase allows authors to address factual errors or answer specific reviewer questions, typically spanning late March to early April; for ICML 2025, this ran until March 31, followed by reviewer discussions ending April 8.[41] The conference maintains a strict dual submission policy, prohibiting simultaneous submissions to other peer-reviewed venues with archival proceedings, though non-archival workshop presentations are permitted.[32] It features a main research track for original technical contributions and a separate position paper track for broader perspectives on ML challenges, both leading to archival publications.[32][43] Ethical considerations are emphasized through mandatory impact statements in submissions, assessing potential societal effects and promoting responsible AI, with reviewers noting any ethical concerns.[32] All authors must also commit to reciprocal reviewing, with at least one qualified author per submission serving on the program committee.[32]Program Activities
The International Conference on Machine Learning (ICML) features a core program centered on the presentation and discussion of rigorously reviewed research papers, selected through a double-blind peer-review process. These activities typically span 5 to 7 days, with the main conference days organized into parallel tracks to accommodate the high volume of contributions, allowing attendees to engage deeply with advancements in machine learning.[44][1] Oral presentations form a key component, where selected papers are delivered as 20- to 30-minute talks by authors, highlighting novel methodologies and empirical findings in areas such as deep learning architectures and optimization techniques.[45] Complementing these are poster sessions, which provide opportunities for extended one-on-one discussions; posters are displayed in dedicated halls, often with spotlight sessions featuring brief 5-minute overviews to draw attention to high-impact work.[46] Invited talks by prominent researchers, such as those on closing the loop between learning and experimentation, offer broader perspectives on field-defining challenges and are scheduled across the main days, typically numbering 4 to 6 per conference.[47] Tutorials, held at the outset of the event, deliver 2-hour educational sessions on foundational and emerging topics, including data-efficient supervised learning and neural operators for scientific computing, equipping attendees with practical insights into techniques like federated learning.[48][49] Ancillary events enrich the program by fostering specialized discourse and networking. Workshops, occurring concurrently at the conference's close, focus on niche themes such as machine learning for healthcare—exemplified by sessions on interpretable models for equitable disease phenotyping—and include invited talks, contributed posters, and panel discussions to explore interdisciplinary applications.[50][51] Social events, including receptions and affinity gatherings, promote informal interactions among researchers, while diversity receptions support underrepresented groups through targeted networking.[46] Career fairs facilitate connections between academics and industry recruiters, and industry expos, introduced in the 2010s, showcase tools and platforms from leading companies to bridge research and practical deployment.[52][53] Since 2020, ICML has adopted hybrid formats to enhance global accessibility, incorporating live streams of oral sessions, invited talks, and tutorials, alongside virtual poster viewing platforms that enable asynchronous engagement regardless of time zones.[54] This structure ensures that both in-person and remote participants can fully interact with the program's intellectual and communal elements.[55]Publications
Proceedings
The proceedings of the International Conference on Machine Learning (ICML) are published in the Proceedings of Machine Learning Research (PMLR), an open-access series dedicated to machine learning conference outputs. PMLR, which evolved from the Journal of Machine Learning Research (JMLR) Workshop and Conference Proceedings, has hosted ICML volumes since the early 2010s, with the transition formalized in 2015 when JMLR rebranded its proceedings arm as PMLR. The series is managed under JMLR and print-archived by Microtome Publishing, ensuring long-term preservation while prioritizing digital accessibility.[21][56][57] All accepted papers, whether selected for oral, spotlight, or poster presentations, are included in the annual PMLR volume corresponding to the conference year, typically released shortly after the event. Each volume compiles full-length papers in a standardized format, adhering to PMLR's LaTeX template for consistency. Supplementary materials, such as additional proofs, experimental details, code repositories, and datasets, are supported and often linked directly from the PMLR entry for each paper, though PMLR does not formally archive them as part of the core volume; authors are encouraged to host these in public repositories like GitHub or Zenodo for reproducibility. These proceedings are rigorously indexed in major academic databases, including Google Scholar, DBLP, and Scopus, facilitating discoverability and scholarly impact.[40][58][59] PMLR's open-access model provides free, unrestricted downloads of all papers via its website, with each assigned a unique Digital Object Identifier (DOI) for precise citation and tracking. This policy, rooted in JMLR's commitment to open science since 2000, has significantly enhanced the global reach of ICML research by removing paywalls, leading to increased citations and broader adoption in academia and industry compared to earlier proprietary publishing models used by ICML. For instance, the shift to open access has correlated with PMLR volumes achieving high visibility metrics.[56][60][56]Notable Contributions
The International Conference on Machine Learning (ICML) has hosted numerous landmark papers that have shaped core algorithms and methodologies in the field. A notable early contribution is the 2004 paper "K-means Clustering via Principal Component Analysis" by Chris Ding and Xiaofeng He, which demonstrated that principal components provide continuous approximations to the discrete cluster indicators in k-means, yielding new lower bounds for the clustering objective and enhanced initialization techniques for dimensionality reduction in unsupervised learning.[61] This work has influenced hybrid clustering methods by bridging principal component analysis with partitioning algorithms, facilitating more efficient data analysis in high-dimensional spaces. In 2015, Sergey Ioffe and Christian Szegedy introduced "Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift" at ICML, proposing a normalization method applied to each mini-batch during training to mitigate internal covariate shift, thereby stabilizing gradients and enabling faster convergence in deep neural networks. With over 64,000 citations as of November 2025, this technique has become ubiquitous in deep learning frameworks, profoundly impacting the scalability of models like convolutional and recurrent networks by allowing deeper architectures without vanishing gradients.[62][63] Advancing reinforcement learning, the 2018 ICML paper "Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor" by Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine presented the SAC algorithm, which maximizes both reward and entropy to promote robust exploration in continuous action spaces, outperforming prior off-policy methods in sample efficiency. Cited more than 15,000 times as of November 2025, SAC has driven progress in robotic control and game-playing agents, inspiring entropy-regularized approaches that balance exploitation and exploration in real-world applications.[64] For efficient computer vision, Mingxing Tan and Quoc V. Le's 2019 ICML paper "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks" proposed a compound scaling coefficient to uniformly balance depth, width, and resolution, achieving top ImageNet accuracy with 8.4 times fewer parameters than previous models like GPipe.[65] Garnering over 25,000 citations as of November 2025, EfficientNet has transformed resource-constrained deployment in mobile and edge computing, influencing scalable architectures in industry and research.[66] Bridging vision and language, the 2021 ICML paper "Learning Transferable Visual Models From Natural Language Supervision" by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Tim Salimans, Pamela Mishkin, Sandhini Agarwal, Lisa Anne Hendricks, Katie Child, Mark Miller, and Ilya Sutskever introduced CLIP, a contrastive model pretrained on 400 million image-text pairs for zero-shot classification across diverse tasks. With over 40,000 citations as of November 2025, CLIP has catalyzed the vision-language foundation model paradigm, enabling flexible generalization and powering applications in content moderation and creative AI.[67] More recently, the 2023 ICML paper "Llama 2: Open Foundation and Fine-Tuned Chat Models" by Hugo Touvron et al. presented an open-source large language model family trained on 2 trillion tokens, achieving competitive performance with closed models while emphasizing safety and alignment, influencing the democratization of AI capabilities. Cited over 5,000 times as of November 2025, it has spurred advancements in accessible generative AI.[68] ICML's legacy underscores trends toward rigorous, impactful research, with proceedings published in PMLR emphasizing verifiable advancements. The conference promotes reproducible results via mandatory checklists requiring detailed experimental setups, dataset descriptions, and code availability since 2020, enhancing the reliability of machine learning findings.[69] Best paper awards frequently spotlight innovations in optimization, such as adaptive methods for non-convex problems, and fairness, including techniques for equitable representation learning, as seen in multiple outstanding paper selections from 2018 onward that address bias mitigation and inclusive AI design.[70][71]Venues
Past Conferences
The International Conference on Machine Learning (ICML) originated as a small workshop in 1980 at Carnegie Mellon University in Pittsburgh, Pennsylvania, USA, attracting a modest audience of researchers primarily from North American institutions.[3] In its early years during the 1980s and 1990s, ICML events remained small-scale gatherings, typically hosted at US universities with attendance in the low hundreds and focused on foundational topics in machine learning among a close-knit academic community.[72] The transition to a formal conference format occurred in 1993 at the University of Massachusetts in Amherst, Massachusetts, USA, marking the beginning of more structured annual proceedings. During the 2000s, ICML began to internationalize, expanding beyond North America to include venues in Europe, Asia, and Australia, which reflected growing global interest in machine learning research.[73] For example, the 2008 edition was held in Helsinki, Finland, signaling Europe's rising role in the field.[73] This period saw steady growth in participation, with events like the 2010 conference in Haifa, Israel, fostering broader international collaboration.[74] The 2010s marked significant expansion, with venues increasingly rotating to Asia and Europe amid rising attendance, which spiked post-2010 due to the field's rapid advancement and increased submissions.[5] Notable examples include the 2015 event in Lille, France, and the 2017 conference in Sydney, Australia, which drew over 2,400 attendees and highlighted the shift toward global rotation away from North American dominance.[75][76] By 2019, in Long Beach, California, USA, attendance reached approximately 6,500, underscoring the conference's scale.[5] The COVID-19 pandemic prompted adaptations, with ICML 2020 and 2021 held virtually—originally planned for Vienna, Austria—resulting in record-high participation of over 10,800 in 2020 and nearly 8,900 in 2021 due to global accessibility.[10][5] Post-pandemic, hybrid formats emerged, as seen in 2022 in Baltimore, Maryland, USA (over 7,100 total attendees), 2023 in Honolulu, Hawaii, USA (nearly 8,000), 2024 in Vienna, Austria (over 9,000), and 2025 in Vancouver, British Columbia, Canada.[77][5][1] This evolution illustrates ICML's transformation from a regional workshop to a premier global forum, with attendance consistently exceeding 7,000 in recent years.[5]| Year | Location |
|---|---|
| 1980 | Pittsburgh, Pennsylvania, USA[72] |
| 1993 | Amherst, Massachusetts, USA |
| 2008 | Helsinki, Finland[73] |
| 2010 | Haifa, Israel[74] |
| 2015 | Lille, France[75] |
| 2017 | Sydney, New South Wales, Australia[76] |
| 2019 | Long Beach, California, USA |
| 2020–2021 | Virtual[10] |
| 2022 | Baltimore, Maryland, USA[77] |
| 2023 | Honolulu, Hawaii, USA[78] |
| 2024 | Vienna, Austria[79] |
| 2025 | Vancouver, British Columbia, Canada[1] |