International Conference On Learning Representations

The International Conference on Learning Representations, or ICLR as it’s commonly known, is a significant academic gathering in the field of machine learning . It typically unfolds annually, usually gracing the calendar in late April or early May. Alongside its esteemed peers, NeurIPS and ICML , ICLR stands as one of the preeminent forces shaping the discourse and advancement within machine learning and artificial intelligence research. It’s not just a conference; it’s a crucible where the future of learning representations is forged.

The conference structure itself is designed to foster deep engagement and broad dissemination of ideas. It features a curated selection of invited talks by luminaries in the field, alongside rigorous oral and poster presentations of peer-reviewed papers. This dual approach ensures that both groundbreaking, high-level perspectives and detailed, specific research findings find their platform.

Since its inception in 2013, ICLR has championed an open peer review process, a model notably influenced by the proposals of Yann LeCun . This transparency in evaluation is a cornerstone of its philosophy, aiming to democratize the scientific publication process. The sheer volume of submissions underscores its growing importance. For instance, in 2019, the conference received 1591 paper submissions, of which a select 500 were accepted for poster presentations, representing a 31% acceptance rate. An even more exclusive group, just 24 papers, earned the distinction of oral presentations, a mere 1.5%. The trend continued, with 2021 seeing 2997 submissions, of which 860 were accepted, maintaining a competitive acceptance rate of 29%. The conference’s origins trace back to 2012, when it was founded by the visionary minds of Yann LeCun and Yoshua Bengio , pioneers who foresaw the transformative potential of learning representations.

The physical locations of ICLR have been as diverse as the research presented, reflecting its global reach and the dynamic nature of the AI community.

• ICLR 2026: Scheduled to convene in the vibrant city of Rio de Janeiro , Brazil .
• ICLR 2025: Set to host attendees in Singapore .
• ICLR 2024: Took place in the historic city of Vienna , Austria .
• ICLR 2023: Welcomed the community to Kigali , Rwanda .
• ICLR 2022: Navigated the challenges of the time with a virtual conference.
• ICLR 2021: Also held virtually, with its primary location in Vienna , Austria .
• ICLR 2020: Faced unprecedented circumstances, leading to a virtual format. Its intended location was Addis Ababa , Ethiopia , a move that itself highlighted important discussions around accessibility and global participation in AI conferences.
• ICLR 2019: Was hosted in New Orleans , Louisiana , United States .
• ICLR 2018: Took place in the scenic city of Vancouver , Canada .
• ICLR 2017: Found its home in Toulon , France .
• ICLR 2016: Was held in San Juan , Puerto Rico , United States .
• ICLR 2015: Convened in San Diego , California , United States .
• ICLR 2014: Journeyed to the stunning landscapes of Banff National Park , Canada .
• ICLR 2013: Marked the inaugural event in Scottsdale , Arizona , United States .

The conference operates on an open access model, with publications readily available on openreview.net, ensuring that the latest research is accessible to a wide audience. This commitment to open science is crucial for accelerating progress in a field that is evolving at breakneck speed.

ICLR is an integral part of a broader ecosystem of machine learning and data mining research. Its focus on learning representations is fundamental to many modern AI advancements. The paradigms explored at ICLR span the entire spectrum of machine learning :

• **Supervised learning :** Where models learn from labeled data, encompassing tasks like classification and regression .
• **Unsupervised learning :** Discovering patterns in unlabeled data, often used for clustering and dimensionality reduction .
• **Semi-supervised learning :** Leveraging both labeled and unlabeled data, a practical approach when labeling is expensive.
• **Self-supervised learning :** A subset of unsupervised learning where the supervision signal is derived from the data itself, often through pretext tasks.
• **Reinforcement learning :** Agents learning through trial and error by interacting with an environment, crucial for decision-making and control.
• **Meta-learning :** “Learning to learn,” where models acquire knowledge that allows them to learn new tasks more efficiently.
• **Online learning :** Models that learn incrementally from a stream of data.
• **Batch learning :** Models trained on the entire dataset at once.
• **Curriculum learning :** Training models by presenting data in a structured, progressive order, mimicking how humans learn.
• **Rule-based learning :** Systems that learn explicit rules.
• **Neuro-symbolic AI :** An emerging area aiming to combine the strengths of neural networks and symbolic reasoning.
• **Neuromorphic engineering :** Designing hardware inspired by the structure and function of the brain.
• **Quantum machine learning :** Exploring the intersection of quantum computing and machine learning.

The problems addressed are equally diverse, ranging from predictive tasks like classification and regression to generative tasks like creating new data through generative modeling . Other key areas include identifying groups in data via clustering , simplifying complex data with dimensionality reduction , understanding data distributions through density estimation , spotting unusual data points with anomaly detection , ensuring data quality via data cleaning , and automating the model selection process with AutoML . The conference also delves into discovering relationships with association rules , understanding meaning through semantic analysis , predicting complex outputs with structured prediction , and the crucial processes of feature engineering and feature learning . Further specializations include learning to rank , inferring rules from data with grammar induction , building knowledge bases with ontology learning , and integrating information from multiple sources in multimodal learning .

The conference proceedings are a testament to the rapid evolution of neural network architectures and learning algorithms. Artificial neural networks and deep learning are central themes, with research on feedforward neural networks , recurrent neural networks (including LSTM and GRU ), convolutional neural networks (with seminal architectures like LeNet and AlexNet often cited), and more recent breakthroughs like Transformers and Vision Transformers . The burgeoning field of Physics-informed neural networks represents an exciting integration of scientific principles with data-driven learning. The exploration of GANs and Diffusion models continues to push the boundaries of generative capabilities.

ICLR also serves as a critical venue for advancements in reinforcement learning , exploring algorithms like Q-learning and policy gradients , and investigating complex scenarios such as multi-agent interactions and learning through self-play . The growing importance of human interaction in AI development is reflected in research on active learning , crowdsourcing , human-in-the-loop systems, and Reinforcement learning from human feedback (RLHF) .

Beyond the algorithms and models, ICLR also fosters discussion on the foundational aspects of machine learning, including computational learning theory , statistical learning theory , and the critical bias–variance tradeoff . The interpretability of models, a growing concern, is also addressed through areas like mechanistic interpretability .

The conference’s influence is further solidified by its inclusion in the pantheon of top AI and machine learning journals and conferences, alongside AAAI , ECML PKDD , NeurIPS , ICML , and IJCAI , and journals like ML and JMLR .

This article, while comprehensive, is a living document, constantly evolving as the field itself does. Like many academic endeavors, it benefits from community input and expansion. It is currently classified as a stub , inviting further contributions to enrich its content.

See also: ICML , NeurIPS , AAAI Conference on Artificial Intelligence , Glossary of artificial intelligence , Outline of machine learning .