AIVEX Proves Its Technological Competitiveness in Core Industrial Physical AI Technologies with Paper Acceptance at ICML 2026

June 12, 2026, Seoul – AIVEX, an industrial Physical AI company, announced that a research paper from its Deep Learning Research Group has been accepted to ICML (International Conference on Machine Learning) 2026, one of the world’s most prestigious AI conferences.

ICML, together with NeurIPS (Neural Information Processing Systems) and ICLR (International Conference on Learning Representations), is recognized as one of the world’s top three AI conferences, serving as a premier academic forum where the latest breakthroughs in machine learning are presented.

The accepted paper, titled “Memory-Distilled Selection for Noise-Robust Anomaly Detection,” introduces a novel anomaly detection method designed to reliably identify defects even in environments with noisy training data, a common challenge in real-world industrial applications.

In manufacturing environments, obtaining perfectly clean training data is highly challenging, and this has long been regarded as one of the key barriers to adopting industrial AI. To address this issue, AIVEX’s Deep Learning Research Group developed the MeDS (Memory-Distilled Selection) algorithm, which enables robust defect detection at both image and pixel levels, even when training datasets contain mislabeled or anomalous samples without accurate ground-truth labels.

In particular, the research introduced a subsampling-based filtering technique that randomly selects subsets of feature data to construct multiple memory ensembles. The research team mathematically demonstrated that the sparsity generated through this process functions as a low-pass filter, enabling effective selection of normal patterns even under highly noisy conditions.

Furthermore, during the knowledge distillation process, where filtered memory scores are transferred to a student network, the method leverages the neural network’s early-learning bias, in which simpler patterns are learned at an earlier stage of training. This approach prevents noise overfitting and significantly improves the model’s ability to identify normal patterns.

In addition, by repeatedly performing a self-selection process in which the model selects clean samples for training based on distilled model scores, the proposed method enables precise pixel-level defect localization. The approach achieved state-of-the-art performance on leading industrial vision AI benchmark datasets, including MVTecAD, VisA, and Real-IAD, demonstrating both high robustness and practical applicability.

AIVEX CEO Minsoo Sung said, “The acceptance of this paper at ICML 2026 is highly meaningful as it demonstrates global recognition of AIVEX’s AI research capabilities and core technology expertise. We will continue to advance industrial Physical AI technologies that address real-world manufacturing challenges and strengthen our global competitiveness.”

Founded in 2020, AIVEX provides an industrial AI platform for manufacturing environments based on AI vision inspection, AI robotics, MLOps, and data platforms. With more than 80% of its employees being engineers, the company continues to strengthen its global technological competitiveness through continuous research and development. AIVEX currently holds a patent portfolio of approximately 80 patents, including 53 registered patents in Korea and overseas.