Christian Szegedy
Last reviewed
Sources
16 citations
Review status
Source-backed
Revision
v3 · 1,911 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
16 citations
Review status
Source-backed
Revision
v3 · 1,911 words
Add missing citations, update stale details, or suggest a clearer explanation.
Christian Szegedy is a Hungarian mathematician and machine learning researcher best known for three foundational deep learning contributions from his decade at Google: the Inception (GoogLeNet) convolutional neural network architecture, the batch normalization training method, and the 2013 paper that introduced adversarial examples [1][5][6][7]. He was a founding scientist of xAI in 2023, and after leaving in 2025 he founded Math, Inc., a startup pursuing what he calls "verified superintelligence" through the automatic translation of mathematics into machine-checkable proofs [13][14][16]. His research papers had accumulated more than 330,000 citations by 2026, placing him among the most cited researchers in artificial intelligence [1].
Christian Szegedy is a researcher whose career bridges two eras of artificial intelligence: the rise of deep learning for computer vision in the 2010s, in which he was a central architect, and the more recent effort to apply AI to formal mathematical reasoning, which he now pursues full time. During more than a decade at Google he was a lead author of the Inception family of architectures, co-authored batch normalization, and was the first author of the work that introduced adversarial examples [1][5][6][7]. He later founded and led Google's autoformalization research group, was one of the founding members of xAI in 2023, and in 2025 left to work full time on AI for mathematics, founding the startup Math, Inc. [13][14][16].
Szegedy was born in Hungary and began his university studies in mathematics at Eotvos Lorand University in Budapest in the early 1990s [2]. He moved to Germany for graduate work at the University of Bonn, where he completed a doctorate in 2005 under Bernhard Korte at the Research Institute for Discrete Mathematics [2][4]. His dissertation, "Some Applications of the Weighted Combinatorial Laplacian," applied spectral and combinatorial optimization methods to matching theory and to the physical design of very large scale integrated circuits, that is, to the layout of computer chips [3].
That background carried him into the electronic design automation industry. From 2005 to 2010 he worked as a research scientist at Cadence Design Systems, developing mathematical methods for timing-driven placement and synthesis, core steps in turning a chip design into a manufacturable layout [1]. In 2010 he joined Google [2].
At Google, Szegedy moved from chip design into the company's fast-growing machine learning effort, eventually becoming a staff research scientist in Google Research [1][2]. He arrived just as deep learning was beginning to dominate computer vision, and over the next several years he became one of the most cited researchers in the field; his publications had accumulated more than 330,000 citations by 2026 [1]. His work spanned image classification, object detection, and network architecture design, and it produced several methods that became standard building blocks of modern neural networks.
Szegedy's first widely noted result was an unexpected one. In the 2013 paper "Intriguing properties of neural networks," he and his co-authors, who included Ilya Sutskever and Ian Goodfellow, showed that image classifiers could be fooled by tiny, carefully chosen perturbations that are imperceptible to humans yet cause confident misclassification [5]. The paper stated the finding plainly: "We can cause the network to misclassify an image by applying a certain imperceptible perturbation, which is found by maximizing the network's prediction error" [5]. It introduced the term adversarial examples and opened a large subfield of research on the robustness and security of machine learning models.
The following year he led the team behind GoogLeNet, the 22-layer network described in "Going Deeper with Convolutions" [6]. Its central idea, the Inception module, applied convolutions at several scales in parallel and concatenated the results, which let the network grow much deeper without an explosion in computation; the GoogLeNet incarnation used only about 6.8 million parameters, roughly nine times fewer than AlexNet [6]. GoogLeNet won the classification and detection tracks of the 2014 ImageNet Large Scale Visual Recognition Challenge with a top-5 error rate of 6.67 percent, and the architecture went on to power Google services such as photo search and Street View imagery analysis [6]. The name is a play on Yann LeCun's earlier LeNet [6].
In 2015 Szegedy co-authored, with Sergey Ioffe, "Batch Normalization" [7]. The technique normalizes the inputs to each network layer using the statistics of the current training mini-batch, which stabilizes and greatly speeds up training, permits higher learning rates, and reduces the need for careful weight initialization. The paper reported that the method reached the same image-classification accuracy as a state-of-the-art baseline using 14 times fewer training steps [7]. It became one of the most widely used and most cited techniques in all of deep learning. Szegedy continued to refine the Inception line in "Rethinking the Inception Architecture for Computer Vision," which introduced Inception-v2 and v3, and in "Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning," which combined Inception modules with the residual connections popularized by ResNet [8][9].
| Paper | Year | Venue | Significance |
|---|---|---|---|
| Intriguing Properties of Neural Networks | 2013 | ICLR 2014 | Introduced adversarial examples [5] |
| Going Deeper with Convolutions (GoogLeNet) | 2014 | CVPR 2015 | Inception architecture; won ImageNet ILSVRC 2014 at 6.67% top-5 error [6] |
| Batch Normalization | 2015 | ICML 2015 | Normalizing layer inputs; matched baseline accuracy in 14x fewer steps [7] |
| Rethinking the Inception Architecture | 2015 | CVPR 2016 | Inception-v2 and v3 [8] |
| Inception-v4, Inception-ResNet | 2016 | AAAI 2017 | Inception combined with residual connections [9] |
| DeepMath: Deep Sequence Models for Premise Selection | 2016 | NeurIPS 2016 | Early neural network theorem proving [10] |
| Autoformalization with Large Language Models | 2022 | NeurIPS 2022 | LLMs translate informal math into formal statements [12] |
Around 2016 Szegedy turned his attention to a goal he would pursue for the rest of his career: applying deep learning to mathematical reasoning. He was a co-author of "DeepMath: Deep Sequence Models for Premise Selection," one of the first demonstrations that large neural networks could be useful for large-scale automated theorem proving, in this case by selecting which existing facts are relevant to proving a new statement [10]. The team included Francois Chollet and Geoffrey Irving.
At Google he founded and led the N2Formal group, whose stated aim was to build an "automatic mathematician." He laid out the research program in a 2020 paper, "A Promising Path Towards Autoformalization and General Artificial Intelligence," presented at the Conference on Intelligent Computer Mathematics [11]. Autoformalization is the task of automatically translating ordinary mathematical writing, expressed in natural language, into a precise, machine-checkable formal language. Szegedy argued that a system that could read mathematics and formalize it, bootstrapping largely from unlabeled text, would acquire general reasoning ability applicable well beyond mathematics. In 2022 his group offered early evidence for the idea in "Autoformalization with Large Language Models," which showed that large language models could translate natural-language problems into formal statements in the Isabelle proof assistant [12].
In 2023 Szegedy joined xAI, the company founded by Elon Musk and publicly announced in July 2023 with a twelve-person founding team [13]. He is generally listed among the company's co-founders, alongside researchers such as Greg Yang and Jimmy Ba, though he has described his own role more modestly: "I would more like consider myself as a founding scientist or a founding engineer than a co-founder really" [14]. At xAI he contributed to the research behind Grok, the company's family of large language models, during a period in which a team of roughly a dozen technical staff operated a very large cluster of graphics processors [14].
Szegedy left xAI in February 2025, one of the first of the founding team to depart, following Kyle Kosic, who had left in 2024 [13]. The rest of the founding group followed over the next year. Greg Yang stepped back into an advisory role in January 2026, and Jimmy Ba and Yuhuai Wu departed in February 2026, by which point reporting described essentially the entire original founding team as having moved on [13].
After leaving xAI, Szegedy returned to autoformalization full time. He became chief scientist at Morph Labs, a startup applying reinforcement learning to formal mathematics that he had earlier backed as a seed investor; he has said he left xAI specifically to pursue the formalization project there [14]. In September 2025 he announced Math, Inc., describing it as a company dedicated to the creation of verified superintelligence via autoformalization, built on the reinforcement learning infrastructure developed at Morph Labs [16]. He frames machine-checked mathematics as essential infrastructure for trustworthy AI: a layer beneath all reasoning where proofs are guaranteed correct without anyone having to read them [14].
The company introduced its first system at the same time: an autoformalization agent named Gauss [15]. Math, Inc. reported that Gauss had completed a challenge posed in January 2024 by the mathematicians Terence Tao and Alex Kontorovich, namely a full formalization of the strong Prime Number Theorem in the Lean proof assistant. After Tao and Kontorovich announced intermediate progress in July 2025, the company said, Gauss finished the remaining formalization in about three weeks, producing roughly 25,000 lines of Lean code containing more than 1,000 theorems and definitions [15]. Math, Inc. said it aimed to increase the total amount of formal mathematical code by 2 to 3 orders of magnitude over the following year [15]. Szegedy has predicted, with his frequent collaborator Francois Chollet, that AI systems will reach superhuman mathematical ability within the next few years [14].