Koray Kavukcuoglu
Last reviewed
May 31, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,833 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,833 words
Add missing citations, update stale details, or suggest a clearer explanation.
Koray Kavukcuoglu is a Turkish computer scientist who serves as Chief Technology Officer of Google DeepMind and, since 2025, as Chief AI Architect of Google. He joined DeepMind in 2012 as one of its early research scientists and rose through the organization to lead research before becoming its CTO. His name is associated with several of the laboratory's best known results, including the deep Q-network that learned to play Atari games, the AlphaGo program, and the WaveNet audio model [1][2][3]. In June 2025 Google's chief executive Sundar Pichai named him to the newly created role of Chief AI Architect, a senior vice president position reporting directly to Pichai, with a remit to speed the integration of Google's Gemini models into the company's products [4][5].
Before his industrial career, Kavukcuoglu trained first as an aerospace engineer and then completed a doctorate in computer science at New York University under Yann LeCun, where his thesis dealt with unsupervised learning of feature hierarchies and with convolutional networks for visual recognition [6][7].
Kavukcuoglu earned undergraduate and master's degrees in aerospace engineering from the Middle East Technical University in Ankara [6][7]. He moved into computer science as a graduate student at New York University, completing a master's degree in 2005 and a doctorate in 2010 [6]. His doctoral advisor was Yann LeCun, a researcher closely identified with convolutional neural networks, and his work also drew on the computational neuroscience tradition associated with Eero Simoncelli at NYU's Center for Neural Science [6][7].
His 2010 dissertation, titled Learning Feature Hierarchies for Object Recognition, addressed how a system can learn useful image features from data without hand designed labels [6]. The thesis brought together sparse coding, a method that represents inputs as combinations of a small number of basis elements, with the layered structure of convolutional networks. The timing placed his graduate work just before the rapid expansion of deep learning, and several techniques he developed during this period reappeared in later research on feature learning.
During his graduate years and in a subsequent post at NEC Laboratories America, Kavukcuoglu published a sequence of papers on unsupervised feature learning [7][8]. The work that became most widely cited is Predictive Sparse Decomposition, often abbreviated PSD. Sparse coding had long been used to find compact representations of data, but solving the sparse code for each new input required an iterative optimization that was slow at test time. PSD trained a fast predictor function alongside the sparse dictionary, so that the system could produce a good approximation of the sparse representation in a single forward pass [8]. This made sparse coding practical as a building block inside larger recognition pipelines.
Kavukcuoglu extended these ideas to images in collaboration with LeCun and others. A 2009 paper on invariant features through topographic filter maps and a 2010 paper on learning convolutional feature hierarchies showed how sparse coding could be applied across the spatial structure of an image and stacked into multiple stages [7]. He also contributed to a survey of convolutional networks and their applications in vision, and he helped build early software for the field, including work on the Torch machine learning library and on hardware accelerated convolutional networks [7][9].
Kavukcuoglu joined DeepMind in 2012, two years before Google acquired the London laboratory in 2014 [4][10]. He arrived as one of the company's first research scientists and worked on the early projects that established its reputation.
The first of these was deep reinforcement learning applied to video games. In a 2015 paper in the journal Nature, titled Human-level control through deep reinforcement learning, Kavukcuoglu was the second author after Volodymyr Mnih on the work that introduced the deep Q-network, or DQN [1]. The system combined a convolutional network with Q-learning, a method from reinforcement learning, and learned to play 49 Atari 2600 games directly from screen pixels, reaching a level comparable to a professional human tester on many of them [1]. The result drew wide attention because a single architecture learned a range of different games without game specific tuning.
Kavukcuoglu was also among the authors of the 2016 Nature paper Mastering the game of Go with deep neural networks and tree search, which described AlphaGo [2]. AlphaGo combined policy and value networks, trained by supervised learning from human games and by reinforcement learning through self play, with Monte Carlo tree search. The program defeated a professional Go player, a milestone that had been considered years away for computer Go [2].
In the same year he was the senior author of WaveNet, described in the paper WaveNet: A Generative Model for Raw Audio [3]. WaveNet generated audio one sample at a time using a deep autoregressive network with dilated convolutions, and when applied to text to speech it produced output that listeners rated as more natural than the parametric and concatenative systems used at the time [3]. The model later informed speech generation in Google products.
Kavukcuoglu became VP of Research at DeepMind, a role in which he founded the deep learning team and oversaw the laboratory's research agenda [4][5][9]. Under his direction the organization pursued a broad program that included game playing systems, protein structure prediction, and language models. He was credited as a contributor to several of these efforts, and his judgment helped set research priorities during a period of rapid output.
In April 2023 Google combined DeepMind with the Google Brain team to form a single unit called Google DeepMind [10][11]. As part of the reorganization the company created a scientific board to guide research direction, and Kavukcuoglu, then VP of Research, was named to lead it [11]. The merger brought together two of the largest AI research groups in one organization led by Demis Hassabis as chief executive [10][11].
Following the formation of Google DeepMind, Kavukcuoglu became its Chief Technology Officer [4][12]. In that role he leads the development of the organization's generative AI models and the work of bringing them into Google's products at scale [12]. His career to that point had run entirely through research environments, from NYU and NEC Laboratories America to more than a decade at DeepMind, and the CTO appointment placed a researcher with deep technical background at the head of the laboratory's engineering and model development [4][9].
The period coincided with the development of the Gemini family of models, Google DeepMind's main line of large multimodal systems. As CTO he represented the technical direction of this work in public settings, including appearances at Google's developer conferences [12][13].
On 11 June 2025 Sundar Pichai announced in a memo to employees that Kavukcuoglu would take on a new role as Chief AI Architect, a senior vice president position reporting directly to Pichai [4][5]. He kept his existing post as CTO of Google DeepMind while adding the new responsibility [4][5]. As part of the change he relocated from London to the area around Google's headquarters in Mountain View, California, to work more closely with the company's product teams [4][5].
The role was framed around closing the distance between research and products. In the memo Pichai described the goal as accelerating how the company brings its leading models into its products, with an aim of more seamless integration, faster iteration, and greater efficiency [4][5]. In practice this meant overseeing how Gemini and related systems are embedded across Google's offerings, from Search and the Android platform to the company's hardware [4][14]. Reporting on the appointment placed it in the context of intense competition in generative AI, where rivals were releasing capable systems quickly and where the speed of shipping models into widely used products had become a measure of standing in the field [4][14].
The creation of the position reflected a structural choice at Google. Rather than leaving model development and product integration to separate chains of command, the company put a single senior technical leader across both, drawing on Kavukcuoglu's familiarity with the models themselves [4][5][14].
Kavukcuoglu is widely regarded as a leading figure in artificial intelligence, and Google's own description calls him one of the world's foremost experts in the field [12]. His standing rests on authorship of several heavily cited papers, including the Nature articles on the deep Q-network and on AlphaGo and the WaveNet paper, each of which became a reference point in its area [1][2][3]. Industry listings of senior technology leaders have included him among prominent chief technology officers in Europe and the wider region [9]. His career is also noted as an example of a research scientist who moved from foundational academic work on feature learning to leadership of a major industrial laboratory and then to a company wide role shaping how AI reaches products [4][9][15].
| Field | Detail |
|---|---|
| Full name | Koray Kavukcuoglu |
| Nationality | Turkish |
| Field | Artificial intelligence, deep learning, reinforcement learning |
| Undergraduate and master's | Aerospace engineering, Middle East Technical University, Ankara |
| Graduate study | MS (2005) and PhD (2010) in computer science, New York University |
| Doctoral advisor | Yann LeCun |
| PhD thesis | Learning Feature Hierarchies for Object Recognition (2010) |
| Earlier post | Research staff, NEC Laboratories America |
| Joined DeepMind | 2012 |
| Known for | DQN, AlphaGo contributions, WaveNet, leading DeepMind research |
| Role at Google DeepMind | Chief Technology Officer |
| Additional role | Chief AI Architect, Google (from June 2025), reporting to Sundar Pichai |