Google Brain
| Company type | Research group |
|---|---|
| Industry | Artificial intelligence, Machine learning, Deep learning |
| Founded | 2011 |
| Founder | Jeff Dean, Andrew Ng |
| Fate | Merged into Google DeepMind (April 2023) |
| Headquarters | Mountain View , United States |
Area served | Worldwide |
Key people | Jeff Dean, Geoffrey Hinton, Zoubin Ghahramani |
| Products | TensorFlow, Google Neural Machine Translation, Imagen, Parti, Magenta |
| Parent | Google AI (until April 2023) |
Google Brain served as Google's foremost AI research group specializing in deep learning before its integration into the larger Google AI organization. Established in 2011, the team merged advanced machine learning research with extensive computing infrastructure. It was responsible for creating significant technologies such as TensorFlow, which democratized access to neural network frameworks, and spearheaded numerous internal initiatives centered on machine learning and natural language processing. In April 2023, Google Brain combined forces with DeepMind, resulting in the formation of Google DeepMind.
History
The inception of Google Brain dates back to 2011 when it started as a part-time partnership between Google fellow Jeff Dean and researcher Greg Corrado.[1] Originally launched under Google X, its achievements prompted reintegration within Google proper; as noted by Astro Teller, Google Brain was responsible for covering all of Google X's costs.[2]
A landmark event occurred in June 2012 when a network of 16,000 processors spread over 1,000 machines successfully trained itself to identify a cat from a dataset of 10 million YouTube images, showcasing significant progress in neural network technology.[1] This breakthrough attracted widespread media attention, including coverage by National Public Radio.[3]
In March 2013, Google brought on board deep learning expert Geoffrey Hinton by acquiring his company DNNResearch Inc. Hinton divided his efforts between academic research and projects at Google.[4]
The April 2023 unification of Google Brain with DeepMind represents Google's strategic move to consolidate AI development efforts under the single entity Google DeepMind.[5]
Team and location
The founding figures behind Google Brain were Google Fellow Jeff Dean alongside Stanford professor Andrew Ng. By 2014, the team roster included Jeff Dean, Quoc Le, Ilya Sutskever, Alex Krizhevsky, Samy Bengio, and Vincent Vanhoucke. The lineup expanded by 2017 to feature Anelia Angelova, Samy Bengio, Greg Corrado, George Dahl, Michael Isard, Anjuli Kannan, Hugo Larochelle, Chris Olah, Benoit Steiner, Vincent Vanhoucke, Vijay Vasudevan, and Fernanda Viegas. Among notable members was Chris Lattner—known for creating Apple's Swift programming language and former Tesla autonomy head—who joined in August 2017 but departed in January 2020 to lead engineering at SiFive.
As of 2021, leadership consisted of Jeff Dean, Geoffrey Hinton, and Zoubin Ghahramani. Prominent contributors included Katherine Heller, Pi-Chuan Chang, Ian Simon, Jean-Philippe Vert, Nevena Lazic, Anelia Angelova, Lukasz Kaiser, Carrie Jun Cai, Eric Breck, Ruoming Pang, Carlos Riquelme, Hugo Larochelle, and David Ha. Samy Bengio exited in April 2021 with Zoubin Ghahramani taking over his duties.
Google Research—which encompasses Google Brain—is headquartered in Mountain View, California. It maintains additional offices globally in Accra, Amsterdam, Atlanta, Beijing, Berlin, Cambridge (Massachusetts), Israel, Los Angeles, London, Montreal, Munich, New York City, Paris, Pittsburgh, Princeton (New Jersey), San Francisco, Seattle, Tokyo, Toronto, and Zürich.
Artificial-intelligence-devised encryption system
In October 2016, Google Brain conducted an experiment illustrating that neural networks are capable of learning secure symmetric encryption. The experiment involved three neural networks named Alice, Bob, and Eve, structured within a generative adversarial network (GAN) framework. Alice was responsible for encrypting messages intended for Bob to decrypt, while Eve attempted to intercept and decipher these messages. Alice and Bob shared a secret key to perform encryption and decryption, maintaining a security advantage over Eve. This study highlighted the potential for neural networks to develop robust encryption techniques.
Image enhancement
In February 2017, Google Brain introduced a probabilistic technique to increase image resolution from 8x8 pixels up to 32x32 pixels. This method expanded upon the pixelCNN architecture and utilized two neural networks: the "conditioning network," which reduces high-resolution images down to 8x8 and maps them to higher resolutions, and the "prior network," which enriches the image by adding fine details based on these mappings. The resulting image at 32x32 resolution is generated as an estimation derived from existing high-resolution data rather than a straightforward enlargement. This advancement demonstrates neural networks' ability to improve image quality, with promising applications in digital imaging and media enhancement.
Google Translate
The Google Brain team improved Google Translate by creating a deep learning system that integrates artificial neural networks with vast multilingual text corpora. In September 2016, they introduced Google Neural Machine Translation (GNMT), an end-to-end model trained on extensive datasets. Unlike the previous Phrase-Based Machine Translation (PBMT) approach, which translated words individually, GNMT analyzes word sequences within their sentence context, significantly enhancing translation fidelity. This resulted in a 24% improvement in similarity to human translations and reduced errors by 60%, particularly benefiting difficult language pairs such as Chinese to English.
To extend these improvements across Google Translate's 103 supported languages, the team developed a Multilingual GNMT system capable of direct translation between multiple languages, including Zero-Shot Translations—enabling translation between language pairs without explicit training data. Additionally, Google introduced neural network models that translate spoken language directly into text in another language without needing intermediate transcription. This capability was achieved by training on extensive paired audio and text datasets, allowing efficient mapping and transformation of audio signals.
Nevertheless, GNMT's processing time grows exponentially with longer sentences, leading Google Brain to deploy an additional 2,000 processors to sustain performance and responsiveness.
Robotics
Google Brain is committed to advancing traditional robotics control by developing machine learning methods that empower robots to acquire new skills autonomously, minimizing reliance on manual programming. Their work also emphasizes cloud robotics, enabling robots to share knowledge and learn collectively through cloud-based platforms. In 2019, Google launched the Google Cloud Robotics Platform to unify robotics, artificial intelligence, and cloud computing for streamlined robotic automation.
Research efforts focus on leveraging deep learning algorithms to assist robots in performing tasks through experience accumulation, simulation environments, human demonstrations, and visual perception. Examples include teaching robots to grasp and throw objects without explicit code instructions and enabling them to pour liquids by learning from multi-angle human demonstration videos.
Google Brain collaborates extensively with industry leaders and academic institutions on robotics innovation. In 2016, they partnered with X to develop real-time robotic grasping systems featuring self-correction capabilities powered by deep learning and large-scale data analysis. In 2020, Google Brain joined forces with Intel AI Lab and UC Berkeley to build AI models that allow robots to learn surgical procedures such as suturing through video-based training.
Interactive Speaker Recognition Utilizing Reinforcement Learning
In 2020, researchers from Google Brain and the University of Lille unveiled Interactive Speaker Recognition (ISR), a system designed to identify speakers from a specific set by requesting only a limited number of personalized words. This ISR framework can be tailored to extract speech segments for Text-To-Speech model training, while also serving as a safeguard against unauthorized voice synthesis attempts.
TensorFlow
TensorFlow is an open-source software library created by Google Brain that facilitates the construction and training of neural networks for machine learning tasks. One notable application is in agriculture, where TensorFlow models trained on images sorted by humans automate crop classification, significantly decreasing manual effort.
Magenta
Magenta builds upon Google Brain’s technologies to generate novel artistic and musical content instead of simply categorizing data. Integrated with TensorFlow, it offers tools that steer neural networks toward creative production. However, research conducted by Valdosta State University reveals that AI still faces challenges in fully capturing human artistic intention, paralleling difficulties observed in automated translation.
Medical Applications
The image recognition prowess of Google Brain has been applied to medical diagnostics by detecting subtle patterns often overlooked by physicians, thereby enabling earlier detection of illnesses. For instance, in breast cancer screening, this technology achieved a false positive rate one-fourth that of human pathologists. Despite this success, the network’s specialized training restricts its ability to identify other conditions present within the same medical images.
Transformer
The transformer architecture for deep learning was introduced by Google Brain scientists in 2017 through the seminal paper Attention Is All You Need. Although Google holds patents on this critical AI framework, it has opted for a lenient enforcement approach, facilitating its widespread adoption across various sectors.
Text-to-Image Model
In 2022, Google Brain launched two text-to-image generation models named Imagen and Parti, positioning them as competitors to OpenAI's DALL-E. Later that year, the technology was extended to support text-to-video synthesis. Following Google's integration with DeepMind, the Imagen project transitioned under the management of Google Deepmind.
Additional Google Products
Technologies developed by Google Brain underpin several prominent Google services, including speech recognition on the Android Operating System, photo search functionality within Google Photos, smart reply features in Gmail, and video recommendation algorithms on YouTube.
Reception
Google Brain has been highlighted in publications such as Wired, NPR, and Big Think, featuring interviews with influential figures like Ray Kurzweil and Andrew Ng who discuss the initiative’s objectives and practical uses.
Controversies
In December 2020, AI ethicist Timnit Gebru departed from Google following disagreements related to her refusal to withdraw a paper titled "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" The study addressed concerns regarding AI expansion including environmental consequences, data biases, and risks of misleading the public. The request for retraction originated from Megan Kacholia, vice president of Google Brain. By April 2021, nearly 7,000 current and former employees along with supporters signed an open letter accusing Google of "research censorship" and condemning the treatment Gebru received.
In February 2021, Margaret Mitchell, head of AI ethics at Google, was dismissed for allegedly breaching company policies by employing automated tools in support of Gebru. This event prompted several engineers to resign in protest. In April 2021, co-founder Samy Bengio announced his departure; although citing personal reasons, insiders linked it to internal conflict within the ethics division. Bengio publicly expressed solidarity with both Gebru and Mitchell.
In March 2022, Google terminated AI researcher Satrajit Chatterjee after he questioned findings published in a Nature article authored by Google's Anna Goldie and Azalia Mirhoseini. The study claimed advancements in AI methods for integrated circuit placement but faced criticism due to insufficient empirical evidence, lack of direct comparisons with existing techniques, and reliance on proprietary datasets hindering reproducibility. Subsequent investigations led to retractions and ongoing scrutiny by Nature. Media reports raised doubts about Google's transparency regarding research results and benchmark disclosures.
California Judge Frederick Chung ruled that Chatterjee provided adequate proof that Google retaliated against him for refusing participation in unlawful activities. His lawsuit alleges that the company exaggerated claims about its AI chip design capabilities.
See also
- Artificial intelligence art
- Glossary of artificial intelligence
- List of artificial intelligence projects
- Noosphere
- Quantum Artificial Intelligence Lab – a joint effort involving Google, NASA, and the Universities Space Research Association