Google Brain (literally translated from Arabic as Google mind) was a deep learning artificial intelligence research team and project under Alphabet Inc. and Google that served as Google's sole AI branch before being incorporated under the Google AI research division (Google Research).^[2] Headquartered in Mountain View, California, United States, the team operated worldwide and maintained additional locations in Cambridge (Massachusetts), London, Montreal, New York City, San Francisco, Toronto, and Zurich.^[49,2] Google Brain was created in early 2010, the early 2010s, or formed in 2011 as a research collaboration initially established by Google Fellow Jeff Dean, Google Researcher Greg Corrado, and Stanford University Professor (and visiting professor) Andrew Ng.^{[1,2,3,48,137,143]} Ng had been interested in using deep learning techniques to solve the problem of artificial intelligence since 2006, and in 2011 began collaborating with Dean and Corrado to build DistBelief, a large-scale deep learning software system on top of Google's cloud computing infrastructure. Google Brain started as a Google X project and became so successful that it graduated back into Google—with Astro Teller stating that Google Brain paid for the entire cost of Google X—subsequently growing into one of the most influential actors in AI research.^[11,143]

Google Brain combined open-ended machine learning research, neural networks, and systems engineering with information systems and Google-scale computing resources, focusing on creating research opportunities in machine learning, natural language processing, and machine translation.^{[1,2,3,48,5,4,74,111,83,137,147]} In 2012, the team achieved its first major result by developing a software-based neural network of 16,000 processors that proved capable of independently recognizing the image of a cat.^[8] The team developed tools such as TensorFlow to enable the public to build and use neural networks, designed the transformer architecture, and carried out numerous internal and external AI projects spanning Google Translate, AI-designed encryption systems, image enhancement, robotics, text-to-image models (Imagen), automatic speech recognition and interactive speaker recognition with reinforcement learning, and medical applications.^{[5,4,74,111,126,50,83,48]}

In 2014, the team included Jeff Dean, Quoc Le, Ilya Sutskever, Alex Krizhevsky, Samy Bengio, and Vincent Vanhoucke. In 2017, team members included Anelia Angelova, Samy Bengio, Greg Corrado, George Dahl, Michael Isard, Anjuli Kannan, Hugo Larochelle, Chris Olah, Salih Edneer, Benoit Steiner, Vincent Vanhoucke, Vijay Vasudevan, and Fernanda Viegas.^[25] Chris Lattner, creator of Apple's Swift programming language and former leader of Tesla's autonomy team, joined Google Brain in August 2017 and left in January 2020 to join SiFive.^[13] Interviews with Google Brain members appeared in media outlets including Wired, National Public Radio, and Big Think, and the team maintained websites at research.google/brain and ai.google/brain-team/.^[39] In April 2023, Google Brain merged with sister company and former subsidiary DeepMind (DeepMind Technologies) to form Google DeepMind.^{[3,4,115,125,142]}

The mission of Google Brain is to improve human lives by making machines smarter.^[116] To achieve this, the team focuses on building highly flexible models capable of independently learning their own features and making optimal and efficient use of data and computing power.^[138] According to Persian and Portuguese sources, this approach fits into the deep learning subfield of machine learning, whereas Russian sources state that it is part of a broader field of machine learning than deep learning, with sources agreeing that it ensures their work solves problems of practical importance. Furthermore, the team's expertise in systems complements this approach, allowing them to build tools that accelerate machine learning research and unlock its practical value for the world.^[116,4,5]

The Google Brain project began in 2011 as a part-time research collaboration between Google Fellow Jeff Dean, Google researcher Greg Corrado, and Stanford University professor Andrew Ng.^[1,50,4,5] Andrew Ng had been interested since 2006 in using deep learning techniques to solve artificial intelligence problems, which led him to collaborate with Dean and Corrado in 2011 to build DistBelief, a large-scale deep learning software system built on top of Google's cloud computing infrastructure.^[6,7,5,117] Google Brain originally started as an ambitious project within Google X, Google's experimental research division, focusing on artificial intelligence and machine learning by exploring the potential of training artificial neural networks with large amounts of data. The project achieved remarkable success and graduated from Google X into Google's main organization as a standalone department, with Astro Teller stating that Google Brain's financial returns covered the entire cost of Google X.^{[7,8,5,117,6]}

In June 2012, The New York Times, followed by National Public Radio and SmartPlanet, reported that a computing cluster dedicated to mimicking human brain activity had successfully taught itself to recognize high-level concepts such as cats from 10 million digital images extracted from YouTube videos without prior instruction, though accounts differ on whether the cluster comprised 16,000 processors across 1,000 computers or 16,000 separate computers.^{[5,8,52,53,6,9,10,117,118,132,115,71]} Early advances from the project also included improving the accuracy of Google's voice search service. In 2013, Andrew Ng left Google to found the online training startup Coursera, and subsequently became the head of Baidu's artificial intelligence group in 2014.^[5]

In March 2013, Google expanded its deep learning research by hiring leading researcher Geoffrey Hinton and acquiring DNNResearch Inc., a company he had established with Alex Krizhevsky and Ilya Sutskever; Hinton stated that he would split his time between university research and his work at Google.^{[9,76,75,119,120,115,10]} Later in 2013, Tomáš Mikolov and his colleagues at Google Brain developed word2vec, a method for learning word embeddings from large collections of text.^[77] Following negotiations between Facebook and DeepMind Technologies in 2013 that concluded without an agreement, Google acquired DeepMind on January 26, 2014, for an undisclosed price, with subsequent reports placing the value at between £400 million and over £500 million ($650 million or €486 million), whereas Chinese media reported the valuation at 4 million to 5 million pounds ($6.5 million).^{[120,72,121,122,123,8,73,40]} In 2015, Google open-sourced TensorFlow, a deep learning framework developed by the Google Brain team that, along with libraries like JAX, made machine learning development more accessible.^[8] In 2017, Google researchers introduced the Transformer architecture in the paper Attention Is All You Need, which was initially developed for machine translation and went on to revolutionize natural language processing, becoming the dominant architecture for large language models.^[10,92] Google Brain's breakthroughs were applied across multiple Google services, including Google Search, Google Translate, and Google Assistant.

Google Brain was initially established by Google Fellow and researcher Jeff Dean and Stanford University visiting professor Andrew Ng, who later left the project to lead the artificial intelligence group at Baidu.^[127,3] In 2014, the team included Jeff Dean, Quoc V. Le, Ilya Sutskever, Alex Krizhevsky, Samy Bengio, and Vincent Vanhoucke. By 2017, team members included Anelia Angelova, Samy Bengio, Greg Corrado, George Dahl, Michael Isard, Anjuli Kannan, Hugo Larochelle, Quoc Le, Chris Olah, Salih Idenir (Salikh Edner), Benoit Steiner, Vincent Vanhoucke, Vijay Vasudevan, and Fernanda Viégas.^[10,11]

Chris Lattner, the creator of Apple's Swift programming language who subsequently led Tesla's Autopilot and autonomous driving team for six months, joined Google Brain in August 2017, shifting his focus to intelligence projects.^[11,12,3] Lattner left Google Brain in January 2020 to join SiFive, where he focused on semiconductor design and development.^[12,13] In 2021, Google Brain was led by Jeff Dean, Geoffrey Hinton, and Zoubin Ghahramani. Other team members around that time 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.^[11] Samy Bengio left the team in April 2021, after which Zoubin Ghahramani assumed his responsibilities.^[13]

The Google Brain Residency Program was a 12-month program geared toward individuals eager to apply their passion to machine learning and artificial intelligence, offering hands-on experience and collaboration with researchers at the forefront of the field. Participants included recent graduates holding bachelor's or doctoral degrees in computer science, physics, mathematics, and neuroscience, as well as individuals with years of industry experience. The program spanned diverse topics including algorithm applications, natural language understanding, robotics, neuroscience, and genetics, with residents making major research contributions in just a few months. Published technical papers resulting from the residency program include "Unrolled Generative Adversarial Networks", "Conditional Image Synthesis with Auxiliary Classifier GANs", "Regularizing neural networks penalizing their output distribution", "Mean Field Neural Networks", "Learning to remember", "To generate high-resolution images with generative adversarial networks", "Multi-task convolutional music models", and "Audio DeepDream: Optimizing Raw Audio with Convolutional Networks".

Google Brain is part of Google Research and is headquartered in Mountain View, California.^[7] It also maintains satellite groups and branches in Accra, Amsterdam, Atlanta, Beijing, Berlin, Cambridge (Massachusetts), Israel, Los Angeles, London, Montreal, Munich, New York City, Paris, Pittsburgh, Princeton, San Francisco, Seattle, Tokyo, Toronto, and Zurich.^{[14,78,15,7,148]}

The Google Brain team's slogan is "Make machines smarter to improve the quality of human life." The team believes that knowledge should be shared by all humankind, so all of its scientific research results are publicly published.

Google Brain conducts research across a broad spectrum of artificial intelligence areas, including seven main research directions: machine learning algorithms and techniques, healthcare, machine learning supporting computer systems, robotics, natural language understanding, music and art creation, and perception simulation. In deep learning, the team focuses on the development of new algorithms, architectures, and training techniques for neural networks. Its work in natural language processing (NLP) involves teaching computers to understand and generate human language. In computer vision, Google Brain works on teaching computers to "see" and interpret images and video. Its research in robotics is centered on developing intelligent robots that can learn and interact with their environment. Additionally, its research in reinforcement learning focuses on teaching agents to make decisions and solve problems in complex environments.

In October 2016 (reported as November 6, 2016 in French sources), Google Brain conducted an experiment related to message and communications encryption to determine whether neural networks are capable of learning secure symmetric encryption.^{[15,79,16,80]} In this experiment, three neural networks (or artificial intelligences) were created: Alice, Bob, and Eve.^[16,80,78] Adhering to the concept of a generative adversarial network (GAN), the goal was for Alice to send an encrypted message to Bob that Bob could decrypt, while the adversary Eve would simultaneously try to intercept the message and evolve her own system to crack the AI-generated encryption.^{[80,16,79,17]} Alice and Bob maintained an advantage over Eve because they shared a key used for encryption and decryption.^[15,79]

The AIs were not given clear instructions or pre-written cryptographic algorithms on how to encrypt their messages, and were only provided with a loss function.^[16] Consequently, if communication between Alice and Bob was unsuccessful during the experiment—with Bob misinterpreting Alice's message or Eve intercepting it—the cryptography would evolve in subsequent rounds so that Alice and Bob could communicate securely. The study proved successful, demonstrating that neural networks can learn to develop their communications from scratch and create their own encryption systems without pre-prescribed algorithms, which could represent an advance for message encryption in the future.^{[15,79,16,124,80]}

In February 2017, Google Brain announced an image enhancement system and identified a probabilistic method to convert images with an 8x8 resolution to a 32x32 resolution, using neural networks to fill in details in very low-resolution images.^{[16,17,81,139,18,19]} Building upon an existing probabilistic model known as PixelCNN, the software generates pixel translations and utilizes two different neural networks to approximate the pixel composition.^{[18,19,13,20,21,82]} The first network, known as the "conditioning network", downsizes high-resolution images to an 8x8 resolution and attempts to create mappings or find a match from the original low-resolution image to these higher-resolution ones.^{[17,18,21,19]} The second network, known as the "prior network", analyzes the pixelated image and uses the mappings to add details based on a large number of high-resolution images.^[17,18] When upscaling the original 8x8 image, the system adds pixels based on its knowledge of what should be in the picture, and the outputs of both networks are combined to create the final translated image.^[21,16] The resulting image is not the original image at a higher resolution, but rather a 32x32 resolution estimation and best guess based on other existing high-resolution images.^[17,18,19] Representing a breakthrough that demonstrated the potential of neural networks to enhance low-resolution images, the technology showed impressive results in real-world tests: when presented with the enhanced and original images, human observers were fooled 10% of the time on celebrity faces and 28% of the time on bedroom photos, whereas conventional bicubic scaling failed to fool any humans.^{[21,82,16,20,19,66,22]}

In December 2017, Google Brain introduced Google NIMA (Neural Image Assessment; German: neurale Bildbewertung), which uses a deep convolutional neural network capable of providing a subjective evaluation of images based on their aesthetics. Strongly resembling human decision-making as the most advanced and comprehensive software in this field to date, it can assist humans with searches and filter large media datasets.^[60,61]

The Google Brain team contributed to the Google Translate project by employing a new deep learning system that combines artificial neural networks with vast multilingual text databases.^[22,23] In September 2016, Google Neural Machine Translation (GNMT) was launched as an end-to-end learning framework capable of learning from a large amount of training data, aiming to enable more accurate, human-like translations and eliminate the need for manual input of word data.^[66,22,23] Previously, Google Translate used Phrase-Based Machine Translation (PBMT), which statistically analyzed text word-by-word to match corresponding words without considering the surrounding phrases in a sentence.^[23,24] In contrast, GNMT evaluates word segments within the context of the rest of the sentence to select more accurate substitutes.^[22,111,147] Compared to older PBMT models, GNMT achieved a 24% improvement in similarity to human translation and reduced errors by 60%, showing significant improvement even in notoriously difficult translations such as Chinese to English.^[21]

While the introduction of GNMT greatly improved translation quality for pilot languages, extending these improvements across all 103 supported languages proved difficult. To address this challenge, the Google Brain team developed the Multilingual GNMT system, which expanded upon the previous framework by enabling translations between multiple languages. This system also enables zero-shot translations, allowing translations between language pairs that the model had never explicitly encountered during training.^{[23,24,22,26]}

Google also announced that Google Translate can perform direct translation using neural networks without transcription, translating speech in one language directly into text in another language without an intermediate speech-to-text conversion.^[140] To train this capability, researchers exposed the system to many hours of Spanish audio paired with corresponding English text, allowing neural network layers that mimic the human brain to link corresponding fragments and manipulate the audio waveform into English text.^[24,25,27] Another drawback of the GNMT model was that translation time increased significantly and exponentially with the number of words in a sentence, prompting the Google Brain team to add 2,000 additional processors to ensure translation remained fast and reliable.^[16,23,24]

In 2017, Google Brain researchers invented the Transformer deep learning architecture, described in the scientific paper "Attention Is All You Need"; Google holds a patent on this widely used architecture but has not enforced it.^[93,94] In February 2018, Google Brain researchers presented an algorithm designed to collect and extract information from multiple texts to write natural-language encyclopedia articles, such as for Wikipedia. However, the developers noted that the generated quality still deviates significantly from human authors, that the algorithm struggles when handling large numbers of sources, and that it is incapable of distinguishing between trustworthy and untrustworthy information.^[67]

Aiming to improve traditional robotics control algorithms where new skills must be manually programmed, robotics researchers at Google Brain are developing machine learning techniques to allow robots to automatically learn and acquire new skills on their own.^[25,26,28] They are also working to develop methods for information sharing among robots so that they can learn from each other during the learning process, an approach known as cloud robotics.^{[26,27,28,85]} As a result, Google launched the Google Cloud Robotics Platform for developers in 2019, combining robotics, artificial intelligence, and cloud computing technologies to enable efficient robotic automation through cloud-connected collaborative robots.^{[26,28,84,27,85]}

Robotics research at Google Brain has focused primarily on improving and applying deep learning algorithms to enable robots to complete tasks by learning from experience, simulations, human demonstrations, and visual representations or data.^{[28,85,86,87,16,30,31,88,32]} For example, Google Brain researchers demonstrated that robots can learn to pick up solid, rigid objects and toss or place them into designated boxes or bins through trial-and-error experimentation in an environment without being pre-programmed to do so.^[27,28] In other research, robots were trained to learn behaviors such as pouring liquid from a cup, learning from video recordings of human demonstrations captured from multiple viewpoints.^[29,86,88] In 2017, the team established three goals and explored three approaches for learning new skills—reinforcement learning, interaction with objects, and human demonstration—aiming to create robots capable of solving more complex tasks than industrial robots through learning and practice.^[62]

Google Brain researchers have also collaborated with other companies and academic institutions on robotics research. In 2016, the Google Brain team collaborated with researchers at X (Google X / X Development) on self-learning robots and hand-eye coordination for robotic grasping, in which robots made about 800,000 grasping attempts to demonstrate how robotics could use experience to teach itself more efficiently.^[31,88,112] Their method enabled real-time robot control to grasp novel objects with self-correction.^[31,88,16] In 2020, researchers from Google Brain, Intel AI Lab, and UC Berkeley developed an artificial intelligence model that allows robots to learn surgery-related tasks, such as suturing, by training on videos of surgical procedures.^[13,87,32]

In 2020, a team from Google Brain and the University of Lille (identified as the University of Lille I in Spanish Wikipedia) in France presented an automatic speaker recognition model named Interactive Speaker Recognition (ISR). The ISR module (or unit) identifies and recognizes a speaker from a given list of speakers solely by requesting or querying a few user-specific words or utterances.^[32,89,36] The model can be modified or altered to select speech segments in the context of text-to-speech (speech synthesis) training.^{[16,90,32,36]} It can also protect data by preventing malicious voice generators and attackers using speech generators from accessing it.^{[16,89,32,36]}

TensorFlow is an open-source software library supported by Google Brain that allows anyone to utilize machine learning by providing tools to train their own neural networks.^{[74,111,112,147]} The tool has been used to develop software and applications utilizing deep learning models that help farmers reduce the amount of manual labor required to sort their crops and produce, accomplished by training a neural network on a dataset of images previously sorted by humans.^[22,111,147]

Magenta is a project that uses Google Brain to create new information in the form of art and music rather than classifying and sorting existing data.^{[22,74,111,147]} TensorFlow was updated with a suite of tools for users to guide, program, and train neural networks to create images and music.^[22,74] However, a team from Valdosta State University found that artificial intelligence struggles to perfectly reproduce human intention in art, similar to the issues faced in translation.^[22]

Google Brain's image sorting and classification capabilities have been used to help detect certain medical conditions by searching for patterns and features that human doctors might not notice to provide an earlier diagnosis.^[22,111] During breast cancer screening, this method was found to have a quarter of the false positive rate (75% less likely to make a false positive diagnosis) compared to human pathologists, who require more time to review each image and cannot dedicate their entire focus to this single task.^[22,74,147] Due to the neural network's highly specific training for a single task, it cannot identify other ailments or conditions present in an image that a human specialist could easily detect.^[22,74,111]

Google Brain is also attempting to determine the time of death of patients with the help of deep learning. For this purpose, the system initially examined 216,221 records of patients who received inpatient treatment for at least 24 hours, utilizing data received from the University of California San Francisco Medical Center and University of Chicago Medicine. Approximately 46 million data points were extracted from the records and their associated medical reports. This allows the software to predict an in-hospital mortality prognosis, a potential readmission within 30 days, or an extended length of stay, enabling doctors to make decisions significantly earlier. In addition, the software is capable of learning through the artificial neural network and has performed best among comparable software solutions so far.^[55]

In 2022, Google Brain announced that it had created two different types of text-to-image models called Imagen and Parti, which compete with OpenAI's DALL-E.^[31,33,34] Later in 2022, at the end of the year, the project was expanded to text-to-video.^[33,35] Following the merger with DeepMind, the development of Imagen was transferred to Google DeepMind.^[95] An example image was created by Imagen 3.0.

Google Brain project technology is currently used across various Google products and the Android operating system, notably for the speech recognition system in Android and for video recommendations on YouTube, where it has played a crucial role in intelligent video recommendations.^{[34,36,65,63,64,128,129,38,76]} Sources differ regarding photo search implementation, stating that the technology is used either for photo search in Google Photos or for the photo search function within Google's social platform Google+.^{[34,36,65,63,64,128,129,38,76]} Additionally, the technology powers Smart Reply in Gmail.^{[34,36,76,65,38]}

Google Brain was initially formed by Google senior researcher Jeff Dean and Stanford professor Andrew Ng, who later joined Baidu's Artificial Intelligence Division.^[127] As of 2014, team members included Jeff Dean, Geoffrey Hinton, Greg Corrado, Quoc Le, Iiya Sutskever, Alex Krizhevsky, Samy Bengio, and Vincent Vanhoucke. Currently, multiple Chinese nationals or people of Chinese descent work on the Google Brain team.

The Google Brain project received significant, thorough coverage across various media outlets, including Wired Magazine, the New York Times, Technology Review, National Public Radio (NPR), and Big Think.^{[39,42,70,72,118,8,9]} These articles featured interviews with key team members Ray Kurzweil and Andrew Ng, focusing on explanations of the project's goals, objectives, and applications.^{[8,37,40,39,9,42,10,41,149]}

Google Brain has had a profound impact on the development of AI. Their research has led to significant advances in fields such as machine translation, image generation, and game AI. Furthermore, their tools and libraries have democratized AI, enabling a broader audience to experiment with and apply deep learning.^[146]

In December 2020, AI ethics expert Timnit Gebru left Google.^[41,96,16] While the exact nature of whether she resigned or was dismissed is disputed, the cause of her departure was her refusal to retract a paper titled "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" and a related ultimatum setting conditions to be met or she would leave.^{[41,9,96,44,45]} This paper explored the potential risks of AI growth such as Google Brain, including environmental impact, biases in training data, and the ability to deceive the public.^{[41,9,96,44,45]} The request to retract the paper was made by Megan Kacholia, Vice President of Google Brain.^{[42,45,136,46]} As of April 2021, nearly 7,000 current or former Google employees and industry supporters signed an open letter accusing Google of "research censorship" and condemning Gebru's treatment at the company.^{[43,46,91,97,47]}

In February 2021, Google fired Margaret Mitchell, one of the leaders of the company's AI ethics team.^{[13,47,45,46]} The company's statement claimed that Mitchell had violated company policy by using automated tools to find support for Gebru.^{[42,47,45,46]} In the same month, engineers from outside the ethics team began resigning, citing Gebru's dismissal as the reason for their departure.^[44,98,14] In April 2021, Google Brain co-founder Samy Bengio announced his resignation from the company.^[14,99,15] Despite being Gebru's manager, Bengio had not been notified prior to her termination, and he posted online in support of both her and Mitchell.^[13,14,15] While Bengio's announcement focused on personal reasons and growth as the reason for leaving, anonymous sources indicated to Reuters that turmoil within the AI ethics team influenced his decision.^[14,15]

In March 2022, Google fired AI researcher Satrajit Chatterjee after he questioned the findings of a paper published in Nature by Google AI team members Anna Goldie and Azalia Mirhoseini.^{[101,102,113,100]} The paper claimed that their AI techniques, in particular reinforcement learning, for the integrated circuit placement problem were superior to prior methods.^{[101,102,114]} However, this claim is contested because the alleged results, especially rapid chip design, were not adequately supported by specific empirical data and were found to be inconsistent with subsequent published research.^{[103,104,105,101,102]} The paper did not report run times of prior and proposed methods on specific inputs, lacked head-to-head comparisons to sufficiently advanced implementations of prior methods, and was difficult to replicate due to proprietary training and test data.^{[104,106,105,103]} At least one initially favorable commentary was retracted upon further review, and the paper was placed under investigation by Nature editors.^[46,47] Additional media coverage expressed skepticism about the research published by Google and noted that Google had not provided the benchmarks long requested by experts.^[47,108]

In a lawsuit contesting his firing, Chatterjee alleged that the Google research paper overhyped and exaggerated the technology.^[109] Google moved to dismiss the lawsuit, but California Judge Frederick Chung ruled that Chatterjee had "adequately supported his claim that Google terminated him in retaliation for refusing to participate in an act that would violate state or federal law."^[47] Nature published an addendum in September 2024, where the Google engineers clarified their methodology and announced that they had open-sourced a software repository to reproduce the methods, but not the proprietary data of their examples.^[108] However, this did not quell the controversy over whether the proposed techniques are in fact an advance over existing methods, as Google did not show comparative results on publicly available benchmarks, the standard in the field and evidence long requested by experts.^[109]