Google Brain

Google Brain was a machine learning research team at Google. It began in 2011 as a project at X, Google's research and development organization, and later operated within Google Research. On April 20, 2023, Google combined the Brain team with DeepMind to create Google DeepMind. Google Brain therefore refers to a historical organization, not the name of Google's current general-purpose AI laboratory.^[1][38][39]

The team worked on research, software, and production applications of deep learning. Its documented work included the DistBelief distributed training system, TensorFlow, word2vec, sequence-to-sequence learning, Google Neural Machine Translation, the Transformer architecture, neural architecture search, large-scale robot learning, and several later vision and generative-model projects.^{[7][10][12][15][18][30][32][36]} Some adjacent projects were produced by other units of Google Research or by collaborations spanning several Google teams. Organizational attribution is consequently narrower than a list of every paper written by a Google employee.

Google Brain was distinct from DeepMind before the 2023 reorganization. It was also narrower than Google Research, which covered many areas of computer science outside the Brain team's machine learning program. Work published after the reorganization belongs to Google DeepMind unless a source is describing the earlier Brain affiliation or technical lineage.^[20][38][40]

Organizational identity

Google DeepMind's institutional history states that Google Brain started at X in 2011 and explored how modern machine learning could affect Google's products and services.^[1] Three institutional profiles identify the principal founding figures:

Scroll sideways for more →

These descriptions establish their founding roles without implying that the project had a single fixed management structure throughout its twelve-year history. Public sources more consistently document technical leadership, paper authorship, and dated team affiliations than a complete year-by-year organization chart.

The name "Google Brain" also changed meaning as the project grew. Early papers list authors from Google and Stanford rather than a formal Brain affiliation. By 2016 and 2017, Google's research blog routinely identified authors and programs as part of the Google Brain team.^[15][20] In 2018, Google renamed its public Google Research channels under the broader "Google AI" label, but that communications change did not end the Brain team. Google continued to publish work with the affiliation "Google Research, Brain Team" through 2022.^[23][36][37]

Origins and early large-scale learning

DistBelief

The early project tested whether neural networks could be trained across Google's computing infrastructure at a scale that was difficult to reach in academic laboratories. Its first major systems result was DistBelief, a framework for distributing neural network training over clusters with thousands of machines. The associated 2012 NeurIPS paper described models with billions of parameters and training across tens of thousands of CPU cores.^[7][8]

DistBelief supported model parallelism, in which parts of one model run on different machines, and data parallelism, in which replicas process different training examples. The system included Downpour SGD, an asynchronous stochastic-gradient method, and Sandblaster, a framework for distributed batch optimization. The paper reported training a network 30 times larger than previously published systems and accelerating a smaller network used for commercial speech recognition.^[7][8]

DistBelief remained an internal Google system. Its tight coupling to Google's infrastructure and its focus on neural networks later became explicit design constraints for TensorFlow, the team's second-generation system.^[10]

Large-scale unsupervised feature learning

In 2012, Quoc V. Le, Marc'Aurelio Ranzato, Rajat Monga, Matthieu Devin, Kai Chen, Greg Corrado, Jeff Dean, and Andrew Ng published "Building High-level Features Using Large Scale Unsupervised Learning." The researchers trained a nine-layer locally connected sparse autoencoder on 10 million 200 by 200 pixel images sampled from YouTube videos. The model had about one billion connections and trained for three days on 1,000 machines, or about 16,000 CPU cores.^[5][6]

The training images had no object labels. When the researchers examined features in the trained network, they found units selective for human faces, cat faces, and human bodies. They then used the learned representation for classification on 20,000 ImageNet categories and reported 15.8 percent accuracy, a 70 percent relative improvement over the previous result cited in the paper.^[6]

Popular accounts often call this work the "cat experiment," but the paper's stated research question was broader: whether high-level, class-specific detectors could emerge from unlabeled data at large scale. The experiment did not show that the system understood cats, and it was not a general object-recognition service. It tested learned feature selectivity and downstream classification under the paper's experimental setup.^[5][6]

DNNresearch acquisition

Google expanded its deep learning research in March 2013 by acquiring DNNresearch, a company incorporated in 2012 by University of Toronto professor Geoffrey Hinton and graduate students Alex Krizhevsky and Ilya Sutskever. The University of Toronto reported that Krizhevsky and Sutskever would move to Google, while Hinton would divide his time between the university and Google. The financial terms were not disclosed.^[9][43]

The acquisition is relevant to Brain's personnel history, but it should not be used to rewrite earlier academic work as Google Brain research. AlexNet, for example, was completed at the University of Toronto before the acquisition. Later work can be attributed to Brain where a team source or paper supports that affiliation. Google's 2016 Brain review, for example, identifies Sutskever, Oriol Vinyals, and Quoc V. Le as three Brain researchers behind the 2014 sequence-to-sequence paper.^[15]

Research model and team programs

Jeff Dean described the team's 2017 mission as "Make machines intelligent. Improve people's lives." The same account says that researchers pursued both fundamental machine learning questions and work conducted with product teams. It lists visual and auditory perception, natural-language understanding, art and music generation, systems, and algorithms among the group's research areas.^[20]

The team used conference publication, open-source releases, and external collaboration as parts of its research model. In a September 2017 account, Dean wrote that one third of the team's papers that year had at least one author from another institution. He also described Distill, TensorFlow Playground, and the People + AI Research group's deeplearn.js as examples of tools intended to make machine learning research easier to inspect or use.^[20]

Google Brain created a residency program in 2016 for people moving into machine learning research. The first cohort had 27 residents. Google's 2017 account reported that they co-authored more than 40 papers, while the second cohort began with 36 residents in July 2017.^[20] The program later expanded beyond Brain and became the Google AI Residency. Google's 2018 program review explicitly describes that change and says residents were then placed across multiple Google teams and offices.^[24]

The size and output figures in Google's annual summaries are dated snapshots, not permanent characteristics of the organization. The 2017 year-end review reported 140 papers for that year, including more than 60 at ICLR, ICML, and NIPS, and described team members across North America and Europe.^[22] Those figures should not be extrapolated to other years or used as a current headcount.

Systems and infrastructure

TensorFlow

Google released TensorFlow as open-source software on November 9, 2015. The release announcement called it Google's second-generation machine learning system and directly contrasted it with DistBelief. TensorFlow used dataflow graphs, supported heterogeneous devices, and was designed to run from mobile devices to distributed clusters. Google released the reference implementation under the Apache 2.0 license.^[10]

The 2016 OSDI paper lists the authors' affiliation as Google Brain. It describes TensorFlow as a system for mapping computation across machines and devices that could include multicore CPUs, GPUs, and custom accelerators. It also documents use in research and production, rather than limiting the system to neural network experiments.^[11]

TensorFlow was a genuine Brain project, but later TensorFlow development involved a much wider open-source community and multiple Google organizations. Describing every later TensorFlow component as Brain work would therefore be inaccurate. The historical claim is that Brain developed and released the original system, not that the team alone authored every subsequent release.^[10][11]

Tensor Processing Unit collaboration

Google Brain contributed to the requirements for Google's first Tensor Processing Unit, but the chip was a cross-organizational hardware project. The team's 2016 review says Brain worked with Google's platforms team to provide specifications and high-level goals for the custom machine learning accelerator.^[15] It does not say that Brain alone designed or fabricated the chip.

The 2017 architecture paper describes the first TPU as a datacenter inference ASIC deployed since 2015. Its central matrix unit contained 65,536 8-bit multiply-accumulate units, with a peak throughput of 92 tera-operations per second and 28 MiB of software-managed on-chip memory. The paper evaluated production neural network workloads and compared the chip with contemporary server CPUs and GPUs.^[33]

Google's 2016 Brain review reported TPU use for RankBrain, neural machine translation, and the AlphaGo match against Lee Sedol. AlphaGo was a DeepMind system. The TPU deployment was an infrastructure contribution to that system, not evidence that AlphaGo was a Brain project.^[15]

Pathways

Pathways was a later Google Research system for orchestrating computations across large accelerator clusters. Its MLSys paper describes a sharded asynchronous dataflow design, support for heterogeneous parallel computations, and experiments across thousands of TPUs.^[34] Jeff Dean was a co-author, but the paper does not present Pathways as the work of Brain alone.

The distinction matters because Pathways linked several layers of Google's research stack. It drew on distributed-systems work associated with longtime Brain researchers, while also serving broader Google Research model programs. It is best described as part of Brain's late technical lineage and Google Research's shared infrastructure, rather than as an exclusively Brain-owned system.^[2][34]

Language and sequence modeling

Word2vec

Tomas Mikolov, Kai Chen, Greg Corrado, and Jeff Dean introduced efficient continuous bag-of-words and skip-gram models for learning word vectors from large text corpora in 2013. The paper reported that high-quality vectors could be learned from a 1.6-billion-word dataset in less than a day and evaluated relationships in the learned vector space.^[12]

A follow-up NeurIPS paper by Mikolov, Ilya Sutskever, Chen, Corrado, and Dean added subsampling of frequent words, negative sampling, and a method for learning phrase representations. The paper also states an important limitation: word-vector composition does not automatically recover the meanings of idiomatic phrases.^[13] The work became known collectively as word2vec, although that name refers to the models and associated implementation rather than one single paper.

Sequence-to-sequence learning

The 2014 NeurIPS paper "Sequence to Sequence Learning with Neural Networks" introduced an end-to-end method in which one multilayer LSTM encoded an input sequence into a fixed-dimensional vector and another LSTM decoded the output sequence. On the WMT 2014 English-to-French task, the authors evaluated both direct translation and reranking of a phrase-based system's hypotheses.^[14]

Google's 2016 Brain review identifies the authors, Ilya Sutskever, Oriol Vinyals, and Quoc V. Le, as Brain researchers and connects the paper to later machine translation work.^[15] Sequence-to-sequence learning was not the only encoder-decoder approach developed in 2014, and later attention mechanisms substantially changed the design. The Brain paper's documented contribution was a general neural method for mapping variable-length sequences to variable-length sequences.

Google Neural Machine Translation

In September 2016, Google announced the Google Neural Machine Translation system, developed through collaboration between Brain researchers and the Google Translate team. The technical report describes an eight-layer LSTM encoder and eight-layer decoder, attention, residual connections, wordpiece units, and low-precision inference.^[15][16][17]

On Google's human side-by-side evaluation of isolated sentences, the report found an average 60 percent reduction in translation errors compared with the previous phrase-based production system. That number describes the paper's evaluation and should not be generalized to every language pair, document type, or later version of Google Translate.^[17]

The 2016 Brain review says the system replaced the previous production algorithms for a growing set of language pairs and separately notes a multilingual model that demonstrated zero-shot translation between language pairs absent from its direct training examples.^[15]

Transformer

The 2017 paper "Attention Is All You Need" introduced the Transformer, an encoder-decoder architecture based on attention rather than recurrence or convolution. Its multi-head self-attention layers allowed positions in a sequence to exchange information while enabling substantially more parallel computation during training than recurrent models.^[18][19]

The paper was written by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Google's technical post attributes the work to those eight researchers and explains the architecture in the context of machine translation.^[18] Google's 2023 reorganization announcement lists Transformers among the collective accomplishments of the Brain and DeepMind organizations, while DeepMind's own announcement specifically associates Transformers with the Brain side of the combined history.^[38][39]

The original paper evaluated machine translation. Later encoder-only, decoder-only, vision, audio, and multimodal systems adapted the architecture in different ways. It is accurate to trace those architectures to the paper, but not to treat every Transformer-based model as a Google Brain project.

BERT and the limits of team attribution

BERT illustrates why Google Research and Google Brain should not be treated as synonyms. Its November 2018 release post identifies the authors as researchers in Google AI Language, and the NAACL paper gives the affiliation "Google AI Language."^[25][26] The model used masked-language-model pretraining to learn bidirectional Transformer representations and reported results on eleven natural-language processing tasks.^[25][26]

Google DeepMind's later institutional history includes BERT in its account of Brain's technical lineage.^[1] Both records can be reported without erasing the organizational ambiguity: BERT was a Google Transformer project announced by Google AI Language, and Google's successor organization places it in Brain's historical line. The original publication does not support describing it simply as an exclusively Brain-authored system.

Vision, robotics, and generative work

Inception and batch normalization

Brain's TensorFlow release post states that DistBelief trained the Inception model that won the 2014 ImageNet classification challenge.^[10] The associated "Going Deeper with Convolutions" paper introduced the Inception architecture and reported the competition result.^[27] The paper involved a large group of Google and external contributors, so the safest attribution is to the Google research collaboration rather than to one individual.

Sergey Ioffe and Christian Szegedy's 2015 batch-normalization paper proposed normalizing layer inputs using mini-batch statistics to make deep-network optimization less sensitive to internal distribution changes. It reported faster training and improvements to an ImageNet classifier.^[28] The method became widely used, but later normalization methods and implementation choices differ from the procedure in the original paper.

Vision Transformer

The 2020 paper "An Image is Worth 16x16 Words" has the explicit affiliation "Google Research, Brain Team." It applied a standard Transformer encoder to sequences of image patches and showed that, with large-scale pretraining, the resulting Vision Transformer could perform competitively on image-classification benchmarks.^[29] The result supported using Transformer backbones in computer vision without retaining a convolutional network as the primary architecture.

Large-scale robot learning

Brain's robot-learning work included experiments in which multiple robots collected grasping experience and shared it for model training. A 2016 project collected more than 800,000 grasp attempts over two months using between 6 and 14 manipulators. A convolutional network predicted whether proposed gripper motions would lead to successful grasps, and the controller repeatedly updated motions from camera observations.^[15][30]

The project was collaborative. Google's technical post thanks colleagues at Google Research and X, and the annual Brain review describes it as collaboration with X.^[15][30] The evidence supports attributing the machine learning research to Brain participants while retaining the cross-team engineering context.

Magenta

Google Brain announced Magenta in June 2016 as a project researching machine learning for music and art. Its initial post says the project would use TensorFlow, release models and tools as open source, and publish demos and technical work.^[31] Brain's 2016 review records music-generation and artistic-style-transfer work under the program.^[15]

Magenta was a research and software project, not evidence that generated media had artistic intent or human authorship. Its historical relevance to Brain is the investigation of generative models and interfaces for creative work.

Neural architecture search

Barret Zoph and Quoc V. Le's 2017 neural architecture search paper carries a Google Brain affiliation. It trained a recurrent controller with reinforcement learning to generate candidate neural-network descriptions and used validation performance as the reward.^[32] The paper evaluated image classification and language modeling. Later AutoML systems used many different search spaces, optimization methods, and efficiency techniques, so the 2017 method should not be treated as a description of all AutoML.

Imagen and other late Brain-labelled projects

The 2022 Imagen project page identifies the affiliation "Google Research, Brain Team." Imagen combined a large pretrained text encoder with cascaded diffusion models for text-to-image generation and introduced DrawBench for human evaluation of text-image alignment.^[36] The paper and project site also discuss dataset and social-bias limitations, which are part of the technical record rather than optional qualifications.

Google Research continued using the Brain Team label for other 2022 publications. The FindIt post, for example, identifies its authors as researchers in "Google Research, Brain Team" and describes a model for object localization from natural-language queries.^[37] These dated affiliations show that Brain remained an active named team after the 2018 Google AI communications change and before the April 2023 reorganization.

Selected chronology

Scroll sideways for more →

PaLM and late-period organization

PaLM, the Pathways Language Model, was a broad Google Research project with dozens of authors. Its 2022 paper describes a 540-billion-parameter dense Transformer trained on 6,144 TPU v4 chips using Pathways.^[35] It reports evaluations across language, reasoning, code, bias, toxicity, and memorization.

PaLM belongs in a history of Brain's late technical program because it used systems and leadership connected to Brain, and Google's successor history includes the PaLM family in that lineage.^[1][2] The paper itself lists Google rather than an exclusive Brain affiliation. The precise description is therefore a Google Research project with substantial Brain lineage, not a model produced only by an isolated Brain unit.

This caution applies generally after 2018. Papers could involve Brain, Google AI Language, Google Research systems teams, product groups, outside institutions, or several of these at once. Author affiliation on the paper and a dated team announcement are stronger evidence than later attempts to assign the work to one internal group.

Production and cross-team work

Google Brain combined research with collaboration on production systems. The documented examples are narrower than the broad claim that Brain powered most of Google's machine learning products:

Scroll sideways for more →

These examples show the team's operating model: research results could become shared infrastructure, open-source software, or components used by product teams. They do not imply that Brain owned the complete product, controlled every deployment, or maintained every later version.

People and attribution

The founding sources support specific roles for Ng, Dean, and Corrado. The DNNresearch acquisition sources support the company's identity, the move of Krizhevsky and Sutskever to Google, and Hinton's continued university appointment.^{[2][3][4][9][43]} Technical publications then identify the people attached to individual projects.

Scroll sideways for more →

This is a project-based list, not a roster of everyone employed by Brain. It also avoids assigning current employers or motives to former staff. Personnel changed over time, and a later job does not alter the affiliation recorded on an earlier paper.

Jeff Dean was the most consistent public technical leader in the available record. His profile says that he helped steer Brain's research direction, and he wrote the team's 2016 and 2017 annual reviews.^{[2][15][21][22]} Andrew Ng's institutional profile identifies him as the founding lead.^[4] Greg Corrado's profile identifies him as a co-technical lead in the founding project.^[3] These dated descriptions are more defensible than inferring a single continuous "head of Google Brain" title across the entire organization.

Relationship to Google Research and DeepMind

Before April 2023, Google Brain and DeepMind were distinct organizations within Alphabet. Brain operated as a team in Google Research, while DeepMind had its own leadership, research program, and public identity. Collaboration did occur. For example, Google's first TPU supported DeepMind's AlphaGo, but the Brain review presents that as an infrastructure use, not as a merger of the two research programs.^[15]

Google Research was broader than Brain. The 2017 Brain research statement calls the team one part of Google's overall research effort, and the 2023 reorganization announcement says Google Research would continue work in algorithms and theory, privacy and security, quantum computing, health, climate and sustainability, and responsible AI.^[20][38]

The 2018 "Google AI" change renamed public research channels and grouped communications under a wider label.^[23] It did not make every Google AI paper a Brain paper. The continued use of "Google Research, Brain Team" on publications through 2022 provides direct evidence that Brain remained a distinct internal affiliation inside the broader umbrella.^[29][36][37]

Reorganization into Google DeepMind

On April 20, 2023, Sundar Pichai announced that the Brain team from Google Research and DeepMind would form a new unit called Google DeepMind. Demis Hassabis became CEO of the combined unit. Jeff Dean became Google's Chief Scientist, reporting to Pichai and serving as Chief Scientist to both Google Research and Google DeepMind.^[38]

Pichai's announcement also states that Google Research would remain a separate organization under James Manyika, with a remit spanning fundamental computer science and applied research areas. Brain's transfer therefore did not mean that all of Google Research became Google DeepMind.^[38]

Hassabis's announcement describes the same combination and names a new scientific board for the unit. It places Zoubin Ghahramani on the research leadership team reporting to Koray Kavukcuoglu and says Eli Collins would join the leads team as vice president of product.^[39] These were announced roles at the time of the reorganization, not guarantees of later reporting structures.

The announcements give Google's stated rationale: concentrate talent, computing infrastructure, and decision-making for more capable general AI systems, while pursuing safety and responsibility.^[38][39] They do not independently establish claims that the merger was caused by one competitor, one product launch, or one internal failure. Such causal claims require evidence beyond timing and should not be presented as fact.

The transition continued after the announcement. A July 2023 Google DeepMind conference post notes that papers initially submitted under a Google Brain affiliation would appear in a Google Research post because many submissions predated the combination.^[40] That note is useful for interpreting 2023 papers: the affiliation at submission can be historically accurate even when the authors' organization had changed by publication time.

How to interpret Google Brain's record

Google Brain's technical record can be described through verifiable papers, software releases, and dated organizational announcements. DistBelief and TensorFlow document the evolution of distributed machine learning infrastructure.^[7][10][11] Word2vec, sequence-to-sequence learning, GNMT, and the Transformer document a progression in language modeling and translation.^{[12][14][17][19]} Vision Transformer, large-scale grasping, neural architecture search, Magenta, and Imagen show that the team also worked in vision, robotics, automated model design, and generative media.^{[29][30][31][32][36]}

Several broader conclusions require restraint. Citation totals change continuously and are not needed to establish that a paper was influential. Product claims should identify the documented collaboration rather than assign an entire Google service to Brain. Personnel claims should follow dated affiliations rather than assume a current role. Later Google DeepMind systems may inherit Brain code, ideas, or staff, but they are not retroactively Google Brain projects.

After April 20, 2023, Google Brain no longer existed as a separate team in Google's public organizational structure.^[38][39] Its software and research continued to be used, and many former Brain researchers joined the combined organization. The correct current link for the successor organization is Google DeepMind, while Google Research remains a separate research organization.

See also

• Google
• Google Research
• Google DeepMind
• TensorFlow
• Transformer
• BERT
• PaLM
• Vision Transformer
• Tensor Processing Unit
• Machine translation

References

1. ^Google DeepMind. "About Google DeepMind," section "Our journey." deepmind.google/about
2. ^Google Research. "Jeffrey Dean." research.google/...jeff
3. ^Google Research. "Greg Corrado." research.google/...gregcorrado
4. ^Stanford Institute for Human-Centered Artificial Intelligence. "Andrew Ng." hai.stanford.edu/...andrew-ng
5. ^Dean, Jeff, and Andrew Ng. "Using large-scale brain simulations for machine learning and A.I." Google, June 26, 2012. blog.google/...g-large-scale-brain-simulations-for
6. ^Le, Quoc V., et al. "Building High-level Features Using Large Scale Unsupervised Learning." ICML 2012. research.google.com/...unsupervised_icml2012
7. ^Dean, Jeffrey, et al. "Large Scale Distributed Deep Networks." Google Research, NIPS 2012. research.google/...scale-distributed-deep-networks
8. ^Dean, Jeffrey, et al. "Large Scale Distributed Deep Networks." Advances in Neural Information Processing Systems 25, 2012. proceedings.neurips.cc/...823815f66102863-Abstract
9. ^Franca, Sara. "Google acquires U of T neural networks company." University of Toronto, March 12, 2013. utoronto.ca/...cquires-u-t-neural-networks-company
10. ^Dean, Jeff, and Rajat Monga. "TensorFlow: Google's latest machine learning system, open sourced for everyone." Google Research, November 9, 2015. research.google/...ystem-open-sourced-for-everyone
11. ^Abadi, Martin, et al. "TensorFlow: A System for Large-Scale Machine Learning." 12th USENIX Symposium on Operating Systems Design and Implementation, 2016. usenix.org/...abadi
12. ^Mikolov, Tomas, Kai Chen, Greg Corrado, and Jeffrey Dean. "Efficient Estimation of Word Representations in Vector Space." ICLR 2013. research.google/...representations-in-vector-space
13. ^Mikolov, Tomas, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. "Distributed Representations of Words and Phrases and their Compositionality." NIPS 2013. proceedings.neurips.cc/...65f3c4923ce901b-Abstract
14. ^Sutskever, Ilya, Oriol Vinyals, and Quoc V. Le. "Sequence to Sequence Learning with Neural Networks." NIPS 2014. proceedings.neurips.cc/...97f410bb7eca942-Abstract
15. ^Dean, Jeff. "The Google Brain Team: Looking Back on 2016." Google Research, January 12, 2017. research.google/...brain-team-looking-back-on-2016
16. ^Le, Quoc V., and Mike Schuster. "A Neural Network for Machine Translation, at Production Scale." Google Research, September 27, 2016. research.google/...translation-at-production-scale
17. ^Wu, Yonghui, et al. "Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation." 2016. arxiv.org/...1609.08144
18. ^Uszkoreit, Jakob. "Transformer: A Novel Neural Network Architecture for Language Understanding." Google Research, August 31, 2017. research.google/...ture-for-language-understanding
19. ^Vaswani, Ashish, et al. "Attention is All you Need." Advances in Neural Information Processing Systems 30, 2017. proceedings.neurips.cc/...bd053c1c4a845aa-Abstract
20. ^Dean, Jeff. "The Google Brain Team's Approach to Research." Google Research, September 13, 2017. research.google/...rain-teams-approach-to-research
21. ^Dean, Jeff. "The Google Brain Team: Looking Back on 2017 (Part 1 of 2)." Google Research, January 11, 2018. research.google/...ooking-back-on-2017-part-1-of-2
22. ^Dean, Jeff. "The Google Brain Team: Looking Back on 2017 (Part 2 of 2)." Google Research, January 12, 2018. research.google/...ooking-back-on-2017-part-2-of-2
23. ^Google AI Communications. "Introducing Google AI." Google Research, May 7, 2018. research.google/...introducing-google-ai
24. ^Google Research. "Highlights from the Google AI Residency Program." October 9, 2018. research.google/...the-google-ai-residency-program
25. ^Devlin, Jacob, and Ming-Wei Chang. "Open Sourcing BERT: State-of-the-Art Pre-training for Natural Language Processing." Google Research, November 2, 2018. research.google/...for-natural-language-processing
26. ^Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding." NAACL-HLT 2019. aclanthology.org/N19-1423
27. ^Szegedy, Christian, et al. "Going Deeper with Convolutions." CVPR 2015. research.google/...going-deeper-with-convolutions
28. ^Ioffe, Sergey, and Christian Szegedy. "Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift." ICML 2015. research.google/...ducing-internal-covariate-shift
29. ^Dosovitskiy, Alexey, et al. "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale." ICLR 2021. openreview.net/forum
30. ^Levine, Sergey, et al. "Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection." International Symposium on Experimental Robotics, 2017. research.google/...and-large-scale-data-collection
31. ^Eck, Douglas. "Welcome to Magenta!" Magenta, June 1, 2016. magenta.tensorflow.org/...welcome-to-magenta
32. ^Zoph, Barret, and Quoc V. Le. "Neural Architecture Search with Reinforcement Learning." ICLR 2017. openreview.net/forum
33. ^Jouppi, Norman P., et al. "In-Datacenter Performance Analysis of a Tensor Processing Unit." ISCA 2017. research.google/...sis-of-a-tensor-processing-unit
34. ^Barham, Paul, et al. "Pathways: Asynchronous Distributed Dataflow for ML." MLSys 2022. research.google/...ous-distributed-dataflow-for-ml
35. ^Chowdhery, Aakanksha, et al. "PaLM: Scaling Language Modeling with Pathways." 2022. research.google/...language-modeling-with-pathways
36. ^Saharia, Chitwan, et al. "Imagen: Text-to-Image Diffusion Models." Google Research, Brain Team, 2022. imagen.research.google
37. ^Kuo, Weicheng, and Anelia Angelova. "FindIt: Generalized Object Localization with Natural Language Queries." Google Research, September 20, 2022. research.google/...n-with-natural-language-queries
38. ^Pichai, Sundar. "Google DeepMind: Bringing together two world-class AI teams." Google, April 20, 2023. blog.google/...april-ai-update
39. ^Hassabis, Demis. "Announcing Google DeepMind." Google DeepMind, April 20, 2023. deepmind.google/...announcing-google-deepmind
40. ^Google DeepMind. "Google DeepMind's latest research at ICML 2023." July 20, 2023. deepmind.google/...-deepmind-research-at-icml-2023
41. ^DePristo, Mark, and Ryan Poplin. "DeepVariant: Highly Accurate Genomes With Deep Neural Networks." Google Research, December 4, 2017. research.google/...nomes-with-deep-neural-networks
42. ^Shallue, Chris. "Show and Tell: image captioning open sourced in TensorFlow." Google Research, September 22, 2016. research.google/...ning-open-sourced-in-tensorflow
43. ^University of Toronto. "Six degrees of separation: how U of T's Geoffrey Hinton is connected to top AI researchers around the world." March 28, 2017. utoronto.ca/...ted-top-ai-researchers-around-world