Greg Corrado

Greg Corrado is a Google researcher, a Google Brain co-founder, and a co-author of the 2013 word-representation papers associated with word2vec. His work spans scalable deep learning, consumer applications, and healthcare AI. In healthcare, he has argued with collaborator Yossi Matias that promising model performance must be followed by rigorous testing and integration with clinical practice.[3][5][6][4]

In a February 2023 essay, Corrado and Matias argued that healthcare ML requires a different development approach from consumer software: “applying ML to health requires thoughtful assessment, ecosystem awareness, and rigorous testing.” They emphasized collaboration with expert health partners to establish useful clinical workflows and earn the trust of professionals and patients.[4]

The two authors reported that six years elapsed between their team's primary algorithmic research on preventing blindness in underserved diabetic populations and a deployment study demonstrating the integrated system's accuracy in community screening. They advocated partnerships with established medical technology companies to accelerate translation, while retaining open research and open software as complements to those partnerships.[4]

Their assessment of Med-PaLM made the distinction between benchmark success and clinical readiness concrete. In February 2023, they described it as a research-grade medical question-answering model rather than a finished product. Passing medical-exam questions, they wrote, “still falls well short of the safety and accuracy required to support real-world use for medical question answering.”[4]

With Michael Howell and Karen DeSalvo, Corrado co-authored the 2024 JAMA article “Three Epochs of Artificial Intelligence for Healthcare” (listed in his Google publication profile). They proposed a framework distinguishing symbolic and probabilistic systems, task-specific deep learning, and foundation models. Their argument paired foundation models' ability to change tasks through instructions with new risks, including hallucinations.[2]

Corrado also co-authored the 2025 Med-PaLM 2 and AMIE papers. Med-PaLM 2 combined base-model improvements, medical fine-tuning, and strategies for reasoning and grounding. AMIE used a simulated self-play environment with automated feedback to train diagnostic dialogue. The AMIE authors qualified their consultation results: the study used patient-actors and synchronous text chat, an unfamiliar format in clinical practice, and further research was required before real-world translation.[2]

Before this healthcare work, Corrado studied physics as an undergraduate at Princeton and pursued graduate studies in neuroscience and computer science at Stanford. At IBM Research, before joining Google, he worked on neuromorphic silicon devices and large-scale neural simulations. He subsequently became a founding member and co-technical lead of Google's large-scale deep neural networks project, bringing experience in biological and artificial neural systems to that work.[3][2]

Corrado co-authored “Efficient Estimation of Word Representations in Vector Space” with Tomas Mikolov, Kai Chen, and Jeffrey Dean, first submitted in January 2013. The team proposed two architectures for learning continuous word vectors from large datasets. In the September 2013 revision, they reported learning high-quality vectors from 1.6 billion words in less than a day, with state-of-the-art results on their syntactic and semantic word-similarity test set.[5]

Their October 2013 paper, “Distributed Representations of Words and Phrases and their Compositionality”, added Ilya Sutskever to the author team. It extended the continuous Skip-gram model through frequent-word subsampling, which the authors reported improved speed and representations, and negative sampling, an alternative to hierarchical softmax. It also addressed a limitation of individual word vectors: the meanings of “Air” and “Canada” do not readily combine into the meaning of “Air Canada.” The team introduced a method for finding phrases and learning vectors for millions of them.[6]

Corrado's application work included SmartReply. Its 2015 technical account describes paired recurrent neural networks: one encoded incoming email into a vector, and the other generated possible replies. Anjuli Kannan developed the sequence-to-sequence experiments with researchers and Gmail engineers; Sujith Ravi's team supplied semantic-intent mapping so suggestions differed in meaning as well as wording. These collaborators' contributions supplied practical mechanisms for turning learned language representations into useful email assistance.[3][8]

The December 2023 Latent Space NeurIPS recap names Corrado alongside Jeff Dean in its Word2Vec segment, revisiting the collaborative research a decade after the original papers.[1][5][6]