John Platt

Platt leads Applied Science at Google Research. His work has crossed machine learning, graphics, circuits, numerical analysis and climate science. Before Google, he was deputy director of Microsoft Research’s Redmond laboratory and director of research at Synaptics. Google’s biography also credits his work on calibrating model outputs, a problem distinct from learning a classifier: the numerical confidence of a prediction needs to correspond to how often such predictions are correct.[1]

Make a large optimization problem manageable

His SMO algorithm breaks support-vector-machine training into the smallest possible quadratic optimization problems. Those small subproblems can be solved analytically, avoiding a general numerical optimization routine inside the main training loop. The method also avoids keeping an enormous matrix in memory. The paper analyzes its computational behavior and tests it against the chunking methods used at the time, making training more practical for large datasets.[2]

Search for better scientific programs

Platt coauthored the Empirical Research Assistance paper. ERA combines a language model with tree search to write and improve software against a defined quality metric. The system explores candidate research ideas, implements them and uses experimental results to decide what to investigate next. The paper tests it across scientific tasks including biological data analysis and forecasting. The result concerns the quality of generated empirical software on those tasks.[3]

In his Latent Space interview, Platt distinguishes optimizing a prediction score from establishing an explanation of the world. A scientist still has to choose a meaningful objective and check whether a result reflects the intended phenomenon. His account of ERA’s origins in automated Kaggle experiments shows why improved benchmark performance and scientific understanding require separate judgments.[4]