Jeff Dean and three Google colleagues created Discovery Loop to automate the research cycle

Jeff Dean and three Google colleagues created Discovery Loop to automate the research cycle

On August 5, Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals announced the creation of Discovery Loop. The company will begin with machine learning research and engineering and will test the approach in its own work.

The Discovery Loop website describes the goal as automating experimental cycles. A researcher formulates a hypothesis, runs an experiment, obtains a result, and chooses what to try next. The evaluation of each experiment becomes an input for the next step.

The company says its systems will propose experiments, run evaluations, and use the results to determine the next step. It plans to use advanced models and large computing resources to run thousands of experiments in parallel.

“There is an opportunity to automate more of the experimental cycle, which has until now required substantial human effort. This will allow more experiments and improve their quality,” Dean told TechCrunch.

This promise can be tested by examining the sequence of experiments: the evaluation of each experiment should change the next hypothesis and the next experiment. Generating a hypothesis starts such a sequence, while the measured result guides what follows.

Discovery Loop will first apply this approach to its own machine learning research and engineering. Its stated long-term goal is to use the approach in other areas of science and engineering where work can be organized as a sequence of tests.