Pushmeet Kohli

Kohli’s doctoral research used graph cuts for efficient inference in computer vision, including reuse of computation across related optimization problems. Before DeepMind, he spent a decade at Microsoft Research and became director of research for its Cognition group. His present unit has developed systems including AlphaFold, AlphaGenome, SynthID and AlphaEvolve. These programs bring learned representations together with the domain’s data, constraints and ways to test a result.[4][2]

Search over programs using an evaluator’s feedback

The coauthored AlphaEvolve paper combines language-model code changes with evolutionary search. Evaluators score candidates, and the system uses that feedback to propose and retain improvements. Its reported applications include data-center scheduling, accelerator circuit simplification and faster training code. The approach is especially useful where a candidate can be tested automatically; the evaluator defines what counts as an improvement and can constrain the discoveries the search can recognize.[6]

Build relationships into the representation

His coauthored graph-networks paper argues that generalizing to new combinations requires representations of entities, relations and composition rules. Graph networks provide a common framework for learning over that structure. The paper connects an architectural choice with a reasoning goal: reusing learned relationships when the particular objects or arrangement change. It combines a review, a position and a framework, with an accompanying software library.[7]

In his Training Data interview, Kohli contrasts domains where software can rapidly check a result with experiments that require physical work. He discusses how model-driven hypothesis generation and search can help scientists while keeping validation central. The practical question is which part of discovery can be accelerated and which evidence is needed before treating a proposed result as established.[5]