AI & data science
MIT group works to make neural networks more interpretable
MIT researcher Andreas and his team are developing tools to explain neural networks and study how training data shapes model behaviour.
What happened
The group’s techniques can generate plain-language descriptions of processing nodes inside neural networks. It also studies how datasets and examples influence model behaviour, a relationship Andreas says lacks a solid scientific theory.
Why it matters
Machine-learning researchers and data professionals building or evaluating models could use interpretability tools to understand systems’ internal processes and anticipate their behaviour, supporting more robust models for specific applications.
What to watch
Teams could check whether model explanations identify relevant internal features, then compare those explanations with model behaviour on representative examples.
