Mechanistic interpretability
We can inspect a neural network's numbers without understanding its computation. Can explanations become reliable enough to predict and check its behavior?
Machine learning
Within Machine learning · Search all subjects
AI alignment
Alignment asks how AI systems can reliably follow intended goals and constraints, including in situations different from those used in training.
Continual learning
An AI system can learn new tasks and lose skills it already had. How can it keep useful knowledge while adapting with limited memory and computation?
Out-of-distribution generalization
A model can predict well on familiar data and fail when conditions change. Which changes can it handle reliably, and how can we know?
Adversarial robustness
Carefully chosen changes to an input can fool a learned classifier. How can we provide reliable protection against a stated class of such changes?
Machine unlearning
Can a trained model efficiently remove the influence of selected training data, with credible evidence of what was removed?
High-precision analog in-memory computing
Several methods correct errors in analog memory calculations. How should they be combined when inputs, noise and conversion costs change?
Uniform-distribution PAC learning of DNF
Can ordinary random labeled examples teach a DNF formula in polynomial time under the uniform distribution?