Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Center for Computational Neuroscience, Flatiron Institute, New York, NY, USA., Department of Physics, Harvard University, Cambridge, MA, USA., Kempner Institute, Harvard University, Allston, MA, USA., Department of Applied Mathematics, Harvard University, Cambridge, MA, USA.
2
0
4
Grokking and double descent aren't as mysterious as you think: they're just the push and pull between representation learning and readout calibration.
Forget in-distribution accuracy, the shape of your model's internal representations is a better predictor of real-world generalization.