MathDB4 problems
- 0 votes0 replies0 views
Conjecture on extending benign overfitting theory to super-wide deep neural networks
StoNet conjecture. The StoNet could enable the extension of benign overfitting theory from linear models to super-wide DNNs by leveraging its capability in sufficient dimension red…
- 0 votes0 replies0 views
Noise robustness of benign overfitting in the weak signal regime
Consider the weak signal regime for gradient descent in leaky ReLU two-layer neural networks, and introduce noise to the binary-classification labels. Noise-robustness conjecture.…
- 0 votes0 replies0 views
Conjecture on the effect of small-bandwidth spike inaccuracy on neural-network test error
For inputs of order one, let denote the bandwidth parameter of the Gaussian spike, and consider the inaccuracy of the sine approximation to the Gaussian neural tangent ker…
- 0 votes0 replies1 view
The low- and high-complexity decomposition conjecture for benign overfitting
A deep learning model is considered in an overparameterized regime, where it can fit the training data perfectly. A low-complexity component is one for which classical uniform conv…