MathDB7 problems
- 0 votes0 replies0 views
Reduced robustness conjecture for plain-network OTAD
Let OTAD denote the optimal transport-induced robust model, and let a plain network mean a network without residual connections. Reduced robustness conjecture. OTAD based on a plai…
- 0 votes0 replies0 views
The robustness–accuracy trade-off conjecture for neural networks
Adversarial training augments a dataset with perturbations constrained to lie within an -ball of the corresponding inputs, and robustness measures performance under such…
- 0 votes0 replies0 views
The improved convergence-rate conjecture for MSI-HG
Let MSI-HG denote the mini-batch semi-implicit hybrid gradient method discussed in the paper, and let be the number of iterations. MSI-HG rate conjecture. Based on experiments,…
- 0 votes0 replies1 view
Hybrid masking-and-compression conjecture for adversarial robustness
The LM-CCA strategy masks certain feature values completely, while the LC-CCA strategy uses compressed values; both are evaluated as counter-adversarial attack strategies for robus…
- 0 votes0 replies0 views
Typical-label requirement for robustness guarantees in trained neural networks
Consider a random deep neural network with a target classifier drawn from the same random distribution as the network's initialization, and suppose that the adversarial robustness…
- 0 votes0 replies0 views
Extension of adversarial robustness guarantees to trained deep neural networks
Consider deep neural networks trained on MNIST and CIFAR10 data, and the adversarial robustness guarantee proved for deep neural networks with random weights and biases: for any…
- 0 votes0 replies0 views
Conjecture on the generalization of information in robust representations
Generalization conjecture. The information conveyed by robust representations has better generalization, and generalization is more of a problem on CIFAR-10 than on MNIST.