Automatically Differentiable Nonlinear Tensor Networks (ADNTNs) for Exponential Compression of Deep Neural Networks

In three linesAutomatically Differentiable Nonlinear Tensor Networks (ADNTNs) compress neural network weights using compact core tensors trained via reverse-mode automatic differentiation. Tested on AlexNet and VGG-16, they achieve per-layer compression ratios from 2000× to 77000× while maintaining or improving accuracy.

Summary generated by Claude — human-verified