H.B. Keller Colloquium

Monday February 1, 2021 4:00 PM

Relaxed Differentiable Search Methods for Deep Neural Networks

Speaker: Jack Xin, Department of Mathematics, University of California at Irvine
Location: Online Event

 Neural network architectures are traditionally hand designed based on a designer's 

 knowledge, intuition and experience. In recent years, data-driven design has been studied as 

 an alternative where a search algorithm is deployed to find an optimal architecture 

 in addition to optimizing network weights. This led to a two-level optimization problem to solve. 

 I shall review a few methods especially differential architecture search (DARTS), then 

 introduce a single-level optimization problem as a relaxation approximation and the associated  

 convergent algorithm RARTS. Through architecture/weight variable splitting and Gauss-Seidel iterations, 

  RARTS outperforms DARTS in accuracy and search efficiency, shown 

  in a solvable problem and the CIFAR-10 image classification based search. 

  The gain over DARTS continues upon transfer to ImageNet (1000 image classes). 

  RARTS is further applied for network architecture compression, such as 

  channel pruning of overparameterized convolutional neural networks.

  In experiments on ResNet-18 and MobileNetV2, RARTS achieves considerable 

  channel sparsity and learns slim deep networks with satisfactory accuracy. 

Series H. B. Keller Colloquium Series

Contact: Diana Bohler at 6262326138 dbohler@caltech.edu