Sourodeep Bhattacharjee

 

Date: Friday, May 03, 2019

Time: 10:00 am - 12:00 pm

Location: 3105, Lambton Tower

 

we have proposed an Estimation of Distribution Algorithm (EDA) based automated architecture search algorithm for convolutional neural networks named VAE-EDA-Q AVS. The proposed method uses Variational Autoencoders which allows uniform and smooth sampling of latent spaces and can efficiently explore the search space of possible CNN configurations. The core algorithm of VAE-EDA-Q AVS was extended to encode arbitrary CNN architectures of dynamically generated depths and to produce optimized offspring candidates of similarly varied CNN architectures.

VAE-EDA-Q AVS was tested on CIFAR10 and CIFAR100 benchmark datasets successfully and compared to various state-of-the-art CNN algorithms. It was observed that VAE-EDA-Q AVS generate CNN models that have 1.5% higher accuracy for CIFAR10 on average compared to all other state-of-the-art algorithms while requiring 25% less parameters and 6% higher accuracy for CIFAR100 with 10% less parameters on average. This indicates that VAE-EDA-Q AVS is able to discover CNN architectures that are simpler in design and yet provides a better classification accuracy.

 

Thesis Committee:

 

Seminar Announcement

 

5113 Lambton Tower, 401 Sunset Ave., Windsor ON, N9B 3P4, (519) 253-3000 ext. 3716, csgradinfo@uwindsor.ca