This is an online seminar. Registration is required.
【Date】2021/Dec/15(Wed) 1100am-1130am (JST)
【Speaker】 Dr.Clifford Broni-Bediako
AutoDNN for Remote Sensing Image Understanding with Gene Expression Programming of Cellular Encoding
Research scientists are gearing up for adopting deep learning methods to their respective domain problems.
AutoDNN, also called automated neural architecture search, aims to automate the architecture design of deep neural networks (DNN) to enable researchers adopt DNN methods with ease, and with little or no expertise in deep learning. This work presents a novel generative encoding, symbolic linear generative encoding (SLGE), that combines the complementary strengths of gene expression programming and cellular encoding to develop modularized convolutional neural networks automatically via a random search and an evolutionary algorithm for remote sensing image understanding tasks. SLGE can evolve modularized architectures with shortcut (skip) connections and multi-branch connections similar to the ones commonly adopted by human experts. The effectiveness of the proposed method is demonstrated by discovering networks that achieve promising results on NWPU-RESISC45, AID, EuroSAT and BigEarthNet scene classification benchmarks and on ISPRS Vaihingen, Potsdam and UAVid semantic segmentation benchmarks. Compared with recent state-of-the-art systems, using fewer parameters under the specified model-size constraint, the results of the best discovered networks are
competitive to, or even exceed, human expert-designed networks for scene classification and semantic
segmentation in remote sensing image understanding.
Public events of RIKEN Center for Advanced Intelligence Project (AIP)Join community