Naslov
Improving understanding of deep learning models for image classification through visual analytics

Autori
Dražić, Ante ; Kuzmanić Skelin, Ana ; Bonković, Mirjana

Vrsta, podvrsta i kategorija rada
Radovi u zbornicima skupova, cjeloviti rad (in extenso), stručni

Izvornik
Workshop on Information and Communication Technologies ; 27th International Conference on Software, Telecommunications and Computer Networks / - , 2019, 54-56

Skup
27th International Conference of Software, Telecommunications and Computer Networks (SoftCOM 2019)

Mjesto i datum
Split, Hrvatska, 19.09.2019. - 21.09.2019

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
deep learning, neural network, image classification, visual information analytics, black-box models

Sažetak
Recent advances in architectures of deep learning models have improved their performance in the field of computer vision with an important contribution to image classification. Understanding performance characteristics of deep learning on these high dimensional datasets are still elusive due to the large number of parameters leading to complex interactions and vagueness in the understanding of an underlying decision processes. To this end, visual representation of learned image features and the decision dataflow enables insights into multi modular structure. In this work we explore key aspects of the convolutional neural network (CNN) as the most studied network architecture of deep learning image classifiers through data and process visualization graphs in a TensorFlow framework on a CIFAR-10 dataset. Comparative performance between three different improvement techniques is given which is presented in such a way as to demonstrate the usability of graphical presentation as a tool for visualizing computational process.

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika, Računarstvo

POVEZANOST RADA