Experiences from applying Convolutional Neural Networks for classifying 2D sensor data
Applying deep neural networks on existing benchmark datasets is different than applying it on your company data where you have different requirements and constraints. This talk shows an example for applying a Convolutional Neural Network on real 2D sensor data, points out its challenges and provides hints.
Tags: Artificial Intelligence, Computer Vision, Deep Learning & Artificial Intelligence, Machine Learning
Scheduled on wednesday 14:50 in room cubus
Matthias Peussner has studied applied system science at the Osnabrück University, Germany. He worked several years as a software developer. Now he is working as scientist for ROSEN Technology & Research Center GmbH in Lingen, Germany. He has special interest in doing data science, developing machine learning models (deep learning) and creating software prototypes using Python.
When being the first in your company to apply a deep learning algorithm on your data you often have to overcome several obstacles. One challenge is to understand your data and to form a training and test dataset. Another one is to get your algorithms and its performance accepted and integrated in your existing processing workflow.
Convolutional Neural Networks have become a standard tool in processing image data. They have shown to reach human-level classification performance on some object recognition tasks.
In this talk I will present my experiences in getting started using a convolutional neural network for classification of 2D sensor data. I will point out the importance of understanding your data and give hints of how to select your train and test datasets according to the requirements. Furthermore, I will show how to get a feature extractor out of the classifier and how to visualize it.