46th #ebaytechtalk: Detecting activity patterns w/ Convolutional Neural Networks

This is a past event

131 people went

Details

46th #ebaytechtalk: Detecting activity patterns in accelerometer data using Convolutional Neural Networks, a talk by Kaitao Yang

Abstract
-----------
Neural Networks have become a state-of-the-art method in machine learning these days. Possible applications include image recognition (faces, objects) and the interpretation of large quantities of data. One of these applications is recognizing different states of activity (such as walking, jogging, cycling, driving, sitting on the train, and sleeping) from accelerometer data as present in modern fitness-trackers. Convolutional Neural Networks (CNNs) are a subcategory of neural networks. Successful variants include LeNet, AlexNet, VGG, GoogLeNet, ResNet, Xception, and MobileNet. All of these were originally designed to process two-dimensional data such as images.

In this talk I will show that, with minor modifications, the VGG model can be used to process accelerometer data, which is one-dimensional. We trained three CNNs using labelled accelerometer data to detect daily activities. All our trained models achieved 98% accuracy. More details of this study include: (1) data pre-processing (streaming, shuffling, augmentation); (3) activation function; (4) loss function; (5) GPU computing; (6) Dutch supercomputer for training a large-scale model; and (7) Google cloud services (storage and virtual machine).

Speaker bio
----------------
Kaitao Yang works as a Data Scientist & Deep Learning Expert at eBay in Amsterdam. Before joining eBay, he worked as a Deep Learning Engineer at Jheronimus Academy of Data Science. He obtained his Professional Doctorate in Engineering (PDEng) degree from the Electrical Engineering department of Technische Universiteit Eindhoven, and his Master degree from the Cognitive Science department of Xiamen University, in 2016 and 2012, respectively. He has a mastery of popular Deep Learning models (CNNs, sequential models, generative models), with experience in applying them to process both structured data (such as SQL and CSV) and unstructured data (images, text, biomedical signals, and stock prices). In addition, he is actively offering Deep Learning lectures. For details, see: https://www.linkedin.com/feed/update/urn:li:activity:6285216778280927233.