Sie sind auf Seite 1von 28

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/321385358

Training Deep Neural Networks in Python Keras


Framework(Tensorflow Backend) with Inertial Sensor Data for Human
Activity Classification

Presentation · November 2017

CITATIONS READS

0 4,633

1 author:

Tahmina Zebin
University of East Anglia
27 PUBLICATIONS   81 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Human Activity Recognition with Inertial Sensors using a Deep Learning Approach View project

Gait Sensor Fusion View project

All content following this page was uploaded by Tahmina Zebin on 03 September 2018.

The user has requested enhancement of the downloaded file.


Training Deep Neural Networks in
Python Keras Framework

Tahmina Zebin, 1School of Electrical and Electronic Engineering,


Tahmina.zebin@Manchester.ac.uk
Deep Learning Models

• Common Machine Learning


Tasks:
• Regression: the output
variable takes continuous
values.
• Classification: the output
variable takes class labels.
• Clustering: Un-supervised

Todays Focus:
• Neural Networks: For Deep
learning
• Labels and Feature Calculation :
for activity Classification
• Raw Time Series : For Sensor
data prediction: Regression
Aim : Hands on Implementation of Deep
Learning Models

Steps Included:
• Data Preparation(Inertial Sensor Data).
• Evaluate Algorithms/Improve Results.
• Visualization and Presentation of Results.
Before we jump into codes
lets look at some building
blocks of a Neural Network
Biological Neuron vs the Artificial
neuron

From Wikipedia
Activation Functions
Different Activation Functions for different
layers?
Cost Function: Gradient
Descent

Batch Gradient Descent Stochastic Gradient Descent


Problems and solutions with
gradient calculation
Overfitting

• Cross-validation
• Regularisation techniques:
– Dropout
– Learning rate decay
– Batch normalization
Steps Associated with
Training a neural Network
Neural Network Architectures
to cover today

1.Deep Feed- 2. Recurrent Long-Short- 3. Convolutional


forward Network Term-Memory Network Neural Network

http://www.asimovinstitute.org/neural-network-zoo/
Work-Flow Diagram for Todays Task
Lets Begin the training
Software-Requisites :

• Anaconda distribution of
Python
3.6(https://anaconda.org/
anaconda)
• Additional Packages and
Libraries: scikit-learn,
Theano, Tensorflow,
Keras, Pytorch
Anaconda Navigator
Spyder IDE (Python 3.6)
Keras Library Framework
Data Pre-processing

• Feature Scaling
• Standardization vs Normalization
• Preparing Data matrix as required by a network
A simple feed Forward
Neural Network
LSTM For Sensor
Data Prediction

1 epoch 10 epoch 50 epoch


LSTM+Dense NN
For Classification(75 epochs)

With Batch Normalisation


Presenting Results from a
Neural Network

• Confusion Matrix
• Recall
• Precision
• F1-Score
Wearable Clinic :The Real World Scenario
2 Medical Conditions:
Schizophernia and
Chronic Kidney Disease

UK biobank:100000
Participants
Adaptive Sensing &
Behavioural phenotyping

• Unsupervised machine learning (e.g. kNN)


• Validate using accelerometer data from Biobank
• Some pilot data collection
• Example: bring forwards appointments based on wearable data
24
Conclusions
• Does depth Matter?
• Hybrid Networks?(LSTM+CNN)
• Transfer Learning?
• Can they be used in Risk Prediction Modelling?
• Real-time implementation?
References

• https://www.superdatascience.com/machine-learning/
• https://machinelearningmastery.com/setup-python-
environment-machine-learning-deep-learning-anaconda/
• http://colah.github.io/posts/2015-08-Understanding-LSTMs/
• https://cloud.google.com/blog/big-data/2017/01/learn-
tensorflow-and-deep-learning-without-a-phd
• https://keras.io/losses/
Thanks for your time 

Questions?

View publication stats

Das könnte Ihnen auch gefallen