SLIDE 1 Fakultät für Informatik
Technische Universität München
Deep Learning in Smart Spaces
Markus Loipfinger
Advisor(s): Marc-Oliver Pahl, Stefan Liebald Supervisor: Prof. Dr.-Ing. Georg Carle Chair of Network Architectures and Services Department of Informatics Technical University of Munich (TUM)
[1]
SLIDE 2 2
Outline
Motivation Analysis
Related Work Design & Implementation Evaluation
Conclusion
SLIDE 3 3
Motivation
Why Deep Learning in Smart Spaces?
- Enable interesting use cases such as facilitating
the daily life routines (e.g. self-adapting rooms)
- Find complex coherences among the data
- Goal: Enable users with little or even no pre-knowledge
to build and train a neural network
- Approach: Provide easy-to-use machine learning
functionality
- Modularize to suitable building block
- Enable Mash-Up
[9] [11]
SLIDE 4
4
Analysis: How does Deep Learning work?
Deep Learning – Feed-forward Neural Network
Forward Pass Backward Pass x1 x2 x3 b w3 w2 w1 a y = a ( x * W + b )
SLIDE 5
5
Analysis
Deep Learning Architectures – Recurrent Neural Network
. . . n e t . . . e t w . . .
SLIDE 6
6
Analysis
Deep Learning Architectures – Deep Belief Network
Classifier
SLIDE 7 7
Analysis
Application of machine learning / deep learning in smart spaces
- Knowledge required in
- Machine learning / deep learning
- Machine learning framework / library
- Design of neural network based on
- Particular problem
→ Each problem requires an appropriate learning algorithm
- Available training data
- Provide machine learning as a service
- Reduce complexity of machine learning
- Users do not require pre-knowledge
- Ensure usability & reusability
- Rapid prototyping
SLIDE 8 8
Related Work
Use cases mentioned in related work
→ e.g Health & Home Care, Comfort, Security, Energy Saving, etc.
- Supporting (disabled) people [2], [3]
→ Security (e.g. fire alarm) by FFNN → Automation (e.g. controlling of all devices) by RNN
- Recognizing human activites [4], [5], [6]
→ Pre-Training of a DBN (unsupervised learning) → Fine-Tuning of the DBN (supervised learning) → Output: 1 out of 10 activities
- Predicting human behaviour [7]
→ Pre-Training & Fine-Tuning of a DBN → Hybrid architecture with a classifier on top of the DBN → RNN: takes previous actions into account
- Smart Grid: Q-Learning, Reinforcement Learning [8]
SLIDE 9 9
Design
Modular Approach
- Separation of the state of a neural network and the respective
learning algorithm
All parameters and hyperparameters of a neural network contained in a configuration file Three machine learning services
- Feedforward Neural Network
- Deep Belief Network
- Recurrent Neural Network
SLIDE 10
10
Insertion: Neural Network - Parameters & Hyperparameters
Neural Network Learning rate Cost function Number of hidden layers Optimization technique Regularization technique Type of activation function Number of training epochs Size of mini-batch Weight initialization Bias initialization
SLIDE 11
11
Functionality
SLIDE 12 12
Implementation
Machine Learning Services
- One service for each neural network
- Based on a context model
- Neural network implementation with the help of TensorFlow
- Configuration file reader
- Prepare training data
→ Bring data into the right shape → each neural network requires other conditions on the input:
- FFNN: (data, label)-pairs of size [batch_size, feature_size]
- r [batch_size, label_size]
- DBN: data of size [batch_size, feature_size]
- RNN: (data, label)-pairs of size [batch_size, num_steps,
feature_size] or [batch_size, label_size]
SLIDE 13 13
Evaluation
Main aspects: Performance, Usability & Reusability
Quantitative Evaluation Latency of the modular approach
➢ Maybe a bit slower due to modularization
Accuracy and training-time
➢ Probably comparable
Lines of Code for implementing the use case by using my machine / deep learning modules
➢ Significantly less
Time for
➢ Implementing the use cases ➢ Creating the corresponding neural networks
Qualitative Evaluation Experience with the concept Usability Effect and application of reusability
SLIDE 14
14
Evaluation
Accuracy / Reconstruction Error – Comparison
Approach Accuracy / Reconstruction Error Lines of Code Service FFNN 97.50 % 2 Regular FFNN 97.58 % 85 Service DBN 0.02 2 Regular DBN 0.02 148 Service RNN 98.50 % 2 Regular RNN 98.44 % 128
SLIDE 15
15
Evaluation
Training Time - Comparison
500 1000 1500 2000 2500 3000 Training Time [s] 200 400 600 800 Training Iterations [x10] Regular FFNN Service FFNN Regular DBN Service DBN Regular RNN Service RNN
SLIDE 16
16 500 1000 1500 2000 2500 3000 Training Time [s] 200 400 600 800 Training Iterations [x10] Regular FFNN Service FFNN Regular DBN Service DBN Regular RNN Service RNN
Evaluation
Training Time - Comparison
SLIDE 17 17 500 1000 1500 2000 2500 3000 Training Time [s] 200 400 600 800 Training Iterations [x10] Regular FFNN Service FFNN Regular DBN Service DBN Regular RNN Service RNN
Evaluation
Training Time - Comparison
16.0 16.2 16.4 16.6 16.8 Training Time [s] 50 86 87 88 89 90 91 92 Training Time [s] 50 860 865 870 875 880 885 890 Training Time [s] 50
SLIDE 18 19
Evaluation
Running Time - Comparison
Regular FFNN Service FFNN Regular DBN Service DBN Regular RNN Service RNN Machine Learning Approach 2 4 6 8 10 12 Runtime [s]
SLIDE 19 20
Evaluation
Running Time - Comparison
Regular RNN Service RNN 3 4 5 6 7 8 9 10 11 Runtime [s] Regular FFNN Service FFNN 0.25 0.30 0.35 0.40 0.45 0.50 Runtime [s] Regular DBN Service DBN 2 3 4 5 6 7 8 Runtime [s]
SLIDE 20 21
Evaluation
Main aspects: Performance, Usability & Reusability
Quantitative Evaluation Latency of the modular approach
➢ Maybe a bit slower due to
modularization Difference of about 0.3 s Accuracy and Time
➢ Probably comparable
Acc: similar, Time: comparable Lines of code for implementing the use case by using my machine / deep learning modules
➢ Significantly less
83, 146, 126 lines of code less (1) Time for
➢ Implementing the use cases
Service: ~30 s, Regular: 5 - 10 min
➢ Creating the corresponding neural
networks Service: ~30 s, Regular: 2 - 5 min Qualitative Evaluation Experience with the concept
➢ Service handling ++ ➢ Code understanding +/0/0 (1) ➢ Service understanding + ➢ Configuration file understanding ++ ➢ Configuration file modifying ++ ➢ Neural network creation + ➢ Neural network training +/++/0 (1) ➢ Understanding without ML/DL +
knowledge Usability ++/++/+ (1) Reusability ++/++/+ (1)
(1)FFNN/DBN/RNN
SLIDE 21 22
Conclusion
- Realization of three machine learning services which
- are easy-to-use
- do not require knowledge in
- machine learning
- a machine learning library
- yield
- a high usability
- reusability
- a good performance
Separation of the state of a neural network and the corresponding learning algorithm → configuration file Structure of the learning algorithms and the neural networks already implemented
SLIDE 22
23
Thank you
SLIDE 23 24
Sources
[1] https://storiesbywilliams.com/tag/ibm-watson-supercomputer/ [2] A. Hussein, M. Adda, M. Atieh, and W. Fahs, “Smart home design for disabled people based on neural networks,” Procedia Computer Science, vol. 37,
[3] A. Badlani and S. Bhanot, “Smart home system design based on artificial neural networks,” in Proc. of the Word Congress on Engineering and Computer Science, 2011. [4] H. Fang and C. Hu, “Recognizing human activity in smart home using deep learning algorithm,” in Proceedings of the 33rd Chinese Control Conference, July 2014, pp.4716–4720. [5] H. D. Mehr, H. Polat, and A. Cetin, “Resident activity recognition in smart homes by using artificial neural networks,” in 2016 4th International Istanbul Smart Grid Congress and Fair (ICSG), April 2016, pp. 1–5. [6] H. Fang, L. He, H. Si, P. Liu, and X. Xie, “Human activity recognition based
- n feature selection in smart home using back-propagation algorithm,” {ISA}
Transactions, vol. 53, no. 5, pp. 1629 – 1638, 2014, {ICCA} 2013. [7] S. Choi, E. Kim, and S. Oh, “Human behavior prediction for smart homes using deep learning,” in 2013 IEEE RO-MAN, Aug 2013, pp. 173–179. [8] D. Li and S. K. Jayaweera, “Reinforcement learning aided smart-home decision- making in an interactive smart grid,” in 2014 IEEE Green Energy and Systems Conference (IGESC), Nov 2014, pp. 1–6. [9] https://www.google.de/search?q=pr%C3%A4sentation+m%C3%A4nnchen+ idee&source=lnms&tbm=isch&sa=X&ved=0ahUKEwihgLvYyPzUAhWIh7QKHZ [10]07D68Q_AUICigB&biw=1855&bih=990#imgrc=dTd6OykL30b4QM: [11] http://www.allnetflats-in-deutschland.de/smarthome
SLIDE 24 25
Configuration File
[Neural Network] type = Feed Forward Neural Network [Input Layer] feature_size = 784 [Hidden Layers] hidden_layer_1 = 100 number_of_hidden_layers = 1 [Output Layer]
softmax_unit = True no_activation = False [Weight] mean = 0.0 standard_deviation = 0.1 seed = 123 [Bias] constant = 0.0 [Cost Function] cross_entropy = False squared_errors = True [Optimization Technique] gradient_descent = True momentum = False adagrad_optimizer = False [Regularization Technique] dropout = False weight_decay = False [Additional Methods] learning_rate_decay = False early_stopping = True k_cross_validation = False [Hyperparameters] activation_fct_tanh = True activation_fct_sigmoid = False activation_fct_relu = False learning_rate = 0.08 number_of_training_epochs = 25 mini-batch = 300 momentum_rate = 0.8 p_keep = 0.75 wc_factor = 0.6 lr_decay_step = 100000 lr_decay_rate = 0.96 early_stopping_rounds = 100 early_stopping_metric_loss = True early_stopping_metric_accuracy = False validation_k = 10 [Parameters] display_step = 10
