This document discusses advances in machine learning and deep learning on IoT devices. It notes that the number of connected devices is growing rapidly and will reach 40 billion by 2020. It then covers different types of machine learning approaches like online learning vs learning from historic data. It also demonstrates several deep learning techniques including neural networks, convolutional neural networks, LSTMs, and autoencoders. Finally, it discusses challenges like computational complexity and potential solutions like IBM's TrueNorth neuromorphic chip.

1. Unless stated otherwise all images are taken from wikipedia.org or openclipart.org
DeepLearning and Advanced Machine Learning on IoT
@romeokienzler

2. • Intro (what I do, what we do, why we do IoT)
• Latest advancements in AI based ML
• Demo
• Hands on

4. Why IoT (now) ?
• 15 Billion connected devices in 2015
• 40 Billion connected devices in 2020
• World population 7.4 Billion in 2016

16. Machine Learning on
historic data
Source: deeplearning4j.org

17. Online Learning
Source: deeplearning4j.org

18. online vs. historic
• Pros
• low storage costs
• real-time model update
• Cons
• algorithm support
• software support
• no algorithmic improvement
• compute power to be inline
with data rate
• Pros
• all algorithms
• abundance of software
• model re-scoring / re-
parameterisation (algorithmic
improvement)
• batch processing
• Cons
• high storage costs
• batch model update

19. 1. API
2. pre-trained model
3. existing pipeline with your data
4. create own pipeline / model
abstraction levels

20. IBM Watson Personality
Insights

21. IBM Watson Natural
Language Classifier

23. DeepLearning
DeepLearning
Apache Spark
Hadoop

24. Neural Networks

25. Neural Networks

26. Deeper (more) Layers

27. AutoEncoders

28. AutoEncoders

29. Convolutional

30. Convolutional
+ =

31. Convolutional

33. Learning of a function
A neural network can basically learn any
mathematical function

34. Recurrent

35. LSTM
“vanishing error problem” == influence of past inputs decay
quickly over time

36. http://karpathy.github.io/2015/05/21/rnn-effectiveness/

40. •Outperformed traditional methods, such as
•cumulative sum (CUSUM)
•exponentially weighted moving average (EWMA)
•Hidden Markov Models (HMM)
•Learned what “Normal” is
•Raised error if time series pattern haven't been seen
before

43. Learning of an algorithm
A LSTM network is touring complete

44. Problems
• Neural Networks are computationally very complex
•especially during training
•but also during scoring
CPU (2009) GPU (2016) IBM TrueNorth (2017)

45. IBM TrueNorth
•Scalable
•Parallel
•Distributed
•Fault Tolerant
•No Clock ! :)
•IBM Cluster
• 4.096 chips
• 4 billion neurons
• 1 trillion synapses
•Human Brain
• 100 billion neurons
• 100 trillion synapses
•1.000.000 neurons
•250.000.000 synapses

46. DeepLearning
the future in cloud based analytics
Storage Layer (OpenStack SWIFT / Hadoop HDFS / IBM GPFS)
Execution Layer (Spark Executor, YARN, Platform Symphony)
Hardware Layer (Bare Metal High Performance Cluster)
GraphXStreaming SQL MLLib BlinkDB
DeepLearning4J
ND4J
R MLBase H2O
Y
O
U
GPUAVX
Intel Xeon E7-4850 v2 48 core, 3 TB RAM, 72 GB HDD, 10Gbps, NVIDIA TESLA M60 GPU
(cu)BLAS
jcuBLAS
S
T
R
E
A
M
S

50. bit.ly/toa16

53. •IBM Cloud Free Tier
•http://ibm.biz/joinIBMCloud
•Google TPU
•http://www.recode.net/2016/5/20/11719392/google-ai-ch
tpu-questions-answers
•IBM Neuromorphic Chip
•http://www.research.ibm.com/articles/brain-chip.shtm
•Recoding of the Talk
•https://www.youtube.com/watch?v=h5_NH3sL0Qw
•Contact Romeo Kienzler on Twitter: @romeokienzler