This document provides an overview of deep learning and convolutional neural networks from David Solomon, an IBM executive architect. It begins with Solomon's background and credentials. It then defines deep learning, describes how neural networks learn feature hierarchies, and lists common deep learning techniques like convolutional neural networks for image recognition and recurrent neural networks for sequential data. The document explains how deep learning can learn complex patterns from large datasets using GPUs for fast training. It concludes with an example using the MNIST dataset of handwritten digits to demonstrate a simple convolutional neural network model in TensorFlow.

1. David Solomon
IBM Executive Architect and Senior Cloud and
Cognitive Evangelist

2. About Me
Focus / Passion
• AI, Cognitive, Emerging Technology
• Analytics
• Data (Architecture, Modeling, Integration)
• Cloud Service Architecture
• Applying the above to real-world business problems
Education & Certification
• M.S. Software Engineering
• B.S, Physics
• Data Mgmt, AI, Cloud, Docker, DevOps, …
Proud Member of
the IBM WolfPack
David Solomon
Technical
Evangelist, IBM
dsdlsolomo
@dlsolomo
Team-
wolfpack

3. • Deep Learning Introduction and
Concepts
• An Overview of Convolutional Neural
Networks (CNN)
• Demo/walk-through of a CNN Example
Session Overview

4. Deep Learning is inspired by the human brain, it attempts to mimic
the activity in layers of neurons in the human brain where thinking
occurs.

5. • a collection of statistical machine learning techniques
• used to learn feature hierarchies
• based on artificial neural networks

6. Activation of a neuron
The output from the neuron is a real number between 0 and 1
The neural net “learns” by tweaking the weights and biases step by step until the
prediction closely matches the correct output, i.e. minimize the “cost value”

7. Deep Learning algorithms learn “Feature Hierarchies” as
they progresses through their hidden layers

9. Simple Moderate Complex
Basic Classifiers:
Logistic Regression
or SVM
Pattern Complexity
Traditional Shallow
Neural Network
Deep Net
• Increased availability of labelled data
• Deep nets take a long time to train
• Availability of high performance GPUs speeds up training of a deep net
• GPU is approximately 250 times faster than CPU, i.e. the difference between
one day of training and over eight months.
DL algorithms learn more complex patterns than is possible with traditional
machine learning algorithms

12. https://www.youtube.com/watch?v=-P28LKWTzrI
CPU
Multiple Cores
GPU
Thousands of Cores

14. • a collection of statistical machine learning techniques
• used to learn feature
hierarchies• based on artificial neural networks
Technique Description Applications
Convolutional Neural
Network (CNN)
Hierarchical classification Image recognition
Recurrent Neural
Network (RNN)
Learns from patterns
time-series or sequential
data
Weather, stocks, speech
recognition
Deep Belief Network
(DBN)
Uses unsupervised
modeling for
classification
Image recognition

15. VISION
pixels -> edge -> texton -> motif -> part -> object
e.g. self-driving cars, reading medical images
SPEECH
sample -> spectral band -> formant -> motif -> phone -> word
e.g. Alexa
NATURAL LANGUAGE PROCESSING
character -> word -> clause -> sentence -> story
e.g. DeepText: Facebook's text understanding engine

16. Machine Learning Deep Learning
Data Volumes Can work on small data
volumes
Needs large data
volumes
Feature Engineering Needs a lot of data
preparation and feature
engineering
Feature engineering is
implicitly performed
during training
Hardware CPU and/or GPU Needs accelerators such
as GPUs
Model Interpretability Models are easy to
interpret (e.g., Decision
Trees)
Difficult to understand
why the model produces
a given prediction

19. Image
Channels with
pixels valued from
0-255
Filter(s)
Convoluted Feature

20. • Models are typically designed using Python and
other languages, using one or more frameworks
(Python coding in particular is key)
• Some knowledge of data science and deep
learning techniques and/or collaboration with
Data Scientists
• Knowledge/experience with key open source
technologies such as Python, Jupyter,
Tensorflow, etc.
• How to interface with deployed models from
relevant languages (e.g., Python, Java, NodeJS,
etc.)

22. • Our example will be using the the MNIST dataset of
handwritten digits 0-9
• This dataset contains 60,000 images and will be used for
training our model
• Since this is a greyscale dataset, we have only one
channel (as opposed to 3)
• We will focus on a simple model with one convolutional
layer
Our Deep Learning CNN Example- MNIST Dataset

23. General Neural Network Design Flow

24. CNN Demo Flow
Create a
Model with NN
Modeler
Train the model
with Experiment
Builder
Explore
TensorFlow in
Python
Deploy the
model
Use HPOs to
train and
compare
multiple
conditions
Deploy the
Tensorflow
Python Model