Overview:
jacques.lethuaut@gmail.com
Deep Learning - Part I
R8L0PN473F
This file is meant for personal use by jacques.lethuaut@gmail.com only.
Sharing or publishing the contents in part or full is liable for legal action.
Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
What is Deep Learning?
● Deep Learning is a specific subset of Machine Learning that utilizes a computational technique
called Neural Networks, in an attempt to replicate how biological intelligence processes
information and makes decisions.
● As the name suggests, the ideas behind Neural Networks draw their inspiration from Neuroscience
and other areas of biological research, and have already been around for many decades.
However, it is only recently, due to the rise in the volume of data and the increased computational
resources available to modern-day researchers, that the whole world has realized the power and
jacques.lethuaut@gmail.com
R8L0PN473F
flexibility of Deep Learning, and the true paradigm shift it represents in the evolution of computing.
● To understand the reasons for the above, we first
need to gain an intuition for how Deep Learning
does things differently from more traditional
techniques in Machine Learning, and why that
approach allows for a more autonomous form of
data-driven decision-making.
This file is meant for personal use by jacques.lethuaut@gmail.com only.
Sharing or publishing the contents in part or full is liable for legal action.
Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
Typical Machine Learning Workflow
● In a typical Machine Learning setting for predictive applications, the workflow would look
something like the following diagram:
Machine Learning
Car
jacques.lethuaut@gmail.com No Car
R8L0PN473F
Input Feature Extraction Classification Output
● Let’s assume we have a classification system that aims to predict
whether an image belongs to a car or not.
This file is meant for personal use by jacques.lethuaut@gmail.com only.
Sharing or publishing the contents in part or full is liable for legal action.
Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
Typical Machine Learning Workflow
● In a typical Machine Learning setting for predictive applications, the workflow would look
something like the following diagram:
Machine Learning
Car
jacques.lethuaut@gmail.com No Car
R8L0PN473F
Input Feature Extraction Classification Output
● In that case, the input data to our model would be images of
various objects, from which we would expect some of the images
to be those of cars.
This file is meant for personal use by jacques.lethuaut@gmail.com only.
Sharing or publishing the contents in part or full is liable for legal action.
Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
Typical Machine Learning Workflow
● In a typical Machine Learning setting for predictive applications, the workflow would look
something like the following diagram:
Machine Learning
Car
jacques.lethuaut@gmail.com No Car
R8L0PN473F
Input Feature Extraction Classification Output
● In the usual Machine Learning workflow, we would feed in these
images, construct a set of features out of the image (also called
Feature Extraction), and provide these features of the image to
our classification model.
This file is meant for personal use by jacques.lethuaut@gmail.com only.
Sharing or publishing the contents in part or full is liable for legal action.
Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
Typical Machine Learning Workflow
● In a typical Machine Learning setting for predictive applications, the workflow would look
something like the following diagram:
Machine Learning
Car
jacques.lethuaut@gmail.com No Car
R8L0PN473F
Input Feature Extraction Classification Output
● Only after these features are provided to our classification model,
will the model be able to make predictions and classify, based on
the features it was provided, if the input image belongs to a car or
not.
This file is meant for personal use by jacques.lethuaut@gmail.com only.
Sharing or publishing the contents in part or full is liable for legal action.
Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
The Problem of Feature Extraction
● The key point to note here is the process of feature extraction which is part of the Machine
Learning workflow.
● It is important to understand that in Machine Learning, the process of feature extraction represents
a bottleneck in the quality of the solution design. The more important / relevant the features we
can think of or create for the prediction problem, the better the model becomes. In other words, in
Machine Learning, the performance of the model is constrained by the quality of the feature set
jacques.lethuaut@gmail.com
R8L0PN473F
we can manually & creatively extract from the data and provide to the algorithm.
● In many of the curated, tabular datasets we have
worked with in Machine Learning, the columns of
the table already provide us with most of the key
features required for every instance, and only a
small amount of feature engineering is required
from there to get the final feature set for good
performance.
This file is meant for personal use by jacques.lethuaut@gmail.com only.
Sharing or publishing the contents in part or full is liable for legal action.
Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
The Problem of Feature Extraction
● However, the problem quickly becomes apparent when we start working with the more
unstructured data modalities present in real-world problem statements, such as images, text,
audio, video and graphs.
● In these cases, it becomes difficult to manually extract the right set of features from data and build
machine learning models that can recognize the right patterns to create good mappings from
input to output.
jacques.lethuaut@gmail.com
R8L0PN473F
● This is a core problem that statistical Machine Learning methods have historically struggled with -
their need for manual / handcrafted Feature Extraction techniques, and this is one of the main
reasons it was difficult to achieve good predictive performance on unstructured data before the
rise of Deep Learning.
● In the next deck, we will see how Deep Learning
solves this problem of Feature Extraction, and
why this has allowed it to achieve generalized
levels of high performance on many types of
unstructured data. This file is meant for personal use by jacques.lethuaut@gmail.com only.
Sharing or publishing the contents in part or full is liable for legal action.
Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
Summary
So in order to summarize:
● Deep Learning is a specific paradigm of Machine Learning that utilizes Neural Networks to replicate
how biological intelligence processes information and makes decisions.
● The issue faced by traditional Machine Learning algorithms is that their performance is reliant on
manual / handcrafted feature engineering techniques to extract the right features from instances for
jacques.lethuaut@gmail.com
R8L0PN473F
good predictive power. This approach does not scale well with the increasing complexity and
unstructured nature of the data prevalent in the modern world, and that is one reason Machine
Learning algorithms have not been very successful at making predictions on images, text, audio and
other unstructured data types.
● One of the reasons for Deep Learning’s rise in
prominence is that it overcomes this problem in
an elegant manner, and we will see how it
accomplishes this in the next slide deck.
This file is meant for personal use by jacques.lethuaut@gmail.com only.
Sharing or publishing the contents in part or full is liable for legal action.
Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
Happy Learning !
jacques.lethuaut@gmail.com
R8L0PN473F
This file is meant for personal use by jacques.lethuaut@gmail.com only.
Sharing or publishing the contents in part or full is liable for legal action. 10
Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
