Artificial Intelligence/Machine Learning/Deep Learning:

‘Bridging the Skills Gap’

Lesson 1: History, Definitions and Basic Concepts
Hello, welcome back to my course on AI and Machine Learning

Today we will address following items:

1. What is the difference between AI, Machine Learning (ML) and Deep Learning (DL) ?
2. History of AI/ML/DL
3. Definitions & Basic concepts of ML/DL (Loss/Cost Function, Mean Squared Error
(MSE), Regression, Classification, Supervised Learning, Unsupervised Learning)
labels are given for all i/p
What is the difference between AI/ML and DL?

Last week we saw that AI (and ML/DL) are hyped under the digitalization umbrella to the
extend, that people believe that their jobs will be taken over by robots in a few years. The
euphoria/fear comes from the lack of knowledge about the subject.

AI ML DL

AI has been around for a while and can be defined as the ‘automation of thought’…for
example a calculator and an xls spreadsheet can be considered as AI

ML: addresses the topic of a computer learning the rules of a system while looking at mere
inputs and outputs. In 1957, Frank Rosenblatt invented the Perceptron, the simplest form
of a Neural Network (NN). The Perceptron created a lot of excitement…but this changed in
1969.

In 1969, Minsky and Papert published a book highlighting the perceptron’s limitations and
that Frank Rosenblatt's predictions had been grossly exaggerated. For example, the
Perceptron was not able to learn a XOR (exclusive OR) function. A XOR function is a logical
operator that outputs TRUE if the 2 inputs differ. Research funding dried up and this led to
the first AI (or NN) winter
 1986: Rumelhart and Hinton solved the XOR problem and showed that backpropagation algo used in
(training of NN) works this was the basis of all subsequent NN and DL progress. deep
learning
In 1989, Yann LeCun (Bell Labs) combined the ideas of convolutional neural networks (CNNs) covnets
and backpropagation and applied them to classifying handwritten digits. The resulting had
network named LeNet was used by the US Postal Service to automate reading ZIP codes on about
mail envelops. 60k
params
NNs however faced another winter in the 1990s and the early 2000s mainly because of the
popularity of Kernel Methods. Kernel Methods are a group of classification algorithms, the
best known of which is the Support Vector Machine (SVM). In fact, in the early 2000s, most
NN papers were rejected at AI conferences. NN were still considered too expensive to train. expensive in
terms of
DL - is DL: DL is a subset of ML that puts emphasis on learning successive layers of increasinglycomputatn
power
neural meaningful representations. In 2006, Geoffrey Hinton launched the term ‘Deep Learning’ to
networks explain new algorithms that let computers ‘see’ and distinguish objects and text in images
only!!
and videos. Suddenly all other ML algorithms were branded as shallow learning.
The rebranding trick and the technological developments of GPUs revived research with
respect to NNs.

his CNN used AlexNet (2012): Alex Krizhevsky (advised by Hinton) created a broad and deep CNN that
60 million won the 2012 ImageNet competition (image classification competition: 1.4M high resolution
params, 60 color images into 1000 categories). 2012 marked the first year where a CNN was used to
million params
achieve a top 5 test error rate of 15.4%.
you need train

Since then many enhancements where developed to make NNs perform better and solve
new problems. By 2015, the winner reached an error rate of 3.6%. NNs have completely
replaced SVMs and Decision Trees in a wide range of applications. In 2016 and 2017, Kaggle
was dominated by gradient boosting machines (where structured data is available) and DL
(used for perceptual problems such as image classification). Most popular libraries are
XGBoost library and keras.

Most industries do not need or use DL. DL is not always the right tool for the job. Today
however, for image classification, speech recognition or machine translation, nobody would
dream of trying to do it without NNs.

Are we close to another AI winter? Unlikely because:

1. AI is still not central to the way we work, think and live

2. AI research has benefited from huge investments over the past 10 years and most of
the research findings of DL have not been applied.

In March 2019, Hinton, LeCunn and Bengio won the Turing Award for their contribution to
deep learning: https://www.nytimes.com/2019/03/27/technology/turing-award-ai.html
Definitions & Basic concepts:

Pre-ML:

Model is
Input Data programmed Output
by a human

ML:
New Input Data
Input Data Computer
‘learns’ the Predict
Output Data Model

Training Predicting

Example: Supervised Learning - Regression Problem label is any real number

Algorithm that predicts the value (k$) of a house in a specific city given its size (m2)
xi is the size of (house)i (m2)
yi is the price of the (house)i (k$)
m: number of training samples - m=6
h(xi)=yip hypothesis or prediction – a function learned during the training phase
allowing us to predict the price of any house given its size.
In order to measure the performance of the computer’s model we introduce a Cost
Function C indicating how far the output is from what we expect. ML learning aims at using
the loss as a feedback signal and adjust the weights of the model parameters (w and b) in
order to minimize the loss (y-yp) minimization problem differentiation problem
function must be continuous and smooth!

Cost Function C: Mean Squared Error (MSE)

1
C(w,b) = ∑6𝑖=1(𝑦𝑝𝑖 − 𝑦 𝑖 ) with yp the predicted housing price and y the actual housing price
6

y = 3.3x - 182
1000

800

600

400
line does not fit
perfectly Cost
200

0
0 50 100 150 200 250 300 350 400

Example: Supervised Learning - Classification Problem

we can
convert this
to
probability
using
sigmoid fn
A Classification Problem is an algorithm that takes as input the (x1,x2) coordinates of a point
and outputs whether the point is to be classified as O or X. The performance of the
algorithm can be defined as the % of the points that are correctly classified. In reality there
are more potential performance KPIs that we will discuss later.

Binary Classification: we only have 2 classes: y={0,1} equivalent to {Obese, not

Obese}
Multi-Class Classification (Softmax) where y={1, 2,…, k} example handwritten
digits classification 0 9

A boundary can also be non-linear

The ML learning classification problem can be defined as follows:

xi: feature (weight of a person in this case)

(xi,yi): training data set – i-th sample
Yi label (or class) – in our case Obese (y=1) or not Obese (y=0)
h(Xi)=yip hypothesis or prediction – a function (probability distribution) learned
during the training phase allowing us to predict the label ynew of xnew, not part of the
training data.

The examples that we discussed are examples of supervised learning because we fed the
computer with features and the associated label. In supervised learning, we teach the
machine ‘by example’. We feed the machine with examples so that the machine can learn
patterns in these examples. Once the patters are understood by the machine, we can have
the machine apply the patters learned on new data.

Another set of algorithms, like clustering (K-Means), uses unsupervised learning. The
computer is fed with unlabeled data. Based on a few parameters you ask the computer to
find some structure in the data.

In reinforcement learning (RL), we let the machine (an agent) learn from experience as it
observes ‘a world’. In RL, data scientists need to design the algorithm, the world in which
the agent operates and the reward structure. The machine’s goal is to select a sequence of
actions that maximizes the reward.
RL brings its own set of challenges such as:

1. How to design a reward structure that takes into account future rewards and avoid
short term thinking? One option is to work with ‘discounted’ rewards
2. How can we let the agent learn from experience without compromising safety? It is
unthinkable that we would let an autonomous vehicle ‘gain experience’ from
operating on a public road. In this case, simulation becomes key.
3. Do we allocate rewards along the process of ‘gaining experience’ or do we only
allocated a reward at the end when the mission is completed successfully?

World

Observation
Actions
& Reward

Agent

Currently, reinforcement learning is mostly a research area.

Next session: Lesson 2

We will dedicate next session to the most widely used optimization technique: Gradient
Descent! It makes sense to look at the math refresher on Calculus and Linear Algebra before
embarking on Gradient Descent. Topics that will be covered are:

1. Gradient Descent, Multivariate Gradient Descent, Stochastic Gradient Descent, Mini

Batch Stochastic Gradient Descent,
2. Optimizers: Momentum, RMSprop, Adam, Adagrad, Newton’s method