presentation.rst

css:	presentation.css
skip-help:	true
data-transition-duration:	1000

My talk at the University of Mauritius

---

id:	title-slide

Introduction to Machine Learning

Dr. Omri Har-Shemesh

30. August 2019, University of Mauritius

---

data-y:	r1000

About me

---

id:	background-physics

Background: M.Sc. in Physics

---

Background: Ph.D. in Computational Science

---

Current: Data Science / Programming @ schmiede.one

---

data-rotate-y:	0
data-x:	3000
data-y:	0

Introduction to Machine Learning

Articifial Intelligence?

---

data-rotate-y:	90
class:	wide-step

What about AI?

---

data-rotate-y:	90
data-y:	r1000

What about AI?

---

data-rotate-y:	90
data-y:	r1000

What about AI?

---

data-y:	0

---

data-rotate-y:	0

---

data-y:	r1000

So what is Machine Learning?

Machine Learning: Field of study that gives computers the ability to learn without begin explicitly programmed.

Arthur Samuel (1959)

A methodology and a collection of algorithms designed to discover and exploit meaningful patterns in raw data.

Omri Har-Shemesh (now)

---

data-y:	r1000

Is this possible?

---

class:	wide-step

Is this possible?

---

class:	very-wide-step

Is this possible?

---

class:	very-wide-step

Is this possible?

---

class:	very-wide-step

Is this possible?

Yes!

---

data-y:	0
data-x:	5000

How is this possible?

Data + Computing Power + Some Clever Ideas

---

class:	quite-wide-step
data-y:	r1000

What is data?

• Numbers (scalars, vectors, tensors)
• Categories
  • Yes/No
  • Very satisfied / Satisfied / ... / Not satisfied
  • Cat / Dog / Building
• Images
• Sound
• Text

• Web logs
• Financial records
• Website interactions
• Customer service contacts
• Social media interactions
• Machine logs
• Medical records
• ...

---

class:	wide-step

What is data?

---

Three types of machine learning

1. Supervised Learning
2. Unsupervised Learning
3. Reinforcement Learning

---

class:	wide-step

Supervised Learning

---

class:	wide-step

Unsupervised Learning

---

class:	wide-step

Reinforcement Learning

---

class:	very-wide-step white-back

---

Two types of supervised learning

1. Regression
   You're predicting a number
2. Classification
   You're predicting a category

---

Example: linear regression

• We have data

X = [x_1, x_2, x_3, \ldots]

y = [y_1, y_2, y_3, \ldots]

• We want to predict, for a given new X, the appropriate y.

---

Example: linear regression

• First step - assume a model:

y \sim aX + b + \epsilon

• Second step - define a loss function

L = \sum\limits_{i=1}^N (y_i - \hat{y}_i)^2

---

class:	wide-step

Example: linear regression

---

Example: linear regression

• First step - assume a model:

y \sim aX + b + \epsilon

• Second step - define a loss function

L = \sum\limits_{i=1}^N (y_i - \hat{y}_i)^2

• Third step - find parameters a and b that minimize L.

\frac{\partial}{\partial a} L = 0 \\ \frac{\partial}{\partial b}L = 0

---

class:	very-wide-step

Demo

---

Potential Problems

1. Data too noisy
2. Model too simple (high bias / underfitting)
3. Model too complex (high variance / overfitting)

---

Bias-Variance Tradeoff

E \left[ \left(y - \hat{f}(x)\right)^2 \right] = \left(E[\hat{f}(x)] - y
\right)^2
+ E[\hat{f}^2(x)] - E[\hat{f}(x)]^2 + \sigma^2

---

What can we do?

1. If data is too noisy - get better data! You can't solve this with ML.
2. If model too simple - complicate the model (easy).

3. Model overfitting the data?

   • Use test set to know you're overfitting.
   • Penalized loss function.
   • Add more data (make it harder to overfit).
   • Fine tune the parameters of the model.
   • Model specific tricks (e.g. Dropout layers).

---

class:	very-wide-step white-back

Are we done?

---

data-scale:	7
data-y:	3000
data-x:	4000

---

data-scale:	1
data-y:	0
data-x:	7000
class:	white-back

Introducing Python

---

data-y:	r1000

pandas

A great package for

1. Data exploration
2. Data cleaning
3. Data wrangling
4. Data visualization (together with matplotlib)
5. Data preparation

---

class:	very-wide-step

scikit-learn

---

class:	quite-wide-step

scikit-learn: small demo

def generate_data(n_points=10, eps=0.1, seed=None):
    if seed:
        np.random.seed(seed)
    if n_points == None:
        X = np.linspace(-2.2, 3.2, 1000)
        n_points = 1000
    else:
        X = np.random.uniform(low=-2, high=3, size=n_points)
        np.sort(X)
    y = X**3 - 2 * X ** 2 + 1.5 ** X + 5 + eps * np.random.normal(size=n_points)
    return X, y

---

class:	quite-wide-step

scikit-learn: small demo

# Simple Example - Linear regression with a linear feature set
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LinearRegression

X, y = generate_data(n_points=100, eps=2)
X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.1)

model = LinearRegression()
model.fit(X_train, y_train)

prediction = model.predict(X)

score = model.score(X_test, y_test)

---

class:	quite-wide-step

scikit-learn: small demo

# Let's do the same with polynomial fit
from sklearn.preprocessing import PolynomialFeatures

pf = PolynomialFeatures(degree=3)
pf.fit(X_train)
X_train_poly = pf.transform(X_train)

model = LinearRegression()
model.fit(X_train_poly, y_train)

---

class:	quite-wide-step

scikit-learn: small demo

# Or easier, combine them in a pipeline!
from sklearn.pipeline import make_pipeline

pipeline = make_pipeline(
    PolynomialFeatures(degree=3),
    LinearRegression()
)

pipeline.fit(X_train, y_train)
pipeline.score(X_test, y_test)

---

class:	quite-wide-step

---

class:	quite-wide-step

---

A small challenge :)

---

class:	very-wide-step

And a small plug...

---

data-y:	-2000
data-x:	4000
data-scale:	15

Thank you!

---

Dr. Omri Har-Shemesh, data scientist @