Introduction to Machine Learning

class: center, middle

## Machine Learning: history, applications, recent successes

Hugo RICHARD

.affiliations[
  ![Criteo](images/criteo.jpeg)
]

.credits[Originally written by Arthur Mensch and Zaccharie Ramzi]
---

## Machine learning crash course

- Applications, history, limits

- General concepts, goals and caveats

- Mathematics:

- Statistics/ Linear algebra / Optimisation

- Programming:

- Numerics in Python / Scikit-learn / Deep learning

---
## Team

Come Fiegel (ENSAE)

Hugo Richard (Criteo)

---
## Course organisation

### <span style="color:blue">`https://data-psl.github.io/lectures2024/`</span>

Wednesday to Friday and Monday:

Course: 9h - 10h30, 11h - 12h30
      Lab: 14h - 17h

- Based on Python notebooks (bring your laptop)
  - Colab
  - Teachers will be available on Discord

---
## Plan of this lecture

- What is machine learning ? A brief history

- Broken locks and recent successes
  - Computer vision
  - Text analysis
  - Speech analysis
  - Game playing
  - Human behavior modelling
  - Biology

---

## What is machine learning ?

- Term that has become common all the way up to general media (since 2015)

- Often heard along the words
  - "Artificial Intelligence"/"Deep learning"

- A century-old concept that I'll try to introduce

---

## What is a machine ?

.center[<img src="images/machine.png" width="85%" />]

Standard *computer science* paradigm. Program a.k.a. algorithm

<div class='column'><img src="images/ada_lovelace.jpg" width="100%" /> Ada Lovelace</div>

<div class='column'><img src="images/von_neumann.gif" width="100%" />Von Neumann</div>

<div class='column'><img src="images/alan_turing.jpg" width="100%" />Alan Turing</div>

<div class='column'><img src="images/edsger_dijkstra.jpg" width="100%" />Edsger Dijsktra</div>

---

## What is an intelligent system ?

.center[<img src="images/intelligence.png" width="85%" />]

If the agent is a machine: *artificial intelligence* (term coined in 1956)

---

## How do we become intelligent ?
.center[<img src="images/learning1.png" width="85%" />]

.center[**Learning by trial and error**]

---

## How do we become intelligent ?
.center[<img src="images/learning2.png" width="85%" />]

.center[**Learning by organizing new knowledge**]
---

## How do we become intelligent ?
.center[<img src="images/learning3.png" width="85%" />]

.center[**Learning by comparison with a reference**]

---
### What do we know about intelligence ?

- Psychology/behavior study: learning processes

- Neuroscience: better understanding of the brain

- Humans use the three type of learning above (trial & error, knowledge, reference)

.left-column[<img src="images/brain.png" width="50%" />]
.right-column[<img src="images/neuron.png" width="100%" />]

---
## How can a machine learn ?

Supervised learning process

.center[<img src="images/learning3.png" width="85%" />]

Implementing a program that learns from supervision

.center[<img src="images/machine_learning.png" width="85%" />]

---
## Machine learning as a field

- A subfield of "artificial intelligence" (computer science), where *intelligence* is acquired from *data*

.center[<img src="images/machine_learning.png" width="85%" />]

- Extract information from data: relies on statistics

- Training relies on optimization and mathematics

---
## A brief history

#### AI boom 1950-1970

- Perceptron: a first attempt to classify images  (1958)

- AI at the time was mostly symbolic reasoning (not from data)

- AI booming, until 1970

#### AI winter 1970-1980

- Too hard a problem ! Researcher were over optimistic.

<div class='column'><img src="images/minsky.jpg" width="60%" /><br>Marvin Minsky</div>

<div class='column'><img src="images/rosenblatt.jpeg" width="60%" /><br>Frank Rosenblatt</div>

<div class='column'><img src="images/shannon.jpg" width="60%" /><br>Claude Shannon</div>

---
## A brief history
#### 1980s : Progress in modelling: neural-networks, training techniques

.center[<img src="images/neural_networks.png" width="35%" />]

#### 1990s-2000s:
      - AI shifts from symbolic reasoning to machine learning (from data)

- Progress in statistics: what can be learned ?

- Models simple enough to be analysed (SVM/generalized linear models)

- First applications (character recognition)

- Progress in behavior modelling: *Netflix recommendation*

---
## A brief history
  #### The 2010 turn:

- Large datasets are acquired (ImageNet + Web expansion)

- Progress in parallel computing (GPU)

#### After 2010: new AI / deep learning boom

- Complex (deep) models start working for image classification (2012)

- For text analysis (Chat GPT 2020, Le Chat 2024)

- Audio processing (2016: speech recognition)

- Scene understanding, face recognition (2014--), video generation (2023)

- Games: solving ATARI (2015--), Go (2016), ...

- Biology: AlphaFold (2020)

---
## The new boom explained

- Computational power
.center[<img src="images/gpu_tpu.png" width="35%" />]

- Powerful programming frameworks
.center[<img src="images/frameworks.png" width="50%" />]

---
## The new boom explained
- Available datasets and growing data corpus

- Money and people

.center[<img src="images/papers.jpg" width="45%" />
<img src="images/revenues.jpg" width="45%" />]

- Privately funds from the GAFAM, public funding from Chinese governement

---
## Machine learning is now everywhere

1. Computer vision

2. Speech processing

3. Text understanding

4. Game playing

5. Behavior modelling

---
## Image processing

.center[
<img src="images/images.png" style="width: 720px;" />
]

---
## Image processing

.center[
<img src="images/vision.png" style="width: 680px;" />
]