ML - Lecture 3

Adversarial ML

Perturbing an image before giving it to the NN can alter Massively the results

---

Supervised Learning

• Regression
• Classification
• Loss Function

The supervisor tells you how wrong you are. Learning means to respect the supervision (determine the right weights such that the input corresponds to the outputs)

Learning Environment: Collection of inputs and the corresponding outputs.

---

Regression

Given the height $h$ of a person we want to predict his weight as a function oh $h$: $w(h)$, from a collection of supervised data

We need an output corresponding to a real number $\in \mathbb{R}$

---

Classification

Determine if a picture represent a number between 0-9.

We need an output corresponding to a set of numbers, more generally corresponding to a category$\in \mathbb{N}$

---

Measure the Learning Process:

In case of supervised learning a measure of the learning process can be represented by the Loss: a sort of distance between the machine-found output and the actual output:

Our objective is to minimize: $|\omega -f(h)|$, defined as the Loss Function.

There exist infinite loss functions another one can be: $(\omega -f(h){)}^2$ or $\log |\omega -f(h)|$

In case of classification the loss function becomes a little more complicated because we are working with discrete numbers (or classes), but we can still find a loss function.

---

Unsupervised Learning

• Linear Separation (data clustering)

Find patterns in the data.

---

Linear Separation

To perform data clustering we can use the geometric distance of two points:

• Given the data point $x$ and two cluster points $o_1$ and $o_2$ with define with $d(x,y)$ the distance between 2 points.
• If $d(x_1,o_1)<d(x_1,o_2)$ than $x$ will belong to the cluster $o_1$ else, to the cluster $o_2$

---

Using the right tool for the job

Using the loss function shown in the picture is actually a bad idea. In this case the machine thinks that the 3 is more similare the the 2 to the reference image.