the subset of AI that learns from data and adapt to changes in the environment

• - An ML-enabled system will be able to react to new inputs (never before encountered) and produce sensible outputs,

• e.g. the Boston Robotics dog will be able to navigate a path it has not seen before, the Tesla self-driving car will be able to drive down any street and the Google Home can respond to new commands and sensor readings.

• - Key idea is that ML algorithms are trained to perform an estimation or decision-making task using training data, and then are expected to perform that task well even on unseen data.

ABSTRACT VIEW OF MACHINE LEARNING: MODEL

• First, define a model linking inputs to output(s)

• This model has a set of parameters denoted 𝜃

The critical task of ML is to find the best values for the parameters 𝜃 using available data

• This is accomplished in the training phase

ABSTRACT VIEW OF MACHINE LEARNING: TRAINING/LEARNING

In training phase, use known inputs and outputs to find θ

branches

classification: spam/not spam

regression: output is continuous value price/probability

ABSTRACT VIEW OF MACHINE LEARNING: TRAINING/LEARNING

In training phase, input and output are not known so θ can not be found, so objective is to find pattern

clustering: groups similarities customers

association: looks for relationship between variables customer bought this and also bought this

ABSTRACT VIEW OF MACHINE LEARNING: TRAINING/LEARNING

In training phase, learning which relies on feedback from the environment in response to an action taken by the machine to discover the model parameters.

Set of all $$S_y$$ is finite

To categorize inputs belonging to a particular class

eg. true/false

Set of all $$S_y$$ is uncountable

To estimate the value of response variable given independent inputs

eg. predicting rainfall

$$C$$

Class of the object that is trying to identify

$$C^{\wedge}$$

Estimate of the class

BINARY CLASSIFICATION TESTS

$$\frac{TN\ +\ TP}{TN+TP+FN+FP}$$

BINARY CLASSIFICATION TESTS

fraction of truly positive samples correctly labelled as positive

$$\frac{TP}{TP+FN}$$

BINARY CLASSIFICATION TESTS

fraction of truly negative samples correctly labelled as negative

$$\frac{TN}{TN+FP}$$

BINARY CLASSIFICATION TESTS

fraction of truly positive decision that are truly positive

$$\frac{TP}{TP+FP}$$

ROC

The larger the area under the curve the ____

the classifier is

Cross Validation

• OTHER FORMS OF DATA PRE-PROCESSING

• - In sometimes have the case of one class being much more prevalent than the others. This can skew the training of the ML model. This phenomenon is known as ________

eg.

Class 0 (Non-Fraudulent): 95% of the samples

• Class 1 (Fraudulent): 5% of the samples

• To deal with this, we can perform downsampling (i.e. randomly sampling only a fraction of the majority class) and then upweighting those samples (i.e. giving them greater weight in the training calculations).

• OTHER FORMS OF DATA PRE-PROCESSING

• In dealing with ________ (i.e. non-numerical data), e.g. names, types, etc., we would normally have to convert these into a numerical representation since most ML classifiers take only numerical inputs.

• One-hot encoding is a method for transforming categorical data to numbers.