Lectures

{0}------------------------------------------------

JESSE KRIJTHE<br>CSE22A - MACHINE LEARNING

Image /page/0/Picture/2 description: The image is a close-up shot of a blue-toned surface with a pattern of small, rounded shapes, possibly grains or particles. A large, flat, and angular structure cuts across the center of the image, creating a sense of depth and texture. At the bottom of the image, the text "Delft University of Technology" is displayed in a clean, sans-serif font, indicating the origin or association of the image with the university.

{1}------------------------------------------------

Human-generated or Computer-generated?

{2}------------------------------------------------

Image /page/2/Picture/0 description: The image shows the text "Easier Task?" in a large, serif font. The text is black against a white background.

Image /page/2/Picture/1 description: A painting of a red apple and a green pear sitting on a table. The apple is on the left and the pear is on the right. The background is green. The table is yellow.

{3}------------------------------------------------

Suppose we want to build a system to solve this task (distinguishing human-generated from machine-generated paintings, or distinguishing apples and pears).

Discuss (in pairs):

• What would you need to construct such a system?
  • What physical tools would you need?
  • What additional information do you need to gather (talk to a biologist? converse with a farmer? use a search engine?)
  • How do you implement the decision making component of the system?
  • What steps would you need to take to develop such a system?

ystem?

to gather (talk to a biologist? ine?)

ing component of the system?

velop such a system?

{4}------------------------------------------------

• Task can seem very simple, but describing rules that a machine can use is very hard (think of all the exceptions, etc.)
• Idea: create a rule that constructs these rules by looking at many examples.
• "Finding patterns" in the data

ts these rules by

{5}------------------------------------------------

Study of algorithms that use example objects (data), to learn mappings from objects to desired outcomes

Image /page/5/Figure/2 description: The image shows a diagram with the word "INPUT" on the left, followed by a list of items: "Painting", "Fruit", "Photo", "Opinion poll", "Patient", "Facebook user", "Google search", "TikTok session", and "Traffic scene". To the right of "INPUT" is a black arrow pointing to a black box labeled "MODEL", indicating a process or system that takes the listed items as input.

Chess board

PREDICTION

FI ""

Human vs. Machine Made Apple vs. Pear Objects in the photo Outcome of an election Diagnosis/Optimal treatment Advertisement Relevant website Next video Brake vs. Don't Brake Best Move

{6}------------------------------------------------

Image /page/6/Picture/1 description: A close-up shot shows a plant with geometric, polygonal leaves. The leaves are predominantly green in the center of each polygon, with white edges that create a mosaic-like pattern. The plant's structure is intricate, resembling a natural fractal. In the upper left corner, the text "THE VERGE" is visible.

A Verge-generated application of a Midjourney prompt for "futuristic succulents."

AI art isn't just an experiment — it's a side hus

By Adi Robertson | @thedextriarchy | Sep 2, 2022, 10:30am EDT |

SHARE

n the past few years, art made by program like Midjourney and OpenAI's DALL-E has gotten surprisingly compelling. These The first stap in choosing the appropriate treatment programs can translate a text prompt into literally (and controversially) award-winning art. As the tools get more sophisticated, those that auna stadt al thas a amanand august stabt

$\equiv$ WIRED

BUSINESS AUG 10, 2 VILL KNIGHT

Sloppy Use of M Learning Is Cau 'Reproducibility Science

Al hype has researchers in medicine to sociology rush techniques that they don't understand-causing a wav results.

POLICY $\setminus$ TECH $\setminus$ YOUTUBE

THE VERGE

TECH 🔻

YouTube's recommo pushed election de content to election

Skeptics got more recommend fraud videos, a study found By Nicole Wetsman | Sep 1, 2022, 11:01am EDT

Image /page/6/Picture/15 description: The image shows three social media icons: Facebook, Twitter, and a share icon.

SCIENCE 🔻

Sale: Get WIRED for $29.99 $5! Subscribe now.

$\mathcal{L}$

Using machine learning to identify undiagnosable cancers

A new model that maps developmental pathways to tumor cells may unlock the identity of cancers of unknown primary.

Bendta Schroeder | Koch Institute September 1, 2022

Image /page/6/Picture/22 description: The image shows a circular diagram with various colored segments around the perimeter. The segments are arranged in a ring, with each segment consisting of multiple smaller, parallel lines. The colors of the segments vary, including shades of red, orange, yellow, green, blue, and purple. In the center of the diagram, there is a smaller circle from which lines radiate outward, connecting to the segments on the perimeter. The background of the image is a gradient of light blue, with some faint circular patterns visible.

IAV

"Machine learning tools like this one could empower oncologists to choose more effective treatments and give more guidance to their patients," says Koch Institute clinical investigator Salil Garg.

Machine learning at the edge: The AI chip company challenging Nvidia and Qualcomm

Image /page/6/Picture/28 description: The image shows a close-up of a smartphone screen displaying a lock screen with a keypad and a fingerprint icon. The background of the lock screen features a dark, swirling pattern resembling clouds or waves. The keypad consists of numbers 1 through 9, along with 0 and a delete option. The fingerprint icon is located at the bottom center of the screen and is illuminated with a blue light.

{7}------------------------------------------------

• Siri / Alexa / Google Assistant
• Recommending movies
• Targeting advertisements
• Game playing
  • Chess, Go
  • Videogames (Atari, DotA, Starcraft, etc.)
• Protein structure prediction
• Image generation
• "Text" generation (Large Language Models)
  • Chatbots
  • Code completion
  • Etc.

behavior.

1. Social Media Algorithms: Platforms like Facebook and Instagram employ machine learning to curate your newsfeed, showing you content that is likely to engage you.

queries.

1. Autonom time dec

e. Healthcare image an

and ensure that your inbox contains legitimate messages.

Fraud Detection: Banks and credit card companies use machine learning to detect unusual. transaction patterns and identify potential fraudulent activity.

52 Iannucci

Default (GPT-3.5)

Generate a list of examples of applications of machine learning in daily life

Machine learning has become an integral part of our daily lives, influencing various aspects 0 A V and industries. Here is a list of examples of its applications:

Personalized Recommendations: Services like Netflix, Amazon, and Spotify use machine. learning to suggest movies, products, and music based on your preferences and past

Virtual Assistants: Voice-activated virtual assistants like Siri, Google Assistant, and Alexa. use natural language processing and machine learning to understand and respond to user

Autonomous Vehicles: Self-driving cars rely on machine learning algorithms to make realtime decisions and navigate safely on the road.

1. Healthcare Diagnostics: Machine learning is used for early disease detection, medical

image analysis, and predicting patient outcomes based on historical data.

1. Spam Email Filtering: Email providers employ machine learning to filter out spam emails

R - Lannuarie Translation: Services like Google Translate utilize machine learning to per

Image /page/7/Picture/31 description: The image shows the Siri icon and the word "Siri" in white text below it. There is also a waveform graphic below the word "Siri".

Image /page/7/Picture/32 description: This image shows a Netflix homepage. At the top, there is a "Top Picks for Jashun" section, which includes images of "Breaking Bad", "Sing", "Fosters", "New Girl", and "Are You Here". Below that, there is a "Trending Now" section, which includes images of "Shameless", "The Office", "Orange is the New Black", "Ozark", "New Girl", and "Stranger Things". Below that, there is a "Because you watched Narcos" section, which includes images of "Surviving Escobar", "Comorroh", "Escobar", "Suburra", "Alias JJ", and "Anthony Bourdain Parts Unknown". At the bottom, there is a "New Releases" section, which includes images of "Beyond Stranger Things", "Moana", "The Mist", "The Babysitter", "Riverdale", and "Doctor Strange".

Image /page/7/Picture/33 description: The image shows a screen capture of a Go game between AlphaGo and Lee Sedol. On the left side of the screen, there is a view of the Go board with black and white stones placed on it. Next to the board, there are two bowls, one containing white stones and the other containing black stones. Above the board, there is a timer that reads "00:05:30" next to the text "ALPHAGO". On the right side of the screen, there is a video feed of Lee Sedol. Below the video feed, there is a timer that reads "00:28:28" next to the text "LEE SEDOL". Above the video feed, there is a logo that reads "Google DeepMind Challenge Match".

Image /page/7/Picture/34 description: The image is a colorful 3D rendering of a protein structure. The protein is depicted as a ribbon diagram, with different colors representing different regions or secondary structures. The color spectrum ranges from blue to red, indicating a gradient along the protein chain. The structure is complex, with several alpha-helices and beta-sheets visible. The protein is folded into a compact, globular shape, with various loops and turns connecting the secondary structure elements. The background is plain white, which helps to highlight the protein structure.

{8}------------------------------------------------

DeepMind's Deep Q-Network used to play Atari's Breakout game

Image /page/8/Picture/2 description: The image contains the words "ay" and "me" in a cursive font. The word "ay" is above the word "me".

{9}------------------------------------------------

• Artificial Intelligence / Signal Processing / Multimedia
• Data Analysis
• Scientific ML

{10}------------------------------------------------

After successfully completing this course, you are able to:

• explain the basic concepts and algorithms of machine learning and their underlying statistical concepts.
• implement, apply and evaluate basic ML algorithms in Python.
• explain the concept of and identify (implicit) bias in data and ML algorithms. (Focus in week 5)

lgorithms of machine statistical concepts.

basic ML algorithms in

tify (implicit) bias in s in week 5)

{11}------------------------------------------------

THE COURSE

{12}------------------------------------------------

Image /page/12/Picture/0 description: The image shows the word "People" in a bold, sans-serif font. The text is black and stands out against a white background.

Lecturers

Image /page/12/Picture/2 description: A portrait of a smiling man with short brown hair and blue eyes, set against a dark blue background within a white circle. He has a light beard and is wearing a gray jacket.

Jesse Krijthe

Image /page/12/Picture/4 description: A woman with short gray hair is smiling at the camera. She is wearing earrings and a dark shirt. The background is a light gray color.

Gosi

TAs

Aditi Rawat Aleksander Buszydlik Aleksandra Andrasz Aratrika Das Arnav Chopra Bernadett Bakos

Gosia Migut

Image /page/12/Picture/9 description: A man with curly dark hair and glasses is centered in a circular frame. He is wearing a dark shirt with a subtle pattern. The background is a plain, light color.

David Tax

Iulia Aldea Kirill Zhankov Rares Bites Sagar Chethan Kumar Saunaq Chakrabarty Yousef Bakri

{13}------------------------------------------------

• Course consists of 8 topics/weeks
• Each week corresponds to a topic
  • Lecture 1 on Monday
  • Lecture 2 on Tuesday morning
  • Lab session on Tuesday afternoon

{14}------------------------------------------------

• Read the reading material (mostly based on Bishop's Pattern Recognition & Machine Learning)
• Attend lectures and revise lecture notes $\bullet$
• Practice weekly theory questions in Weblab (<u>https://weblab.tudelft.nl/cse2510/2024-2025</u>)
• Practice implementation questions in lab notebooks $\bullet$
• Engage with colleagues on <u>https://answers.ewi.tudelft.nl/</u> $\bullet$
• Exam: mix between theory and implementation questions $\bullet$
• To prepare for exam: Weblab practice exams

Image /page/14/Picture/9 description: The image shows the word "ace" in a bold, sans-serif font. The color of the word is a bright, slightly desaturated orange.

{15}------------------------------------------------

Image /page/15/Picture/1 description: The image is a book cover with the title "PATTERN RECOGNITION AND MACHINE LEARNING" in bold, gold letters. Below the title, the author's name, "CHRISTOPHER M. BISHOP," is written in blue. The background of the cover features an abstract pattern of swirling brown and gold colors, resembling the surface of water with reflections.

Freely author

Some w other r

The ma the lea

Please use the reading materials!

Freely available from Brightspace and the author's website.

Some weeks we supplement the book with other reading materials.

The math can be challenging, but this is part of the learning goals.

Image /page/15/Picture/7 description: The image contains the word "of".

{16}------------------------------------------------

• https://weblab.tudelft.nl/cse2510/2024-2025/
• Should be available after the lecture
• <u>Highly</u> recommended to make them
• Helps you understand and practice with the material
• Many are similar to the style of the exam questions

{17}------------------------------------------------

• Each week, we provide an assignment in the form of a Python notebook that contains insight and implementation questions.
• Suggest you work on it in pairs
• TAs and lecturer are available to help you with questions
• We use Queue for management during the weekly scheduled lab sessions
• Answers will not be provided, so use the TAs to discuss exercises
• One of the exam questions will be based directly on the lab assignment $\bullet$
• This week's lab: focussed on familiarising yourself with (matrix computations in) Python

the form of a Python notebook n questions.

{18}------------------------------------------------

• 3 hour Weblab exam on campus
• 100% of the final grade
• Note: make sure you have enrolled in Weblab before the exam (and check it is the correct edition of the course)
• Mix of theory and implementation questions
• Access to Python documentation
• <u>No</u> access to your notes
• See Practice Exams for examples

Weblab before the tion of the course)

uestions

{19}------------------------------------------------

• 1. Ask a colleague
• 1. Ask the TAs
• 1. Ask on answers platform <u>https://answers.ewi.tudelft.nl/</u>
• 1. If this failed and the question is unique to your situation: email us at ml-cs-ewi@tudelft.nl

is unique to your situation: nl

{20}------------------------------------------------

{21}------------------------------------------------

Image /page/21/Figure/0 description: The image is a diagram illustrating different machine learning topics. The diagram starts with "Machine Learning" at the root, which branches into "Supervised" and "Unsupervised". "Supervised" branches into "Classification" and "Regression". "Classification" branches into "Generative" and "Discriminative". "Generative" branches into "Linear" and "Non-linear". "Linear" branches into "LDA" and "Nearest Mean". "Non-linear" branches into "k-NN", "QDA", "Naive Bayes", and "Parzen". "Discriminative" branches into "Linear" and "Non-linear". "Linear" branches into "Logistic Regression" and "SVC". "Non-linear" branches into "Decision Trees" and "MLP". "Unsupervised" branches into "Clustering" and "Dimensionality reduction". "Clustering" branches into "Hierarchical" and "K-means". "Dimensionality reduction" branches into "PCA". There are also two labels, "Non-parametric" and "Parametric".

Image /page/21/Picture/1 description: The image contains two rounded rectangles. The top rectangle is orange and contains the text "metric". The bottom rectangle is blue and also contains the text "metric".

{22}------------------------------------------------

• Try to give more concrete examples during lectures
• Revised presentation of difficult notation for backpropagation and PCA.

ples during lectures t notation for

{23}------------------------------------------------

• Math notation may seem opaque sometimes: start practicing now when things are still simple, or we'll regret it later.
• Focus of the course is on good understanding of the fundamentals, we build on this in later courses to work on more complex models and interesting applications.
• Let us know when you think we can improve something (preferably during the course, but also after).

sometimes: start still simple, or we'll

understanding of the n later courses to work teresting applications.

can improve e course, but also

{24}------------------------------------------------

Multimedia content (text, images, speech, video) is an important source of knowledge that is useful to people.

Building systems that allow people to make use of that knowledge, and evaluating such systems from both a technological and human point of view.

Machine Learning Topics covered:

• K-means clustering for e.g. image geo-location, visual words representation, audio-visual video segmentation,
• Gaussian Mixture Model for e.g. Acoustic Modelling
• Naive Bayes Classifier and Hidden Markov Models (not covered in ML) for e.g. Automatic Speech Recognition
• Linear classifier for e.g. video shot change detection,
• k-nearest neighbour classifier for e.g. phoneme recognition from Speech
• Parameter training for e.g. weight training for late multi-modal data fusion
• System Evaluation e.g. accuracy, precision, recall, f-measure, ROC, AUC. $\bullet$ but also metrics for information retrieval systems Mean Average Precision (not covered in ML)

imedia Analysis

K-means clustering of interest points

Image /page/24/Picture/15 description: The image shows three frames of a person walking. The person is wearing a jacket and pants. The person's head and legs are circled in blue. The labels c2, c4, and c1 are present in the image.

Image /page/24/Picture/16 description: The image shows two cropped sections of a grayscale image. The first section shows a person walking, with two blue circles highlighting areas on their leg. The circles are labeled 'c1' and 'c2'. The second section shows another person walking, with two white circles highlighting areas on their leg and waist. The circles are labeled 'c3' and 'c4'.

Image /page/24/Picture/17 description: The image shows a person walking. The person is wearing a jacket and pants. There are blue circles around the person's waist and arm. There are white circles around the person's feet.

Image /page/24/Picture/18 description: The image shows a person walking. The person is wearing a suit and tie. The person is walking from left to right. There are three blue circles on the person's head, shoulder, and waist. There is a white label "c1" near the person's waist.

Image /page/24/Picture/19 description: The image shows a series of six grayscale images depicting a person walking. Each image has blue circles highlighting points of interest on the person and the ground. Some of the circles on the ground are labeled with 'c1', 'c2', 'c3', and 'c4'. The title above the images reads 'Classification of interest points'.

Image /page/24/Figure/20 description: The image shows a diagram with text labels. On the left side, there are two rectangular shapes labeled "R" and "R" with a text label "Range overlap" and an arrow pointing to the overlapping area between the two rectangles. On the right side, there is a graph with the y-axis labeled "z". The graph is divided into three sections labeled "Shot 1", "Shot 2", and "Shot 3". The graph shows some representative feature variation. The boundaries between the shots are labeled "Shot boundary".

Image /page/24/Figure/21 description: The image shows a diagram illustrating the computation of P(o | q) using a two-component Gaussian Mixture Model (GMM) with Expectation-Maximization (EM). On the left, there is a scatter plot with blue points and contour lines, representing the GMM fit. An arrow points from this plot to a Markov Model (M) diagram on the right. The Markov Model consists of six states (1 to 6) represented by circles, with arrows indicating transitions between states. The transition probabilities are labeled as a_ij, where i and j are the state numbers. Below the Markov Model is an 'Observation Sequence' labeled o1 to o6, each represented by a rectangle. Dashed lines connect the states to the observation sequence, with emission probabilities labeled as b_i(o_i), where i is the state number and o_i is the observation at that state.

Image /page/24/Figure/22 description: The image shows a waveform at the top and a graph at the bottom. The waveform has dashed vertical lines and arrows pointing to O1, O2, and O3. The graph shows four curves with different colors. The legend indicates the parameters of each curve: blue curve with mu=0 and sigma^2=0.2, red curve with mu=0 and sigma^2=1.2, yellow curve with mu=0 and sigma^2=5.2, and green curve with mu=-2 and sigma^2=0.5.

Slide courtesy of Hayley Hung

Image /page/24/Picture/24 description: The image contains the name "Wesley Hung" in a stylized font at the bottom.

{25}------------------------------------------------

QUESTIONS?

{26}------------------------------------------------

Back to Machine Learning...

Learning...

{27}------------------------------------------------

Image /page/27/Picture/0 description: The image shows the word "Example" in a serif font.

Image /page/27/Picture/1 description: A still life painting shows a red apple and a green pear sitting on a table. The apple is on the left and the pear is on the right. The background is a dark green color.

{28}------------------------------------------------

Machine Learning tries to identify regularities in the world, by learning from examples ("data"). These regularities should generalise: work beyond the specific examples the model has seen before.

{29}------------------------------------------------

• 1. Many tasks are too complicated to explicitly encode by hand: learn a mapping from input to output using examples.
• 1. Learning about the world using data, to use objective evidence to base our decisions on.

Connections to statistics, signal processing, optimisation

to explicitly encode n input to output

g data, to use decisions on.

{30}------------------------------------------------

(On average, how bad is my model)

Е0[[д, 0 Loss function

Image /page/30/Picture/4 description: The image shows the word "Model" with the text "(Mapping from inputs to outputs)" below it. An arrow points from the word "Model" to the expression "(g, O)]". Another arrow points from the expression "(g, O)]" to the word "Object" with the text "(Thing that I want to make a prediction for)" below it.

(How much does a mistake "cost"?)

{31}------------------------------------------------

Image /page/31/Picture/0 description: The image shows a large number of red and yellow apples floating in water. The apples are arranged in a dense, grid-like pattern, filling the frame. The water is dark and reflects the colors of the apples. In the bottom right corner, there is a glimpse of a metallic structure, possibly part of a machine or container holding the water and apples.

{32}------------------------------------------------

(On average, how bad is my model)

Е0[[д, 0 Loss function

Image /page/32/Picture/4 description: The image shows the word "Model" with the text "(Mapping from inputs to outputs)" below it. An arrow points from the word "Model" to the expression "(g, O)]". Another arrow points from the expression "(g, O)]" to the word "Object" with the text "(Thing that I want to make a prediction for)" below it.

(How much does a mistake "cost"?)

{33}------------------------------------------------

Image /page/33/Figure/1 description: The image shows a diagram of a process flow. The process starts with an "Object", which leads to "Feature values". The "Feature values" then lead to a "Prediction", which in turn leads to a "Performance measure". There are three arrows pointing upwards, one below the arrow between "Object" and "Feature values" labeled "Representation", one below the arrow between "Feature values" and "Prediction" labeled "Model", and one below the arrow between "Prediction" and "Performance measure" labeled "Evaluation".

{34}------------------------------------------------

• How do I represent my object to use it in a mathematical function? What measurements do I do to represent the object?
• We call these measurements the features (variables, covariates), and depict them using a vector. We call this vector space the feature space. For example:

$\begin{bmatrix} x^{1} \\ x^{2} \\ \vdots \\ x^{M} \end{bmatrix}$

$\begin{bmatrix} 0.1 \\ -100 \\ 25 \\ 0 \end{bmatrix}$

{35}------------------------------------------------

What measurements can I do that will help me distinguish the Apple from the Pear?

{36}------------------------------------------------

Image /page/36/Picture/0 description: The image is a scatter plot of different types of fruits, with the x-axis representing weight and the y-axis representing roundness (length/width). The plot contains a variety of fruits, including apples and pears, each represented by a point on the graph. The apples are generally clustered towards the right side of the plot, indicating a higher weight, while the pears are scattered more towards the top, suggesting a higher roundness or length-to-width ratio. The fruits are of different colors, including red, green, and yellow.

{37}------------------------------------------------

Image /page/37/Picture/0 description: The image is a scatter plot of fruits, with the x-axis representing weight and the y-axis representing roundness (length/width). There are two types of fruits in the plot: apples and pears. Apples are represented by orange circles, while pears are represented by blue triangles. The plot shows that apples tend to be heavier and rounder than pears.

{38}------------------------------------------------

Image /page/38/Figure/1 description: The image is a scatter plot with two axes. The x-axis is labeled "Weight" and the y-axis is labeled "Roundness (Length/Width)". There are two types of data points plotted on the graph: triangles and circles. The triangles are blue and the circles are orange. The triangles are generally located towards the left side of the graph, while the circles are generally located towards the right side of the graph.

{39}------------------------------------------------

Image /page/39/Figure/1 description: The image shows a diagram of a process with four stages. The first stage is labeled "Object" and is represented by a gray rectangle. An arrow points from the "Object" stage to the second stage, which is labeled "Feature values" and is also represented by a gray rectangle. Below the arrow connecting the first two stages is the word "Representation" with an arrow pointing up to the arrow. An arrow points from the "Feature values" stage to the third stage, which is labeled "Prediction" and is represented by a gray rectangle. Below the arrow connecting the second and third stages is the word "Model" with an arrow pointing up to the arrow. An arrow points from the "Prediction" stage to the fourth stage, which is labeled "Performance measure" and is represented by a gray rectangle. Below the arrow connecting the third and fourth stages is the word "Evaluation" with an arrow pointing up to the arrow.

{40}------------------------------------------------

Image /page/40/Figure/1 description: This image is a scatter plot with two features, "Roundness (Length/Width)" and "Weight". The plot contains two classes of data points, represented by blue triangles and orange circles. The data points are separated by a decision boundary, which is a piecewise linear function. The region to the left of the boundary is colored light blue, and the region to the right is colored light orange. The x-axis represents "Weight", and the y-axis represents "Roundness (Length/Width)".

A function that assigns an output to every location in the input space

g: \mathbb{R}^2 \to \{\text{apple}, \text{pear}\}\

Image /page/40/Figure/4 description: The image contains the character sequence ".r}" in bold font.

{41}------------------------------------------------

Image /page/41/Figure/0 description: The image is a scatter plot with two axes. The y-axis is labeled "Roundness (Length/Width)". There are two categories of data points, triangles and circles. The triangles are blue and the circles are orange. The triangles are mostly in the upper left of the plot, and the circles are mostly in the lower right of the plot. A dotted line separates the two categories of data points. The area above the line is light blue, and the area below the line is light orange. The word "Decision" is written in a cursive font near the dotted line.

Image /page/41/Picture/2 description: The image shows a decision boundary separating two classes of data points. The x-axis is labeled "Weight". There are several orange circles and a blue triangle plotted on the graph. The decision boundary is a dotted line that separates the blue triangles from the orange circles. The area above the decision boundary is shaded blue, while the area below the decision boundary is shaded orange. The text "Decision Boundary" is written along the dotted line.

{42}------------------------------------------------

• Choosing informative features for the task makes it easy to discriminate between different classes
• If we do not measure the right things, it is impossible to have a model that performs well.
• See also: Bayes error, later this week

for the task makes it ifferent classes

things, it is t performs well.

eek

{43}------------------------------------------------

Image /page/43/Figure/1 description: The image shows a diagram of the steps involved in machine learning. The diagram consists of four boxes labeled "Object", "Feature values", "Prediction", and "Performance measure", connected by arrows. An arrow points from "Object" to "Feature values", from "Feature values" to "Prediction", and from "Prediction" to "Performance measure". There are also three arrows pointing upwards to each of the arrows connecting the boxes, labeled "Representation", "Model", and "Evaluation", respectively. An arc above the boxes is labeled "Label" and has a human head icon in the middle. The title of the diagram is "Steps involved in Machine Learning".

{44}------------------------------------------------

• Measure the performance: how well does the mapping solve the task?
• Measure this using a "loss function"
  • Specific to the problem we are trying to solve
• Idea here: perhaps we should count the number of mistakes.

Image /page/44/Picture/5 description: The image contains the word "ping?" in a serif font. The word is written in black and is slightly slanted to the right.

Image /page/44/Picture/6 description: The image shows a two-dimensional scatter plot with two distinct categories of data points, represented by blue triangles and orange circles. The plot is divided into two regions by a dotted line, with the upper-left region shaded in light orange and the lower-right region shaded in light blue. The x and y axes are represented by thick black lines, with an arrowhead indicating the positive direction. The blue triangles are mostly concentrated in the light orange region, while the orange circles are mostly concentrated in the light blue region. There are a few data points that are located in the 'wrong' region, i.e., blue triangles in the light blue region and orange circles in the light orange region.

Image /page/44/Picture/7 description: The image shows a two-dimensional scatter plot with two distinct categories of data points, represented by blue triangles and orange circles. The plot is divided into two regions by a dotted line, with the upper-left region shaded in light blue and the lower-right region shaded in light orange. The x and y axes are represented by black lines, with an arrow indicating the positive direction of each axis. The blue triangles are mostly concentrated in the light blue region, while the orange circles are mostly concentrated in the light orange region. However, there are some exceptions, such as a blue triangle in the light orange region and an orange circle in the light blue region.

{45}------------------------------------------------

Image /page/45/Picture/1 description: The image compares two systems, labeled "System 1: many small mistakes" and "System 2: one big mistake," in terms of their predicted and real weather conditions. Each system has two columns, "Predicted" and "Real," displaying weather icons. System 1 shows a mix of correct and incorrect predictions, with a green box highlighting a row where the predicted and real conditions match, labeled "Only correct answer." System 2 has mostly correct predictions, except for one row highlighted with a red box, labeled "Only incorrect answer," where the predicted sunny condition is incorrect compared to the real rainy condition. The image also includes the "TUDelft" logo at the bottom left and the number "67" at the bottom right.

mance Metric

Slide courtesy of Odette Scharenborg

Image /page/45/Picture/4 description: The image shows the word "harenborg" in a serif font. The letters are all lowercase.

{46}------------------------------------------------

Image /page/46/Figure/1 description: The image shows the steps involved in machine learning. The steps are object, feature values, prediction, and performance measure. The object is labeled by a person. The representation, model, and evaluation are used to improve the performance measure.

{47}------------------------------------------------

Image /page/47/Picture/0 description: The image shows the word "Learning" in a serif font. The word is in black and is centered in the image.

Image /page/47/Figure/1 description: The image is a scatter plot with two axes. The vertical axis is labeled "Roundness (Length/Width)" and the horizontal axis is labeled "Weight". There are two types of data points plotted on the graph: blue triangles and orange circles. The triangles are generally located towards the left side of the graph, while the circles are generally located towards the right side of the graph.

{48}------------------------------------------------

• What about remembering? $\bullet$
• Going beyond just remembering each specific instance, to generalising to new instances.
• We want to learn an input-output mapping that works for previously unseen objects as well.

"To think is to forget a difference, to generalize, to abstract."

• Jorge Luis Borges in "Funes the Memorious"

{49}------------------------------------------------

Image /page/49/Figure/1 description: The image is a scatter plot with two axes labeled "Roundness (Length/Width)" on the vertical axis and "Weight" on the horizontal axis. There are several data points plotted on the graph. Some data points are blue triangles, some are orange circles, and one is a red star. The blue triangles are mostly located in the upper left quadrant of the graph, while the orange circles are mostly located in the lower right quadrant. The red star is located in the upper middle of the graph.

{50}------------------------------------------------