aps360 artificial intelligence fundamentals
TRANSCRIPT
![Page 1: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/1.jpg)
APS360 Artificial Intelligence Fundamentals
Lisa Zhang
Lecture 1; May 6, 2019
![Page 2: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/2.jpg)
Agenda
First Hour:
I IntroductionI MotivationI Logistics
Second Hour:
I Biological and artificial neuronsI Building an artificial Pigeon
![Page 3: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/3.jpg)
Welcome to APS360!
![Page 4: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/4.jpg)
Introduction
I Instructor: Lisa ZhangI Email: [email protected]
I Please prefix email subject with ‘APS360’I Office hours: Thursday 3pm-4pm and by appointment
![Page 5: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/5.jpg)
About your instructor
Before I started teaching, I was. . .
I a masters student doing research in Machine LearningI a senior data scientist at an advertising technology companyI a startup founder of a data visualization companyI a software developer intern in various Silicon Valley companies,
e.g. Facebook, ContextLogic (Wish)
I studied. . .
I machine learning at UofT (supervised by Prof. Richard Zemel,Prof. Raquel Urtasun)
I pure math at UWaterloo
Ask me about anything outside of class, or empty office hours!
![Page 6: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/6.jpg)
About you
Introduce yourself to the people sitting in your pod:
I What is your name?I What is your area of study?I Why are you here?
![Page 7: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/7.jpg)
Survey: demographics
I Year of study:I 40% - 2rd yearsI 40% - 3rd yearsI 10% - 4th yearsI 10% - other
I Area of study:I Engineering ScienceI Electiacal & Computer EngineeringI Mechanical engineeringI Others
![Page 8: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/8.jpg)
Why did you take this course?
I AI Minor (50%)I AI Certificate (30%)I Practical Skills (65%)I For Fun (50%)
![Page 9: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/9.jpg)
Previous ML and Neural Networks courses?
I Almost everyone have not
![Page 10: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/10.jpg)
Summer Activities
I Taking another course ~60%I Working full time ~70%
![Page 11: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/11.jpg)
Interest in Machine Learning
I Application to another field: ~70%I Becoming a data scientist: ~40%I Machine Learning Research: ~30%
![Page 12: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/12.jpg)
Programming Proficiency
I Python Experience: ~70%I Jupyter Notebooks: ~30%I PyTorch: ~2%
I will assume that you have Python experience. (e.g. loops,conditionals, functions, etc)
![Page 13: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/13.jpg)
What do you know about AI?
What is the difference between:
I Artificial Intelligence,I Machine Learning, andI Deep Learning?
![Page 14: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/14.jpg)
AI vs ML vs DL
Artificial Intelligence: Create intelligent machines that work andact like humans.
Machine Learning: Find an algorithm that automatically learnsfrom example data.
Deep Learning: Using deep neural networks to automatically learnfrom example data.
![Page 15: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/15.jpg)
Relationship
![Page 16: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/16.jpg)
![Page 17: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/17.jpg)
Why machine learning?
For many problems, it is difficult to program the correct behavior.
![Page 18: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/18.jpg)
Types of Machine Learning Problems
I Supervised LearningI RegressionI Classification
I Unsupervised LearningI Reinforcement LearningI (. . . and more)
![Page 19: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/19.jpg)
Supervised Learning Task
Supervised Learning: learning a function that maps an input toan output based on example input-output pairs
Examples:
I Age prediction given a headshot:I Input: headshot imageI Output: person’s age
I Sentiment classification given a tweet:I Input: tweet textI Output: whether the tweet is happy or sad
![Page 20: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/20.jpg)
Supervised Learning Task
I Regression: when the output is a continuous valueI e.g. age prediction
I Classification: when the output is a categorical valueI e.g. sentiment classification
![Page 21: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/21.jpg)
Unsupervised LearningUnsupervised Learning: learning the structure of some(unlabelled) data
Example:
I clusteringI generating new imagesI style transfer
![Page 22: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/22.jpg)
Reinforcement LearningReinforcement Learning: learning what actions to take tooptimize long-term reward.
Example:
I playing a video game
![Page 23: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/23.jpg)
Deep Learning Caveats: Interpretability
Figure 1: from https://xkcd.com/1838/
![Page 24: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/24.jpg)
Deep Learning Caveats: Adversarial Examples
![Page 25: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/25.jpg)
Deep Learning Caveats: Fairness
![Page 26: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/26.jpg)
Example
![Page 27: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/27.jpg)
Course Coverage
I We will focus exclusively on neural networks and deep learning.I Mostly supervised learningI Some unsupervised learningI A tiny bit of reinforcement learning
![Page 28: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/28.jpg)
Course Philosophy
I Top-down approachI Learn by doingI Explains the entire system firstI Details in future courses
I We will introduce very little mathI Focus on implementation and software skillsI Focus on communication skills
![Page 29: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/29.jpg)
Course Website
https://www.cs.toronto.edu/~lczhang/360/
![Page 30: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/30.jpg)
Course Components
I Lectures: Monday (2 hr), Thursday (1 hrs)I Labs: Thursday (1 hr), lead by a TA, in the first half of the
courseI Project: Implementation project in the second half of the
courseI Notes/Readings: Notes and readings are posted weekly.I Any material covered in lectures / tutorials / readings is fair
game for the midterm, and final term test.
![Page 31: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/31.jpg)
Teaching Assistants (in alphabetical order)
All your TAs are machine learning researchers.
I Andrew JungI Huan LingI Farzaneh MahdisoltaniI Jake Snell
![Page 32: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/32.jpg)
Grade Breakdown
I Labs: 15%I Project: 30%I Midterm: 20%I Final Term Test: 35%
![Page 33: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/33.jpg)
Labs
I One per week in the first half of the courseI Done individually – must be your own workI You will be given lab time to complete the lab workI Typically, you can submit your lab by the following Weds 9pm
Late Policy:
I Penalty-free grace period of 1 hours past the deadlineI Penalty of 20% between 1 hour and 24 hours past the deadlineI No assignments are accepted after the 24 hour deadlineI The submission time is based on Quercus clock, and the latepolicy is strict
If you have accessibility concerns, speak to me as early as possible.
![Page 34: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/34.jpg)
Software
I Python 3.6I NumPyI PyTorchI Jupyter Notebooks
All lab handouts will be Jupyter notebooks
All assignments are tested on the Google Colab environment.
![Page 35: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/35.jpg)
Course Project
Work in a group of 3 to build a useful machine learning system.
Everyone must contribute to all parts of the project to earn a grade.
I Project Proposal: 3%I Progress Meeting with TA Mentor: 3%I Progress Report: 4%I Presentation: 10%I Project Repository: 10%
![Page 36: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/36.jpg)
Midterm
I June 20th, 6pm-8pmI Length: 110 minutesI Location: TBDI No aids permitted
![Page 37: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/37.jpg)
Final Term Test
I August 1th, 6pm-8:30pmI Length: 150 minutesI Location: TBDI No aids permitted
![Page 38: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/38.jpg)
Schedule (Before Reading Week)
Week Content Lab
1 Introduction, Basic Neural Networks PyTorch2 Neural Network Training Cats vs Dogs3 Multi-class Classification Data Collection4 Convolutional Neural Networks Gesture Recognition5 Deconvolutions and Autoencoders Autoencoder6 Language Models and RNNs Spam Detection7 Midterm Week N/A
![Page 39: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/39.jpg)
Schedule (After Reading Week)
Week Content
8 Text Generation9 Generative Adversarial Networks10 Reinforcement Learning11 Ethics in AI12 Exam13 (Project)14 Presentations
![Page 40: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/40.jpg)
Away
I I am away May 14 - 26I Your TA Jake will be covering the lectures in my absense
![Page 41: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/41.jpg)
Questions?
![Page 42: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/42.jpg)
Using Pigeons to Detect Cancer
A new study suggests that the common pigeon can reliablydistinguish between benign versus malignant tumors and, in doing so,could help researchers develop better cancer screening technologies.https://www.scientificamerican.com/article/using-pigeons-to-diagnose-cancer/
![Page 43: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/43.jpg)
Training Pigeons
https://www.youtube.com/watch?v=flzGjnJLyS0
1. Show an image of a magnified biopsy to a pigeon2. Pigeon pecks at one of two answer buttons (cancer or
not-cancer)3. If pigeon picks correctly, reward pigeon with a tasty food pellet
![Page 44: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/44.jpg)
Training an Artificial Neural Network
We need to answer similar questions:
1. How will we reward the pigeon/network?2. How do we train the pigeon/network quickly and efficiently?3. How do we know the pigeon/network didn’t just memorize the
photos?4. Are there ethical issues in trusting a pigeon/network to detect
cancer?
![Page 45: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/45.jpg)
How do pigeons work?
![Page 46: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/46.jpg)
Neuron
![Page 47: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/47.jpg)
Neuron (Simplified)
![Page 48: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/48.jpg)
Neuron Anatomy
I The dendrites, which are connected to other cells thatprovides information.
I The cell body, which consolidates information from thedendrites.
I The axon, which is an extension from the cell body that passesinformation to other cells.
I The synapse, which is the area where the axon of one neuronand the dendrite of another connect.
![Page 49: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/49.jpg)
Synapse
![Page 50: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/50.jpg)
Synapse
I Small voltage difference between inside and outside of cellI When a neuron receives “information” in its dendrites, the
voltage difference along that part of the cell lowers.I If the total activity in a neuron’s dendrites lowers the voltage
difference enough, the entire cell depolarizes and the cell fires.I The voltage signal spread along the axon and to the synapse,
then to the next cells.
![Page 51: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/51.jpg)
Neural Decoding
What does it mean when a particular neuron fires?
Neuron can fire in response to. . .
I retinal cellsI certain edges, lines, angles, movementsI hands and faces (in primates)I specific people like Jennifer Aniston (in humans)
![Page 52: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/52.jpg)
Grandmother Cell
I A neuron that represents a complex but specific concept orobject
I Its existence is contested
![Page 53: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/53.jpg)
Distributed encoding
The idea that neuron firing patterns encode information only in adistributed fashion
![Page 54: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/54.jpg)
Artificial Pigeon Brain
I Start with an output “grandmother cell” that represent theconcept that we want to predict.
I Also, start with input neurons that activate with each pixelI Connect input to outputs
![Page 55: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/55.jpg)
Biological Neuron Connectivity
![Page 56: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/56.jpg)
Artificial Neuron Connectivity
I add a hidden layer that don’t have specific meaningI fully-connected, feed-forward network
![Page 57: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/57.jpg)
Modelling Individual Neurons
I x1, x2, ... = the neurons activation of input layer neuronsI h1 = the neuron activation of a hidden layer neuronI y = the neuron activation of the output layer neuron
h1 = σ(b1 +∑
i w1,ixi)
I w1,i = a weight summarizing the connectivity of neurons xiand h1
I b1 = a bias summarizing the activation requirement of theneuron h1
![Page 58: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/58.jpg)
Linear Algebra
I x - input layer activationsI h - hidden layer activationsI y - output unit
h = σ(b + Wx)
I W = a matrix of weightsI b = a vector of biases
![Page 59: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/59.jpg)
But what do the neurons mean?
I Use xi to encode the inputI e.g. biopsy imageI like the neurons that are connected to the receptors in the eye
I Use y to encode the output (of a binary classification problem)I e.g. cancer vs. not cancer
![Page 60: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/60.jpg)
Start with a problem (task)
I Input: An 28x28 pixel imageI Output: Whether the digit is a small digit (0, 1, or 2)
I output=1 means that the digit is smallI output=0 means that the digit is not small
![Page 61: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/61.jpg)
Let’s write some code!
![Page 62: APS360 Artificial Intelligence Fundamentals](https://reader030.vdocuments.net/reader030/viewer/2022012809/61bf49b0dd200c7ffb07fed3/html5/thumbnails/62.jpg)
Next Class:
I We’ll build an artificial pigeon together