learning applications of machine
TRANSCRIPT
![Page 1: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/1.jpg)
Applications of Machine Learning
Malcolm Reid Jr.
![Page 2: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/2.jpg)
Agenda
1. What is Machine Learning?
2. Intro to basic machine learning algorithms
3. Real-world applications of machine learning
4. Questions
![Page 3: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/3.jpg)
Artificial Intelligence vs. Machine Learning
● Artificial intelligence (AI) - the field of computer science that studies endowing machines with
human-like intelligence
● Machine learning (ML) - the field of artificial intelligence that studies machines that can learn from
data.
![Page 4: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/4.jpg)
Canonical Definition of ML
“the study of algorithms that improve their performance P at some task T with experience E” - Tom
Mitchell
Thus, for any learning task to be well-defined we need <P, T, E> to be well-defined
![Page 5: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/5.jpg)
Taxonomy of Artificial Intelligence
https://towardsdatascience.com/types-of-machine-learning-algorithms-you-should-know-953a08248861
![Page 6: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/6.jpg)
A Brief Note on Dimensionality
● Most ML processes high-dimensional data
● When we talk about dimensions in this presentation, remember we’re talking about in feature
space, not about spatial dimensions (though it may be easiest to illustrate a concept using spatial
dimensions)
● Assume that we can always measure within a dimension whether by Euclidean distance or by
Hamming distance
![Page 7: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/7.jpg)
Supervised ML Framework
![Page 8: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/8.jpg)
Basic ML Terminology
● Each discrete example we pass into our machine learning model is called either an example or an
instance● Each coordinate of an instance is called a feature● The correct classification for an example is called its label● Usually in ML, we talk about vectors and vector spaces. I’ll denote a vector with bold (e.g. x)
○ xi = the ith example
● In supervised learning, we have a training set (i.e. the set of examples that our model is trained on)
and a testing set (i.e. the set of examples that the model is tested on. These are withheld at training
time)
![Page 9: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/9.jpg)
Basic ML Terminology (contd…)
● A model is the thing we train. It takes an example and returns a prediction of its label. Also known
as a classifier
![Page 10: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/10.jpg)
ML Terminology Example
● Instance/example = a photo of a cloud
● Feature = pixel in photo
● Label = type of cloud
● Training set = set of m photos used for training our model
● Testing set = set of n - m photos used for testing our model
https://en.wikipedia.org/wiki/Cumulus_cloud
![Page 11: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/11.jpg)
Nearest Neighbors (aka “lazy learner”)
● Problem: Assume you have instances of the form (x1
, y1
) … (xn, y
n). We want to predict the label for
xi
● Simple solution: Computer stores all of the instances. When passed in new instance to predict, it
finds the closest one it has seen and outputs its label
![Page 12: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/12.jpg)
Nearest Neighbors Example
● Assume we’re in two-dimensional space. I’ll give you a training set of n points (Cartesian
coordinates), for you to simulate nearest neighbors. Then I’ll give you a test example and you
classify it.
![Page 13: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/13.jpg)
Training Examples
Feature 1 Feature 2 Label
1 3 In-bounds
2 5 In-bounds
0 0 Out-of-bounds
2 1 Out-of-bounds
![Page 14: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/14.jpg)
Quiz
● How would nearest neighbors classify (2, 3) if trained on the examples from the previous slide?
![Page 15: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/15.jpg)
Improvements to Naive Nearest Neighbors
● Assign a label based on more than just 1 nearest point (k-NN)
● Assign a label based on a weighted average of the k nearest points
● Assorted data structures and algorithms to make nearest neighbors more efficient from a
performance perspective○ E.g. Maybe we don’t need to store all of the instances
![Page 16: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/16.jpg)
Voronoi Diagram
● A given color represents the set of points closest to a point in the dataset
https://en.wikipedia.org/wiki/Voronoi_diagram
![Page 17: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/17.jpg)
Decision Trees
● Decision trees are tree-like models that ask a question at each node and choose a branch or
outcome depending on the answer to that question
![Page 18: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/18.jpg)
https://www.geeksforgeeks.org/decision-tree/
![Page 19: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/19.jpg)
Linear Regression
● Finds a line of best fit in the points
https://en.wikipedia.org/wiki/Linear_regression
![Page 20: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/20.jpg)
Perceptrons
● Very loosely based on neuron in the human brain
https://towardsdatascience.com/perceptron-the-artificial-neuron-4d8c70d5cc8d
![Page 21: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/21.jpg)
Neural Networks
● With a single perceptron, we can only train relatively simple models
● What if we combine multiple perceptrons?
x1
x2
![Page 22: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/22.jpg)
Deep Learning
● Why stop there? Deep neural network has at least 2 hidden layers
https://towardsdatascience.com/applied-deep-learning-part-1-artificial-neural-networks-d7834f67a4f6
![Page 23: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/23.jpg)
“Interesting” ideas in Deep Learning
● “ML is as much an art as it is a science” - my graduate school ML professor
● One example is dropout
![Page 24: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/24.jpg)
Dropout
Srivastava, Nitish & Hinton, Geoffrey & Krizhevsky, Alex & Sutskever, Ilya & Salakhutdinov, Ruslan. (2014). Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Journal of Machine Learning Research. 15. 1929-1958.
![Page 25: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/25.jpg)
Reinforcement Learning (RL)
● A sub-field of ML focusing on how a computer can make choices to maximize some reward function
![Page 26: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/26.jpg)
Ensemble Approaches
● What if we run several different machine learning algorithms and see what they predict
● In practice, this is very popular
![Page 27: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/27.jpg)
Applications of ML
![Page 28: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/28.jpg)
Canonical Definition of ML
“the study of algorithms that improve their performance P at some task T with experience E” - Tom
Mitchell
Thus, for any learning task to be well-defined we need <P, T, E> to be well-defined
![Page 29: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/29.jpg)
Spam Filtering
• T : classify a new mail message as either spam or not spam
• P : minimize misclassification costs (could be differential costs!)
• E : previously manually filed messages
![Page 30: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/30.jpg)
Breast Cancer Detection
• T : classify a mammogram (i.e. an X-ray of the breast) as either malignant or benign
• P : minimize misclassification costs (could be differential costs!)
• E : previously classified mammogram images
![Page 31: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/31.jpg)
Object Detection
• T : classify an image of an object
• P : minimize misclassification costs
• E : set of manually labelled images of objects
![Page 32: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/32.jpg)
Personalized Medicine
• T : given a patient’s health data, predict his or her probability of various outcomes
• P : minimize misclassification costs
• E : medical history and outcomes of prior patients
![Page 33: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/33.jpg)
Face Recognition
• T : label a photo of a face with a name from a set of names
• P : minimize misclassification costs
• E : dataset of face photos with associated names
![Page 34: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/34.jpg)
Face Generation?
Karras, T., Aila, T., Laine, S., & Lehtinen, J. (2017). Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196.
![Page 35: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/35.jpg)
Language Translation
● Huge amount of training data for models -- i.e. the corpus of works translated into multiple
languages
● E.g. Google Translate
![Page 36: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/36.jpg)
YouTube Recommended Videos
● Recommendation systems (e.g. for YouTube or Netflix) often use some form of ML
● Might use clustering to find movies that are similar to each other
● Netflix ML Competition
![Page 37: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/37.jpg)
Self-driving Cars
https://www.wired.com/story/waymo-self-driving-cars-california/
![Page 39: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/39.jpg)
Where can you learn more?
● Online resources:○ https://www.fast.ai/○ https://towardsdatascience.com/○ MOOCs (Coursera, Udemy, Udacity)○ https://techdevguide.withgoogle.com/paths/machine-learning/○ Kaggle
● Talk to me after the presentation!
![Page 40: Learning Applications of Machine](https://reader033.vdocument.in/reader033/viewer/2022051412/627dd23748b8391b64474768/html5/thumbnails/40.jpg)
Parting Thoughts on ML