Skip to content
This repository has been archived by the owner on Aug 27, 2020. It is now read-only.

Latest commit

 

History

History

04_diy_neural

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
images
images
 
 
README.md
README.md
 
 
data-tutorial.md
data-tutorial.md
 
 

README.md

DIY Neural Network

Session A: Data Collection

Objectives:

  • Understand the full story of building a ML model for classification or regression.
  • Understand how data is formatted and downloaded including CSV and JSON.
  • Consider how to frame the problem and collect data.
    • Understand critical questions to ask (e.g. Who is this for? What’s the context?)
    • Understand the questions to ask about sourcing and collecting data.
    • Learn how to prepare a data set, including how to normalize and properly format it.
  • Diagram the components of a two layer "vanilla" neural network.

Tutorials

Related Projects

Reading and Viewing

Assignment 4A due Wednesday, Sept 24, 9am

  1. Find a dataset that interests you. Some ideas:
    • Something you find online. For example, take a look at Kaggle and you can find a list of datasets we've compiled on the dataset wiki.
    • Find a dataset that you collect yourself or is already being collected about you. For example, personal data like steps taken per day, browser history, minutes spent on your mobile device, sensor readings, and more.
    • A dataset that you collect by crowdsourcing data. One way to approach this is to create a google form with questions and ask friends and fellow students to complete the form.
  2. After you have found or collected your dataset, write a blog post that addresses these questions:
    • What is (are!) the data?
    • What format is the data in? (CSV, JSON, PDF, or . . )
    • What are the dimensions of the data (rows and columns)?
    • What are the "variables" (also known as "data items"). In a CSV these would be the column headings. Do you recognize the data types (numbers, strings, images, etc.)? * Is there missing, incorrect, or otherwise problematic data?
    • How and why was this data collected?
    • For whom is this data accurate or useful? What is this data unrepresentative of? (Who is missing and left out of the data?)
    • Knowing what you know now about machine learning, what will a model trained on this data help you do? Are there are alternative (non-machine learning) methods you could use instead?
  3. Link to the dataset and your post on Assignment 4a Wiki.

Session B: Training the Model

Objectives:

  • Learn steps to construct a vanilla neural network and train a classification model with ml5.js.
  • Understand Neural Network architecture
    • What is a Perceptron?
    • What is a multi-layered perceptron?
    • Activation Functions
  • Understand the terminology of the training process
    • Training
    • Learning Rate
    • Epochs
    • Batch size
    • Loss
  • Understand the difference between training and inference

Perceptron and Artifical Neural Networks

Color Classifier

ml5.js examples

Assignment 4b Due Sunday September 29 at 6pm

  1. Read 10.1 - 10.4 of The Nature of Code chapter 10.

  2. Train a model using ml5.NeuralNetwork() from a CSV or JSON dataset. Following are suggested exercises for expanding the Color Classifer or Titanic Survival Dataset examples. Feel free to pick one of the ideas below or invent your own.

    • Follow the Data Wrangling tutorial for the Titanic dataset and clean the data yourself. Try different ways for changing numbers to string labels and filling in missing data. Upload your own CSV, does the example still work? What changes?
    • Go back to the original titanic dataset and add another column from the raw data to the cleaned version. Upload a new CSV and retrain the model - what changes?
    • Try adjusting the "hyperparameters" of either example. Document what changes during the training and prediction stages. Below are some options can try.
    • If you are feeling ambitious, you can also try to train a model with your own dataset.
let nnOptions = {
  activationHidden: 'sigmoid', // 'relu', 'tanh'
  learningRate: 0.25, // any number!
  hiddenUnits: 16, // any number!
  modelLoss:  'categoricalCrossentropy', // 'meanSquaredError'
}

let trainingOptions = {
  epochs: 32,
  batchSize: 64
}

  1. Whatever exercise you choose, document the process of working with the data and the ml5.js library in a blog post. It's absolutely ok if nothing works, write a post that explains what you tried, document the errors, and theorize as to what might have gone wrong -- is the bug in your code, the data, the ml5 library, the web editor? For any terms that are unfamiliar, search for information online and write about what you tihnk they mean in your post.

  2. In addition to documenting your exercise, reflect on your experience using a new feature of an open source library. One of the big challenges of designing a higher level library for machine learning is finding the right balance between hiding the lower level details and allowing for flexibility / customization. Thinking about the difference between programming all the elements of the model (e.g. the Perceptron example and 3blue1brown's Neural Network video) and working with the ml5.neuralNetwork() function, how did this balance sit with you? Is the code easy for you to follow? Do you understand what the library is doing? What are some roadblocks you anticipate for using the library with your own data? Do you have any ideas for improvements or modifications to the ml5.js library?