3.2 Asynchronous

Rule-based thinking

Building models through coding (programming rules into the machine)

Data-driven thinking

Building models through training (rules are learned by the machine)

The reading “Teaching machine learning in elementary school” uses the instructional paradigms of:

• Learning ML by Design

• Learning ML by Teaching

• Their research is grounded in the learning theory of Constructionism

What is Machine Learning?

An AI technology that helps computers learn from data and make decisions without specific instructions for every task. Instead of programming a computer with every detail of what a cat looks like, we can show it thousands of images of cats and other animals. The computer then learns to identify cats based on patterns in those images.

ex. Google searches, Youtube recommendations, video games, self-driving cars

Where does AI fit in the big picture?

What are the similarities between human learning and AI Machine Learning?

Humans:

• Learn by seeing patterns in what we experience.

• We get better at things by practicing and learning from mistakes

• Once we learn something, we can apply it to new situations.

Machines:

• Learn by find patterns in large amounts of data

• Algorithms also improve over time learning from successes and failures.

• Machine learning tries to apply what it has learnt to new situations.

What are the differences between human learning and AI Machine Learning?

Humans:

• We learn in complex ways.

• We can learn new things quickly with only a few examples.

• We are very flexible learners.

Machines:

• ML relies on math and statistics.

• Machines usually need to see thousands or even millions of examples.

• ML is usually designed to be good at one specific task.

Artificial Neural Nets

There are two types:

• Simple Neural Networks: has an input layer, a hidden layer (where calculations occur), and an output layer, suited for basic tasks like pattern recognition.

• Deep Learning Networks: Deep learning neural networks, however, have multiple hidden layers, enabling them to tackle more complex problems, such as image recognition and language translation.

Machine Learning in Elementary School: Training

Historical Data → Machine Learning Algorithms → Predictive Model

Machine Learning in Elementary School: Validating

New Data → Predictive Model → Predictions

What does the term “black box” mean?

Describes ML models that, while effective at making predictions, are complex and hard to understand internally. The analogy to a "magic box" explains how the internal decision-making of the ML model's decision process remains opaque even though we train these models with specific data (like images of cats and dogs).

To help students comprehend the black box concept, researchers in the class reading (Shamir and Levin) had students create a simple Artificial Neural Network (ANN) with only one neuron, using Scratch. This single-neuron model illustrates basic neural network principles, making the complex inner workings of more layered ANNs easier to grasp.

The perdictive model is often referred to as _______.

A black box

What does the analogy to a “magic box explain?

The analogy to a "magic box" explains how the internal decision-making of the ML model's decision process remains opaque even though we train these models with specific data (like images of cats and dogs).

How did Shamir and Levin do to help students comprehend the black box concept?

(Shamir and Levin) had students create a simple Artificial Neural Network (ANN) with only one neuron, using Scratch. This single-neuron model illustrates basic neural network principles, making the complex inner workings of more layered ANNs easier to grasp.

Truth Tables

AND and OR truth tables, comparable to search queries (the text a user enters into a search engine to find information).

ex. “1986 AND Stephen King” versus “1986 OR Stephen King.”

This helped students understand how logic can be structured in the single neuron network to predict outputs based on different inputs.

How was the Structure of a Single Neuron explained?

To understand the structure of Single-Neuron ANN, a simple neuron model consisting of two inputs (represented by two switches) and an output that resembles a light bulb was developed and explained in the video.

What is the key takeaway from the Single Neuron Network Project?

Neurons modify their outputs using weights—numerical values that influence the neuron's response.

The key takeaway from this section is the concept of network weights, which represent the predictive model constructed through the machine learning process rather than the specific mathematical details of how the ANN achieves this goal.

Why would building an ANN predictive model be unclear sometimes?

Many people use AI systems like ChatGPT without understanding how they were developed or how the learning process works.

What is the purpose of the Single Neuron toolkit?

To enable students to write the code for a neuron.

This hands-on approach enhances their computational thinking skills. It provides a clearer understanding of how a neuron learns, offering insights into what happens inside the “black box” of machine learning. This method can serve as an effective educational tool for middle school students, helping them gain foundational knowledge in machine learning by actively engaging with a transparent, rule-driven model.

Use Phase (Module 1): Intro to ML

Engaged in simple activities to learn the basics of ML and how it differs from rule-based programming. 

Modify Phase (Module 2): Practicing the ML process

Completed Code.org AI for Oceans learning activity. 

Create Phase (Modules 3 and 4)

Module 3: Constructing data-driven ML system

Completed Machine Learning for Kids activities.

Module 4: Constructing a rule-driven ML system

Constructed Single Neuron ML system in Scratch

Traditional CT Practice: Decomposition

ML CT Practice:

Data split: data is decomposed by splitting the dataset into a training dataset and a validation dataset, 

Feature Selection: identify the features in a complex data entity such as an image, it is useful to initially break down the data item into subparts.

Traditional CT Practice: Pattern Recognition

ML CT Practice:

Category selection: classifying the training dataset into the problem categories the ML will be trained to recognize. This requires recognizing a pattern. 

Data filtering: use data that are relevant to the classified categories, the patterns relevant to the anticipated features need to be identified.

Predict: The need to predict the category of a data entity in the validation dataset based on features.

Traditional CT Practice: Abstraction

ML CT Practice:

Feature selection: properly classify the data entities, only those input dimensions that contain the relevant information for solving the problem must be selected.

Traditional CT Practice: Algorithm Creation

ML CT Practice:

Algorithm creation: The machine model is based on a neural network algorithm which needs to be either programmed or given.

Learning ML by Design

focused on algorithm construction, improving students’ understanding of computational perspectives, especially when building models like a single neuron network.

Learning ML by Teaching

emphasized data-driven practices, enhancing skills in data manipulation and category creation, such as building a chatbot.