BUILD AND EXPLORE
NEURAL NETWORKS

Artificial Intelligence, or AI, is changing the way we live, communicate, and learn. You've probably heard or even used AI in situations like chatting with ChatGPT, using virtual assistants like Siri or Alexa, or viewing personalized movie recommendations on Netflix. But what exactly is AI, and how does it work?

Artificial Intelligence involves creating computer programs that can recognize patterns, make predictions, and perceivably behave intelligently, somewhat similar to how humans think. Humans are naturally skilled at tasks like recognizing faces, understanding speech, or reading handwriting. Traditionally, programming a computer to do these tasks meant writing out every single instruction explicitly, step-by-step.

But let's take an example: Imagine teaching a computer to recognize handwritten numbers. Every person writes numbers differently, making it nearly impossible to manually program specific instructions to handle every style of handwriting.

With Machine Learning, a specific type of AI, we teach computers to learn on their own by providing lots of examples. Let's say we want our AI to recognize handwritten numbers from 0 to 9.

Traditionally, you'd have to write a complex set of instructions describing exactly how each digit (0-9) should look, specifying the curves, points, angles, and strokes involved. Given how differently each person writes numbers, this becomes nearly impossible to program explicitly.

Instead of programming every detail, we give our AI thousands of images of handwritten numbers, each labeled with the correct digit. Our AI examines these examples and learns subtle patterns by identifying common features like loops, curves, and straight lines, seemingly on its own.

Let's imagine you are the AI. Right now, you have no idea what any digits look like, you just know there are ten possibilities: 0, 1, 2, 3, 4, 5, 6, 7, 8, or 9. Let us begin training.

Round 1: Examine the image of a handwritten "0."

Since you've never seen a number before, you randomly guess one of the ten options, say, "1." Immediately, you're told the guess was wrong, and that the correct answer is "0." So, you start looking for a pattern. You notice the "0" has a loop with empty space in the middle.

Round 2:

Examine a different handwritten "0." This time, remembering the loop pattern you associated with "0," you confidently say "0." You're correct! You now reinforce your understanding that a loop likely indicates a "0."

Round 3:

Examine an image of a handwritten "9." You guess "0" again. You're told that it is actually a "9." Surprised, you adjust your methodology: if you see a loop, it could be "0" or "9," but a line attached to that loop makes it more likely a "9."

Now let's do that a thousand times, with a thousand different examples of digits. After seeing a hundred different "9"'s, you get pretty good at recognizing it. The more unique handwriting you see, the better you are at identifying numbers that are written messy or with a different font.

In reality, AI doesn't think exactly like humans do. Machines don't literally "see" or "think" about loops or lines in the casual way we do. Instead, their decision-making relies on mathematical patterns and numerical calculations. However, the fundamental idea remains simple: AI systems improve their predictions by analyzing many examples and learning from their mistakes and successes.

In machine learning, these small discoveries and refinements happen inside something called a neural network. Imagine a neural network as a network of paths (called nodes and connections) that represent your thought processes.

• Input Nodes: These are like your eyes, seeing the handwritten digits.
• Hidden Nodes: Think of these as your "thinking steps," detecting features like loops, curves, or straight lines.
• Output Nodes: These give your final answer—which digit(s) you think you've recognized.

Each node and connection represents a small rule or piece of evidence. For example, one hidden node might activate more if it detects a loop, another might activate if it sees a vertical line, and another might check if there's space within the loop. The combination of these activations leads the neural network to eventually prefer one digit more than the others.

Let's walk through using a real neural network. I will show you how to use the tool I made. Let's start by selecting a dataset. We can choose from the Boston Housing, California Housing, Diabetes Progression, and Iris classification datasets. For specific information, click the "Info" button next to the dropdown menu after choosing the dataset.

Next, let's discuss Test Size. Test size tells the neural network how much of this data we want to use for training (learning). This is key because it helps determine how well our model will perform on new, unseen data.

Now that we know our dataset and our testing/training separation, we can start building our input layer. Click on the input dropdown menu "Input Parameters." Here you can select inputs from the dataset that you wish to provide to your network for learning.

The next part is Hidden Layers. Hidden layers are the intermediate layers between our input layer and the output layer. They allow the network to capture complex relationships in the data by breaking down the inputs into multiple, simpler steps.

Finally, we can press "Train Model." Once finished, if we click on the "Loss Graph" header, we will see our loss epoch graph. Loss measures how far the neural network's predictions are from the actual values, and epochs are simply the number of times the model goes through the entire training dataset.

At the top of the sidebar, you will see the final Mean Absolute Error (MAE) percentage for regression datasets, or accuracy for classification datasets. By experimenting—changing inputs, adding or removing layers, or tweaking other settings—you discover how each choice impacts the neural network's ability to learn and predict accurately.

By experimenting, changing inputs, adding or removing layers, or tweaking other settings, you discover how each choice impacts the neural network's ability to learn and predict accurately.