Perceptron

This is an example of how a single perceptron works, it was one of the first codes I made to study neural networks, and the code was made using TypeScript.

The project is already compiled to JavaScript, but if you feel like changing something in TypeScript, of course, you will need to have TypeScript compiler installed, and for that, you will need to have a packet manager such as npm or yarn installed.

Making TypeScript code changes

1. Create a new folder;

2. Inside the new folder, clone the repository;

git clone https://github.com/vinicius-goncalves/perceptron-ts.git

3. Then install the latest TypeScript version:

npm install -g typescript

4. Verify if TypeScript was installed correctly:

   tsc -v

   or

   tsc --version

5. Open with your code editor and then make the changes. Then, open a terminal inside the root project folder and use the following command to compile from TypeScript to JavaScript:

tsc

5. Open the index.html and then visualize the console.

This is an example of how a perceptron works, so you need to know the basics of how a neural network works to understand the steps done here.

The project has a "Perceptron.ts" class, you will need to instance this in the app.ts like this:

const perceptron: Perceptron = new Perceptron(2);

Where "2" it's the number of inputs. In this example, I'm using AND logical port - so, we just need two inputs.

After instantiating the class, we need to train the perceptron, and the Perceptron prototype has a train() method.

perceptron.train(inputs, label, lr?);

To understand this, we need to know: Exists a model for training "TrainerSet", that it must have the inputs, where this is an array of numbers, and a label, where is the expected result - this must be a number. Example:

Inside the TrainerSets.ts:

const ANDTrainer: TrainerSet[] = [
    { inputs: [1, 1], label: 1 },
    { inputs: [0, 1], label: -1 },
    { inputs: [1, 0], label: -1 },
    { inputs: [0, 0], label: -1 }
]

Now, we can loop it for some thousand epochs getting a random set inside the ANDTrainer array to train using the train() method. Something like this could be thought:

for(let i: number = 0; i < 20_000; i++) {

    const { inputs, label }: { inputs: number[], label: number } =
        getArrayRandomValue(AND);

    perceptron.train(inputs, label);
}

After that, we can predict the results, to know if the perceptron is getting errors or not - and we can use the predict() method of the Perceptron prototype to do it.

//AND SET, so, expected row matrix: [1, -1, -1, -1]

const predictions: number[] = [
    perceptron.predict([1, 1]), //1
    perceptron.predict([0, 1]), //-1
    perceptron.predict([1, 0]), //-1
    perceptron.predict([0, 0]), //-1
];

If we send a log to the console, we can see the predictions:

console.log(predictions); //[1, -1, -1, -1]