The goal is to implement a Perceptron in JS that classifies MNIST handwritten digits.
features.csv: Each row is a grayscale image the digit with 8x8 pixels.
labels.csv: Contains the labels for the rows in features.csv, from 0 to 9, corresponding to the digit.
Run npm install to install dependencies.
Open the file exercise.js and fill the code snippets indicated by comments, starting from STEP 1.
After that, you can test the code by running node exercise.js and observing the output in the console.
Pro Tip: Pay attention to the Delta parameter in the console messages. It should decrease with each iteration, indicating that the learning is converging.
If you get this right, you should see something like that printed to the console: Score on test data: 99.27%.
The gradient calculation isn't returning a valid number.
It is possible that the Delta will increase in some iterations, by it should not happen on most of them. If you notice that Delta is increasing most of the time, your gradient calculation might be wrong.
It's alright buddy. You can peek the solution in the answer.js file.