6 projects to understand basic algorithms in Machine Learning, using Lodash and Tensorflow.js.

A quick introduction to K-Nearest Neighbors (KNN) algorithm using Lodash.

See 01-plinko folder

• implementing K-Nearest Neighbors (KNN) algorithm with Lodash methods.
• testing the algorithm and interpreting bad results.
• creating random test and training datasets.
• generalizing KNN, gauging accuracy and printing a report.
• investigating optimal k values.
• updating KNN for multiple features.
• applying normalization with MinMax.
• handling objective feature selection with KNN.

Another implementation of K-Nearest Neighbors (KNN) algorithm with Tensorflow.js.

See 02-house-pricings folder

• loading CSV data.
• running K-Nearest Neighbors (KNN) algorithm with Tensorflow.js.
• analyzing results and reporting error percentages.
• applying standardization.
• debugging calculations with node --inspect-brk and Chrome debugger.
• adding features to get a better accuracy.

A gradient descent algorithm in linear regression using Tensorflow.js.

See 03-gas-mileage-calculator folder

• loading CSV data and creating a LinearRegression class.
• creating a basic gradient descent implementation with arrays of data and for loops.
• calculating mean squared error (MSE) slopes, updating coefficients and interpreting results.
• understanding matrix multiplication.
• vectorizing the gradient descent algorithm with Tensorflow.js helpers.
• gauging model accuracy by implementing the coefficient of determination.
• applying standardization and massaging learning rates.
• refactoring for multivariate regression.
• tracking MSE history to update learning rate automatically.
• plotting MSE history with nodeplotlib.
• refactoring towards batch and stochastic gradient descent for better performances.
• making predictions with the model.

A gradient descent algorithm in logistic regression to handle natural binary classification.

See 04-smog-test folder

• updating linear regression for logistic regression by adding the sigmoid equation.
• implementing a test function to gauge classification accuracy.
• supporting variable decision boundaries.
• refactoring with cross entropy to update learning rate.
• plotting cost history with nodeplotlib.

A multinominal logistic regression to handle multi-value classification.

See 05-fuel-efficiency folder

• refactoring to multi-column weights.
• classifying continuous values in 3 ranges of fuel consumption: low, medium and high.
• training a multinominal model.
• handling conditional probability distribution with the softmax equation.
• implementing accuracy gauges with argMax() and calculating accurary.

Another multinominal logistic regression with a larger dataset.

See 06-handwriting-recognition folder

• flattening image data from MNIST handwritten digit database.
• encoding label values and implementing an accuracy gauge.
• debugging the calculation process with node --inspect-brk.
• dealing with zero variances.
• creating memory snapshots with Chrome debugger.
• releasing references to MNIST dataset.
• optimizing Tensorflow.js memory usage with tf.tidy().
• measuring footprint reduction.
• plotting cost history and improving model accuracy.

Based on Machine Learning with JavaScript by Stephen Grider (2021).