Machine Learning Classroom

We are going to build cool ml projects, that is the main goal.

for the next two months

Only Sat–Sun. 2 days per week, 8 weeks total, okay.

Weekend	Core Goal	One-Line Task	Essential Resource
1	Python & NumPy	“Write a 2-layer NN in pure NumPy.”	Hands-On ML, Ch. 2–9
2	Linear → Logistic	“From scratch logistic regression on Iris.”	The Hundred-Page ML Book, §1–5
3	Loss Landscapes	“Add L2, dropout, early-stopping.”	Deep Learning, Ch. 7
4	Convolutions	“Pure NumPy CNN on CIFAR-10.”	CNN from Scratch Notebook
5	Autograd Engine	“Build micro autograd in <200 lines.”	MicroGrad Walkthrough
6	Optimizers	“Implement Adam & RMSprop.”	Optimization for DL course notes
7	Generative	“Train a VAE on CelebA 64×64.”	VAE from Scratch
8	Deploy & Invent	“Wrap your best model in a REST API.”	FastAPI ML Serving Guide

pip install numpy matplotlib tqdm requests fastapi uvicorn

After 8 weekends: GitHub repo with

Topics	Resources
Python Refresher	• Python for Everybody (free course)
Linear Algebra Crash	• 3Blue1Brown Essence of LA (YouTube)
Build a Perceptron	• Perceptron from Scratch Notebook
NumPy & Matplotlib	• Official NumPy Tutorial
Feed-Forward Neural Net	• Neural Networks from Scratch Book (PDF)
Train on MNIST	• MNIST in NumPy
Backprop Deep-Dive	• Calculus on Backprop
Implement SGD & Adam	• Optimization Algorithms Code
Regularization Tricks	• Dropout Paper
Build ConvNet	• ConvNet from Scratch
Experiment on CIFAR-10	• CIFAR-10 Dataset
Invent Something Small	• Ideas List

Artifact	Level	What It Means
8 clean notebooks	Beginner → Intermediate	Each notebook trains a model from scratch in pure NumPy: 2-layer NN, logistic reg, CNN, VAE, etc.
Custom autograd engine	Intermediate	~200 lines of code that can compute gradients for any feed-forward graph you design.
Working CNN on CIFAR-10	Intermediate	70 %+ accuracy without PyTorch/TensorFlow; proves you understand convolutions, backprop, and optimization.
VAE that samples faces	Intermediate	Generates 64×64 celebrity faces; you coded the re-parameterization trick yourself.
Personal GitHub portfolio	Showcase	One repo, clear timeline, reproducible notebooks → instant credibility for any future collaborator.
Mental model	Advanced	You can open any new paper and implement it in raw NumPy because you’ve already done the moving parts.