What You Actually Need When You're Staring at a Broken Model
I spent three years building ML systems before someone finally handed me a single-page reference that didn't make me want to throw my laptop out a window. Most cheat sheets are bloated. They pack every formula, every architecture variant, every hyperparameter range into one document until it becomes useless clutter. The Machine Learning Cheat Sheet Minimalist strips all that away and leaves only what you actually reach for under pressure. The idea is simple enough that it almost feels insulting at first. One side of A4. Front and back. No decorative headers, no "top 10 algorithms" listicles, no motivational quotes. Just the decision trees, the formulas you forget the moment you need them, and the parameter ranges that actually work in production. I built mine after a client migration project where I spent 45 minutes frantically searching through three different documentation sites for the correct bias initialization formula for a custom LSTM variant. That was the last time I relied on scattered references. The core structure covers four blocks: loss functions and their gradients, optimization landscapes and learning rate behavior, regularization signatures, and model selection decision paths. That's it. Everything else is noise.
Loss functions come first because they're the thing that breaks most often. Cross-entropy with the class-balanced variant, binary vs categorical distinction marked in red, and the MSE decomposition showing why it fails on imbalanced classification. Beginners always conflate these. The sheet shows the gradient forms side by side so you can see at a glance which one will explode with large predicted probabilities and which one stays bounded. Optimization occupies the second quadrant. Adam, RMSprop, SGD with momentum — each gets one row showing the effective learning rate behavior, the default beta values, and the one line about when Adam's default settings cause generalization gaps that SGD recovers from. This is the counter-intuitive part most guides skip. Adam converges faster but frequently lands in sharp minima. SGD with momentum finds flatter minima that generalize better, even if the training curve looks slower. I learned this the hard way on a medical imaging task where the Adam-trained model hit 94% validation accuracy and then dropped to 71% on held-out distribution-shifted data. Switching to SGD with cosine annealing brought it back to 89%. The cheat sheet flags this pattern with a small sidebar note rather than a full explanation, because if you need the full explanation you probably aren't ready to be tuning optimizers manually anyway. Regularization gets treated with similar surgical precision. Dropout rates by layer type, L1 versus L2 behavior on weight sparsity, early stopping patience heuristics based on validation loss oscillation patterns, and the label smoothing formula with its temperature parameter. The edge case here is worth mentioning specifically. Label smoothing at too high a value — above 0.2 for most classification tasks — starts collapsing your calibration. I encountered this when deploying a fraud detection model where the output probabilities needed to map directly to risk scores for the operations team. The model's AUC was fine, but the Brier score was terrible because the softened labels destroyed the probability estimates. Lowering the smoothing parameter to 0.05 fixed it without touching the architecture.
The model selection decision tree is the longest section and the most useful. It maps input characteristics — sequential, spatial, tabular, graph-structured — to algorithm families with branching conditions. Does your data have temporal ordering? Go recurrent or transformer. Is it image grid data? Convolutional path. Tabular with mixed types? Gradient boosting or a shallow network with embeddings. The trap most people fall into is starting at the wrong branch. I've seen practitioners try to force a vision transformer on tabular data because it's trendy, then wonder why a simple XGBoost baseline outperforms it by six percentage points. The decision tree prevents this by forcing a gate check at each step. Hyperparameter ranges form the appendix. Learning rate brackets per optimizer, batch size guidelines tied to memory constraints, dropout rate banding per layer depth, and the weight decay to learning rate ratio that works across most Adam variants. These aren't sacred numbers. They're starting points that have survived enough deployment cycles to be worth remembering. Here's what most people don't understand about minimalist references like this. They only work if you already know enough to recognize when you're stuck. A beginner will look at this sheet and see gaps everywhere. That's the point. The gaps are where you learn. The sheet doesn't teach you machine learning. It teaches you what to check when your model is misbehaving and you've exhausted the obvious fixes. Debugging workflow is the actual skill being developed here.
Get the Full Details
I keep the current version in my editor as a PDF, open on a secondary monitor during any model training session. It typically cuts my diagnosis time from an hour of aimless Stack Overflow browsing down to ten minutes of targeted parameter adjustment. Some weeks go by where I don't even look at it because nothing is breaking. Then something breaks badly and the difference between knowing where to look and not knowing is measured in lost weekends. Download the Machine Learning Cheat Sheet Minimalist as a printable two-page PDF with the corrected gradient formulas and updated regularization tables from the linked repository. The latest version accounts for the 2025 shift toward adaptive regularization in transformer pretraining pipelines, which standard sheets still get wrong. Don't expect this to replace documentation. It replaces the panic moment when you're three hours into an experiment and can't remember whether your learning rate schedule should be warmup-first or decay-first. In those moments, the minimalist approach is the only thing that helps.
