Modern Deep Learning Foundation Playground
Six focused, browser-based laboratories for the concepts that benefit most from direct experimentation. Every laboratory now has its own full page, stable workspace, and focused learning objective.
Six complete learning laboratories
Performance Evaluation Metrics
Inspect the confusion matrix, move the classification threshold, and connect precision, recall, F1, ROC, and AUC on one consistent dataset.
- Confusion matrix
- Threshold analysis
- ROC and AUC
Overfitting and Regularization
Compare training and validation behaviour while model capacity, regularization strength, and data conditions change.
- Bias and variance
- Regularization
- Generalization gap
How Does a CNN Work?
Explore a complete CNN pipeline with real benchmark samples, editable kernels, padding, stride, pooling, and layer-level activations.
- MNIST and CIFAR-10
- Activations and pooling
- Tensor shapes
Self-Attention and the Transformer Principle
Trace token embeddings through Query, Key, Value projections, scaled dot-product attention, multi-head mixing, and residual updates.
- Q, K, and V
- Multi-head attention
- Residual stream
Explainability: Understanding Model Decisions
Compare local feature attribution methods and inspect how evidence supports or opposes a model prediction.
- Feature attribution
- Local explanations
- Explanation stability
Transfer Learning and Fine-Tuning
Freeze and unfreeze a pretrained network block by block, then observe trainable parameters, adaptation, expected accuracy, and forgetting risk.
- Frozen layers
- Partial fine-tuning
- Catastrophic forgetting