AI Lab is an experimental learning tool that allows you to directly manipulate and visually understand machine learning, deep learning, and reinforcement learning in one place. Each function comes with a ready-to-run demo, and changing parameters reflects the results in real-time visualizations, allowing you to quickly grasp the concepts. Furthermore, the data entered into the experiments is not stored on the server.

[What can you do?]
- Understand concepts by sight: Visualize key concepts such as decision boundaries, loss curves, filter responses, and error heatmaps.
- Experiment by hands: Adjust parameters using sliders/dropdowns and immediately see results.
- Experience model efficiency: Intuitively compare accuracy, capacity, and error tradeoffs by applying pruning and quantization.
- Optimized for learning/classes/demo: Quickly apply to classes, study sessions, and in-house seminars with lightweight examples.

[Provided Tools (by Category)]
1) Machine Learning
- Linear Regression: Change weights, bias, and learning rate, visualize MSE/residuals.
- Logistic Regression: Adjust decision boundaries/probability contours, and L2 regularization strength.
- Decision Tree: Observe split criteria/maximum depth/overfitting effects.
- K-Means: Animate cluster changes according to K value/initialization/iteration.
- KNN: Change classification boundaries according to K value and distance measure.
- SVM: Visualize margin and support vectors by adjusting linear/margin/C parameters

2) Deep Learning
- XOR Learning Demo: Observe the learning process of a nonlinear problem using a multilayer perceptron
- Fitting Experiment: Compare underfitting and overfitting according to model capacity and regularization strength
- CNN: View convolution/pooling flow and confirm channel-specific responses
- CNN Filter Test: Check results by directly changing kernels such as edge and blur
- Mini LLM: Experience an ultra-light text model (mini demo for understanding input/output flow)

3) Reinforcement Learning
- Grid World: Convergence process of value and policy iteration, visualizing policy/value maps
- N-Slot Experiment (Multi-Armed Bandit): Comparison of exploration and exploitation strategies such as ε-greedy and UCB

4) Optimization/Model Compression
- Sparse Matrix Compression Simulator
Encode in various storage formats, including COO/CSR/CSC/RLE/Dictionary/Bitmap

· Numerical verification of compression size, restoration consistency, and compression ratio

JSON output support
- Pruning Simulator

· Default 6×12, density slider (minimum 50% to 100%)

· Visualize the zeroing process using size-based pruning (weight threshold)

· Impact assessment using sparsity (%) and error metrics
- Quantization Simulator

Default 6×12, quantize float (−1 to 1) weights to integers

Bitwidth 2–8 (default 4), modes: Uniform Symmetric, Uniform Asymmetric, Row-Dynamic, Log2, Binary, Ternary

· Simultaneous display of integerization matrix, restoration matrix (3 decimal places), and error heatmap

· Provides metrics such as MSE, mean error, PSNR (dB), and bitrate, and exports to JSON

[For Learning] Helpful Visualization Points
- Decision Boundary & Probability Distribution: Visualize the classifier's judgment on screen.
- Loss/Error Curves: Track changes due to learning rate/regularization adjustments.
- Filter Response Map: Intuitively understand how the CNN kernel responds to images.
- Efficiency Heatmap: Identify areas where errors increase during pruning/quantization at a glance.

[Recommended for]
- Students/Beginners: Those who want to quickly understand mathematical formulas, graphs, and then hands-on experience.
- Instructors/Mentors: Those who need demo-based lecture/seminar materials.
- Engineers/Researchers: Those who want to easily sketch ideas and validate concepts.

[Data/Privacy Information]
- The input data used in the experiments is not stored on the server.