A comprehensive guide to client-side, zero-install machine learning in ForgeCurves. Train classification and regression models, optimize deep neural networks, evaluate with diagnostic ROC/confusion plots, and export to production ONNX and standalone scripts.
ForgeCurves provides native implementations of modern statistical learning algorithms operating on column-major matrices and tabular datasets with zero cloud roundtrips:
Dual coordinate descent optimizer supporting linear and non-linear Radial Basis Function (RBF) kernel Gram matrices with automatic hyperparameter scaling.
Bootstrap aggregation of decision trees with randomized feature splits, out-of-bag error estimation, and feature importance rankings.
Lasso (L1), Ridge (L2), and Elastic Net regression solvers featuring full cyclical coordinate descent regularization path estimation.
Iteratively Reweighted Least Squares (IRLS) solver for binomial and multinomial logistic classification with analytical deviance metrics.
The built-in nnet subsystem allows designing, configuring, simulating, and training feedforward and recurrent neural networks directly in the browser runtime:
feedforwardnet, fitnet, patternnet, cascadeforwardnet, and recurrent Elman networks (newelm) with arbitrary layer dimensions.
Hyperbolic tangent (tansig), Sigmoid (logsig), Linear (purelin), Softmax (softmax), and ReLU (poslin) with exact analytical derivatives.
Gradient Descent with Momentum (traingdm), Resilient Backprop (trainrp), Scaled Conjugate Gradient (trainscg), and Levenberg-Marquardt (trainlm).
Automatic range normalization (mapminmax), standard normalization (mapstd), and one-hot index vector conversions (ind2vec / vec2ind).
Evaluate model quality using high-performance vector diagnostics rendered dynamically via ForgeCurves' SVG plotting engine:
Interactive multi-class confusion grid showing true positives, false alarms, and class-wise precision and recall metrics.
Receiver Operating Characteristic curve generation with trapezoidal numerical integration of the Area Under Curve (AUC).
Logarithmic performance decay plots tracking training, validation, and test mean squared error (MSE) over gradient epochs.
Benchmark suite results measured on ForgeCurves client-side runtime across core machine learning workloads:
| Workload Description | Dataset Scope | Mean Time (ms) | Throughput / Rate | Status |
|---|---|---|---|---|
| Linear SVM Fit (Dual Coordinate Descent) | 1,000 samples × 10 feats (50 iters) | 290.92 ms | 3,437 samples/s | Accelerated |
| Non-Linear RBF Kernel Matrix | 1,000 × 1,000 Gram matrix (γ=0.1) | 1,213.68 ms | 0.82 M-pairs/s | Accelerated |
| Random Forest Fitting (TreeBagger) | 10 decision trees, 1,000 × 10 | 132.40 ms | 75.5 trees/s | Accelerated |
| Lasso Path Optimization | 20 λ regularization steps | 1,351.13 ms | 14.8 steps/s | Optimal |
| MLP Forward Batch Inference | 10,000 samples (20 → 50 → 10) | 766.87 ms | 13,040 samples/s | WASM SIMD |
| Neural Network Backprop (traingdm) | 1,000 × 10 samples, 50 epochs | 9,274.82 ms | 5.39 epochs/s | Converged |
| Levenberg-Marquardt Step (trainlm) | Hessian step (JTJ + λI) | 174.35 ms | 5.73 steps/s | Linear Solve |
| ROC Curve & Trapezoidal AUC | 10,000 continuous predictions | 95.84 ms | 104,344 preds/s | Instant |
| Confusion Matrix Computation | 10,000 predictions, 5 classes | 252.64 ms | 39,582 preds/s | Instant |
Ensure zero vendor lock-in. Models trained inside ForgeCurves can be exported instantly to production environments without runtime dependencies:
Compiles trained models into standard Open Neural Network Exchange (ONNX) binary schemas compatible with ONNX Runtime, PyTorch, and TensorRT (benchmark: 0.80 ms / 1,243 graphs/s).
Synthesizes standalone GNU Octave / MATLAB® scripts embedding trained weight tensors directly, capable of executing independently in any headless environment (benchmark: 0.67 ms).
Inspect layer topologies, sparsity ratios, bias distributions, and training telemetry in a graphical modal inspector.