Smart Digit Learner ML+ Pro

Draw digits, save/reload datasets, tune classic ML models, train a dense neural net, and test a tiny CNN that preserves 2D pixel structure.

Single-file ✔️ FNN ✔️ CNN ✔️ Dataset I/O ✔️

Draw Pad

Use mouse or touch. Thick centered strokes work best.

28×28 vectorizer

Brush 18px

Dataset & Results

Choose the active model, manage your dataset, then predict or evaluate.

Label (0–9)

Prediction Model

Total samples

Active model

Gaussian Naive Bayes

Feature size

784

Dataset state

Empty

Top prediction

—

Merge imported samples

🧠 Ready. Draw a digit, enter its label, then click Add Sample.

Samples per Digit

Feature Builder & Model Settings

Classic ML, logistic regression, and the feedforward network use the feature builder. The CNN uses raw 28×28 pixels directly to preserve spatial layout.

Feature Builder

Global features for Naive Bayes, KNN, Logistic Regression, and FNN.

Feature mode

Pool size

Neighbor radius

L2-normalize feature vectors

Gaussian Naive Bayes

Continuous Gaussian likelihoods with variance and prior smoothing.

Variance floor

Prior smoothing

KNN

Distance-based voting in feature space.

Neighbors (k)

Distance metric

Use weighted voting

Softmax Logistic Regression

Mini-batch gradient descent with decay and regularization.

Epochs

Learning rate

Decay

L2 regularization

Batch size

Feedforward Neural Network

Dense hidden layer with ReLU and softmax output. Uses selected feature vectors.

Hidden units

Epochs

Learning rate

L2 regularization

Tiny CNN

3×3 convolution filters, ReLU, global average pooling, and softmax classifier.

Conv filters

Epochs

Learning rate

L2 regularization

Average Digit Heatmaps

Each tile shows the average learned appearance for that digit.

PCA 2D Projection

Projects raw samples into 2 dimensions so you can inspect clustering.

Confusion Matrix

Training-set evaluation using the selected model.

Run evaluation to see accuracy and the confusion matrix.