НЕЙРО-СТАНЦИЯ «МНИСТ-1»
Show HN: Living Neural Network on the Webpage
Last checked:
What is НЕЙРО-СТАНЦИЯ «МНИСТ-1»?
An interactive browser-based neural network trainer that implements a real 784→30→10 network for MNIST handwritten digit recognition. Users can train the network on actual MNIST data (50,000 samples) and test it by drawing digits. The tool provides visualization of network architecture, training progress, and learned neurons with configurable hyperparameters.
НЕЙРО-СТАНЦИЯ «МНИСТ-1» pricing
Pricing model: Freemium
НЕЙРО-СТАНЦИЯ «МНИСТ-1» pros
- Runs entirely in the browser with no installation required
- Real implementation of forward/backward propagation and SGD training on actual MNIST data
- Interactive visualization of network architecture, training progress, and what individual neurons learned
- Can manually test with hand-drawn digits for real-time recognition feedback
НЕЙРО-СТАНЦИЯ «МНИСТ-1» cons
- Limited to fixed network architecture (784→30→10) with only hidden layer size customizable
- No model export or save functionality visible
- MNIST-only scope—no support for other datasets or image types
- Minimal documentation or beginner tutorial; Russian language elements may limit accessibility
Frequently asked questions about НЕЙРО-СТАНЦИЯ «МНИСТ-1»
What dataset does it use?
MNIST handwritten digits (50,000 training samples, 10,000 test samples from NIST, USA, 1998). Each image is 28×28 pixels.
Can I draw my own digits to test?
Yes. Draw a digit with the pen tool and the network recognizes it in real-time. Before training, the network guesses randomly.
What is the network architecture?
A 3-layer network: 784 input neurons (28×28 pixels), 30 hidden neurons (configurable), and 10 output neurons (digits 0–9).
What is the training algorithm?
Stochastic Gradient Descent (SGD) with configurable learning rate (η=1.0), batch size (20), and number of epochs (15).