I need to build a project for school in Java, and I challenged myself to building a Neural Network from scratch. This led to the development of a nn engine, which I plan on expanding here in the future
There's a new and completely reworked version of this project on https://github.com/c4vxl/jNN
The core engine introduces a new data type called Tensor, which can be thought of as a multidimensional matrix supporting various data types. It is capable of performing a wide range of mathematical operations, including element-wise computation, matrix multiplication, reshaping, transposing, and more.
Module Class:
The Module class serves as a base for any model or layer that requires serialization or saving.
Linear Class:
The Linear class implements linear transformation by performing weight multiplication and adding an optional bias which can booth be learned during training.
The goal of this project was to create a neural network, capable of identifying any kind of handwritten digit on an 28*28 pixel canvas.
- 784 inputs (for 28*28 pixels)
- 10 outputs (Numbers from 0-9)
- 2 hidden layers
- 12 neurons per hidden layer
This adds up to a total of 818 neurons in the entire model.
It has been trained on the train split of the MNIST Dataset (600000 images)
The user can draw a digit on a canvas, and the AI will attempt to predict the digit in real-time.
Image
This feature enables users to create a personalized dataset by drawing digits multiple times. The generated dataset can be used to fine-tune the model, helping it adapt to the user's handwriting style for improved accuracy.
Image
The benchmarking tool allows users to evaluate the model's performance on any dataset of their choice.
Image
Here, users can train their own model using custom parameters. The following training parameters can be configured:
- Shuffle Batches:
Randomizes the order of training samples to improve generalization. - Number of Epochs:
Defines how many times the entire dataset is used for training. - Logging Rate:
Determines how often the training progress is logged. - Validation Rate:
Specifies how often to validate the model's performance on the validation split. - Learning Rate:
Controls how much to adjust the model’s weights during each update. - Weight Decay:
Regularizes the model by penalizing large weights, helping to prevent overfitting. - Loss Function:
Switch between meanSquareLoss or crossEntropyLoss
This feature lets users load and explore any dataset. Each image is displayed alongside its corresponding label.
Image
Users can directly download datasets from Hugging Face through this feature. It provides options to browse available datasets or search for a specific one.
Image
It's mine, don't copy!
btw Kotlin > Java