Interactive Graphical Neural Networks

This web application is a graphical interface for creating neural networks at the level of individual nodes and edges. Users can create network architectures, train their network, and observe how the model learns.

Go to quickstart for instructions to run a quick demo.
Check out usage for more detailed instructions on how to use the app.

Press Set under Set Layers.
Press Train under Train Model.
Observe the model's performance and new weights on the network!

Creating the Neural Network Architecture

Graph Editor: The graph editor contains all of the nodes and edges of the network. Nodes can be moved around with the mouse, and edges can be created by connecting the black dots on the nodes. Nodes and edges can be deleted by selecting and pressing delete.
Hidden Layer Nodes: Add nodes by dragging and dropping a node from the Drag and Drop section.
Input Node: The graph contains one input node, which takes the x value when running or training the model.
Bias Node: The graph contains one bias node, which takes the value of 1.
Output Node: The graph contains one output node, which is the result of running inference on the network for a given input.
Set Layers: create a network from text by writing the architecture in the layer sizes separated by x's and pressing Set.

Training the Neural Network

Hyperparameters: These parameters can be adjusted to change the behavior of the model.

Learning Rate: Choose the learning rate used to train the neural network.
Activation Functions: Choose the activation functions for the nodes in the network.
Epochs: Choose the number of training iterations to perform.
Batch Size: Choose the amount of data points to use in each training iteration.

Training the Model: The model can be trained to learn a function from a generated dataset.

Function: Choose the function that the network will learn.
Domain: Choose the domain of the input.
Variance: Choose the amount of noise added to the data.

Running the Model: Run inference and backpropagation on a given test point.

Test Point: Choose the input to run inference and backpropagation on.
Inference: Predict the output of the network for the given x value.
Backprop: Adjust the weights of the network to minimize the error between the predicted output and the actual output.

View Results: Visualize the model's performance with graphs.

Loss: Shows the error of the model over each epoch.
Results: Shows the ground truth function (blue), the generated dataset (purple), and the model's predictions (red).