AI theatre
About this app
Explore how a neural network recognises digits and learns through four linked 3D scenes. The AI tutor selects a route through neurons, recognition, training and backpropagation.
The idea
A neural network is built from many simple neurons. Each computes a weighted sum of its inputs, adds a bias and passes the result through an activation function. Stacked in layers, such neurons can recognise patterns such as handwritten digits. Learning means changing the weights so that an error measure goes down. Backpropagation efficiently computes how much each weight contributes to the error, and gradient descent moves the weights in the opposite direction.
What you see
The AI theatre links four existing 3D apps into a story in four acts: “Neuron” (a single artificial neuron), “Recognise” (a drawn digit flows through a trained network), “Training” (a network learns digits in the 3D lab) and “Error” (backpropagation with real numbers). A bar at the top switches between acts, and each act is the full, usable app.
With the AI tutor
The AI tutor can stage the acts: depending on your question it picks a path through the stages, switches acts and drives each app with speech bubbles, a pointer and camera moves. Without the tutor you walk through the acts yourself. The given order works well: first understand what a neuron computes, then see how a trained network decides, and only then how it learns.
Frequently asked questions
- How does a neural network recognise handwritten digits?
- The digit image is broken into pixel values that feed the first layer. Each layer computes weighted sums and activations and passes them on. The output layer produces a value for each digit 0 to 9; after a softmax function these can be read as probabilities.
- How does a neural network learn?
- The network makes predictions on training examples and compares them with the correct answers. A loss function summarises the error, and backpropagation computes its derivative with respect to every weight. An optimiser such as gradient descent then adjusts the weights in small steps so the loss decreases.
- What does backpropagation do?
- Backpropagation applies the chain rule backwards through the network. Starting from the error at the output, it computes layer by layer how much each weight contributed to that error. The optimiser then uses these gradients for the update.
- What can I do in this app?
- You use the act bar to switch between the four stages Neuron, Recognise, Training and Error and operate each one like the underlying app. With the AI tutor you can have the path through the acts explained to you.
Subject: Machine Learning
More from Machine Learning
- AdaBoost
- Anatomy of a prompt
- Data privacy traffic lights
- Decision Tree
- Ensemble: Bagging vs. Boosting
- Finding faces: masks versus learning from examples
- k-Means
- k-Nearest Neighbors (kNN)
A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps