# Neural Networks in Virtual Reality

> My master’s thesis: a web app that configures a neural network and a metaverse world that reads those parameters and builds a functional 3D model of the network – a “3D brain”.

- **Role:** Master’s Thesis · CTU FIT
- **Year:** 2025
- **Tech:** TypeScript, React, Three.js, Neural Networks, Resonite, WebGL
- **Repository:** https://github.com/ondraulehla/neural-network-visualizer
- **Live:** https://ulehla.dev/neural-network
- **Canonical page:** https://ulehla.dev/projects/neural-vr

**Problem.** Neural networks are taught as flat diagrams – hard to feel the architecture, the weights, or how a signal actually propagates through the layers.

**Solution.** A web application that encodes a network’s architecture and parameters, paired with an immersive 3D/VR environment that decodes them and generates a living 3D model used to drive objects in the scene.

**Outcome.** An end-to-end system connecting web technology, neural networks and virtual reality – and an interactive version you can run right here on this site.

---

## Context

This is my master’s thesis at the Czech Technical University, Faculty of
Information Technology (2025), supervised by doc. Ing. Mgr. Petr Klán, CSc.
The brief: build a system that models neural networks in virtual reality – a
web app encodes the network’s parameters, and a visual program in a 3D
metaverse reads that data and constructs a working 3D model of the network
(a “3D brain”) that can control objects in the virtual world.

## Problem

Neural networks are almost always presented as static, flat diagrams. That
representation hides the things that actually build intuition: the shape of the
architecture, the sign and magnitude of individual weights, and how an input
signal propagates and transforms layer by layer. It’s especially unintuitive
for newcomers.

## Solution

The thesis has two halves that talk to each other:

- **A web application** – the interface for defining a network: its layers,
  neurons, activation functions and weights. It runs the forward pass and
  backpropagation, generates datasets, and _encodes_ the whole network into a
  compact description.
- **A virtual 3D world** built on the [Resonite](https://resonite.com)
  metaverse platform using visual programming. It reads and decodes the web
  data and automatically **generates a 3D model** of the network in an
  immersive space, where the model becomes a control element for scene objects.

Together they demonstrate automated generation of a neural-network model from
parameters set in the browser – a bridge between web technologies, AI and VR.

## Outcome

- A complete pipeline: configure in the browser → encode → generate the network
  in VR → use it to drive the 3D scene.
- The forward pass, backprop and dataset generation are all real, running code.
- **Revived on this site:** the original visualization site went offline, so I
  rebuilt its interactive core right here – [open the Neural Network Lab](/neural-network)
  to train a network live and watch a 3D model of it light up with its own
  activations.

## What I learned

Making something _legible_ is its own engineering problem. Getting a network to
train is the easy half; turning weights, activations and signal flow into
something a person can look at and immediately understand – in 2D and in 3D –
took more design iteration than the math ever did.