The purpose of this exploration was to see how accurate a map could be created by AI. The end result is a fully detailed city map that can be used for many different purposes.
To accomplish this, I used two pre-trained models:
Swiftkey’s QuickDraw Model – A dataset of 50 million human-drawn images grouped into 345 categories. This model was trained on Google’s Cloud TPU (Tensor Processing Unit). This model allows me to draw and classify simple objects such as “chair,” “teddy bear,” or “microwave.”
Image Colorization – A model that takes a black and white image as input and outputs a colored image. This model was trained on Google’s Cloud TPU (Tensor Processing Unit).
I then wrote code to combine these two models and create the map.
The map is based on a procedural generation algorithm that uses machine learning to create a city, and then renders it using the Unreal Engine. It was trained using OpenAI’s text generator, GPT-2. The map is completely non-functional, and was created as a demonstration of how machine learning is being used to generate content.
OpenAI recently released a new version of their AI-powered text generator. I have been experimenting with it to see what kind of content it can produce, and also how well it can capture the style of different authors.
So far, I have tried it out with a range of classic novels: from Charles Dickens and Edgar Allan Poe, through to Rudyard Kipling and Jane Austen. The results were interesting.
The output can be quite chaotic and hard to follow, but the AI seems to pick up on the style of each author really well. At times, it sounds almost human!
As a next step in this experiment, I wanted to see if the AI could generate something non-textual, such as an image.
I came across a recent paper from DeepMind which described a way to “generate images from captions”. In other words, it used AI to match sentences with images. To try this out with OpenAI’s text generator, I needed a way to create an image from any sentence or phrase – something that isn’t an easy task! So instead I decided to take a slightly different approach: let’s use AI to write the caption for an image we have already created! This would be more like taking an existing painting and trying
This week we’re excited to announce the Codex, an open source dataset of 5 million human-labeled 3D training examples that’s orders of magnitude larger than any existing public dataset. The Codex is designed to help researchers train large-scale language models such as GPT-3 [1] to understand the visual world.
The Codex was generated by combining publicly available 3D assets with a new deep learning technique for rendering novel views of 3D scenes. We created more 27 million novel renderings from 1.5 million assets and asked humans to label them with captions, object types, and object attributes in order to capture the rich diversity of real-world scenes. We then filtered the results to create a final dataset of 5 million high quality visual observations, each paired with a natural language description.
We believe this is the first time anyone has used deep learning to generate realistic views at this scale, and that it’s possible because recent advances in machine learning have made it possible to render 3D objects with photo-realistic detail using neural networks [2]. While our work focuses on applying these techniques to AI research, we think they could also make 3D graphics more accessible in their own right, particularly for creative applications like animation where the cost of creating original art can
The OpenAI Codex is a project of the artificial intelligence research company OpenAI. The project aims to create a database of images, text, and other data that its AI systems can learn from. It will be available to the general public for free.
The Codex is the first of its kind, and it has the potential to be one of the most important tools in AI research. OpenAI is offering the Codex as a public service to encourage researchers around the world to use it for their own projects. The hope is that by sharing this data, the quality of AI applications will increase across the board.
There are two ways to access the Codex: through any web browser or through a mobile app (iOS or Android). The web version includes an image search tool that allows you to find images related to your query by drawing them on screen. You can also import images from your computer into the Codex for further exploration.
The mobile version of the Codex gives you access to all of its features on your phone or tablet. You can use it as a reference tool while traveling, at work, or just relaxing at home. You don’t have to be connected to Wi-Fi or cellular data plan in order use it!
OpenAI has published several papers about the Codex and
OpenAI is a non-profit artificial intelligence research company. The company’s mission is to ensure that artificial general intelligence benefits all of humanity. Our work combines the best techniques from machine learning and systems design to build flexible AI systems. Our focus areas include:
Artificial Intelligence Safety
Efficient and robust AI systems
Environment modeling and planning
Learning algorithms, tools, and infrastructure for large-scale distributed training, including reinforcement learning
Visual intelligence (e.g., generative models, perception)