There are many awesome examples out there where you can get a very direct feeling for what Machine Learning is. I'll collect a couple of them here.
Image Input Data
Math Symbol Recognition
The write-math.com web service allows you to recognize mathematical symbols automatically. It is described in my bachelors thesis. The HWRT data is available, also in its rendered form as HASYv2 dataset.
![write-math.com](../images/2017/12/write-math.png)
Drawings
Quickdraw lets you recognize drawings. They built a database of 50 million drawings.
![Quickdraw](../images/2017/12/quickdraw.png)
Using this kind of data, you can create an application which recognizes what was drawn and improves the drawing. autodraw.com does so:
Deep Dream
Deep Dream is a technique which needs a few more words to explain. You can read my paper Creativity in Machine Learning if you are interested in a very high level overview. If you just want to play with it, have a look at deepdreamgenerator.com.
![Deep Dream of a Moon jelly](../images/2017/12/Aurelia-aurita-3-0099.jpg)
Style Transfer
You have two images: A style image (e.g. by a painter) and a source image. You want the source image to be in the style of the style image:
Source image:
![Original image of a highland cattle](../images/2017/12/Highland-cattle-1.jpg)
Style image:
![Style image](../images/2017/12/starry-night.jpg)
Output image:
![Applied style transfer](../images/2017/12/Scottish-highland-cattle-1-style.jpg)
https://deepart.io seems to be a web service for this kind of machine learning. I didn't try it, though.
Super-Resolution
You have a small image and want the same image, but with higher resultion? Search for "super resolution":
Text Input Data
Translation
DeepL.com is much better than https://translate.google.com, but also more restricted. As always, you can find some... interesting... translations:
![DeepL Fail](../images/2017/12/deepl.png)
Sentiment Analysis
![Stanford NLP: Sentiment analysis](../images/2017/12/stanford-nlp.png)
Text To Speech (TTS)
Lyrebird.ai has the most impressive TTS system I have seen so far (although Googles Tacotron 2 audio samples are impressive as well).
Other ML
Clustering
Besides my small k-means clustering example, there is Tensorflow Projector
2D Embeddings
Mapping datapoints in 2D makes it easier to find what you are looking for. Have a look at fontmap:
![Fontmap: Organizing fonts in 2D map](../images/2017/12/fontmap.png)
Classifiers
Karpathy made a couple of interactive examples which show the decision boundaries of classifiers:
![Tensorflow Playground: See how neural networks learn](../images/2017/12/playground-tensorflow.png)
Honorable Mentions
- howhot.io was a service which lets you upload an image with a face and rate how hot the person is. It is no longer available.
- Microsoft Chatbot Tay went racist (source)
- A visual introduction to machine learning
- Martin Thoma, 2016: Creativity in Machine Learning
- Alex Rogozhnikov, 2016: Gradient Boosting explained
- Alex Rogozhnikov, 2016: Hamiltonian Monte Carlo explained
- GridWorld: Dynamic Programming Demo
- Wattenberg et al, 2016: How to Use t-SNE Effectively