This Review gives an overview of interesting stuff I stumbled over which is related to machine learning.
New Developments
Human-strength and Super-human strength programs
Super-human strength programs are programs, which surpass even the best human (in the long run) in a specified task. Human-strength programs behave similarly to an (untrained) human.
Although those are not new, seeing them as a list (source) was new to me. However, except for the games and lip reading, I doubt that we are there yet. Interesting, nonetheless:
- Games
- Lip reading: 2016, Lip Reading Sentences in the Wild (YouTube)
- Geolocation by photos, PlaNet - Photo Geolocation with Convolutional Neural Networks
- Speech transcription: 2016, Achieving Human Parity in Conversational Speech Recognition
- Translation: 2016, Zero-Shot Translation with Google’s Multilingual Neural Machine Translation System
- Driving: 2016, Waymo
Live Demos and Websites
- universe.openai.com: Related to the OpenAI gym.
- Project Malmo: Train RL agents in Minecraft
- VISIIR: VIsual Seek for Interactive Image Retrieval - classifying food
- Image-to-Image
Publications
- High-Resolution Image Inpainting using Multi-Scale Neural Patch Synthesis and Code
- The Game Imitation: Deep Supervised Convolutional Networks for Quick Video and YouTube playlist
- Deep Nets Don't Learn Via Memorization
Software
Interesting Questions
Miscellaneous
Trump QA idea
I was just watching this clip and wondered how well a question answering system would work which is trained on Trump speeches. Very often, when reporters / journalists / moderators ask Trump a question, he answers with "I am the [most / best / least] [positive / negative statement]. [Inconsistent answer follows]". The answers themselves would almost certainly be hilarious. Second, one could make an experiment and ask people if Trump actually answered a question like this.
Meetings
- London, 4. December 2016: Data Visualization Challenge
- Barcelona, 5. December 2016 - 10. December 2016: Neural Information Processing Systems (NIPS) (Link)
- Mannheim, 7. April 2017: DataFest Germany