Get a hang of reinforcement learning for games

Check this out: Introduction to AI for Video Games (Reinforcement Learning) by Siraj Raval

And this: Lecture 10: Reinforcement Learning in CS188 Artificial Intelligence, Fall 2013 (University of California, Berkley)

Also this lecture on Deep Reinforcement Learning from Stanford CS231n

And this playlist: Introduction to Reinforcement Learning by Deep Mind

If you are interested in creating Games in Unity, this is also something you shall check out https://github.com/Unity-Technologies/ml-agents

And this: Playlist (Unity Machine Learning)

This Lecture from MIT was also interesting (and the DeepTraffic assignment was fun)

DeepSchool.io

There is a new kid on the block in terms of online courses on Deep Learning.

DeepSchool.io is a set of Jupyter notebooks that teach you the basics and different concepts you need in order to get started and being productive in Depp Learning. They are also videos supporting the notebook, although not for every notebook yet.

It differs from fast.ai in that the videos are shorter and the notebooks are mostly self-explanatory.

The goal of the project is to make Deep Learning accessible to everyone, make it practical, make learning open source and fun.

These are the topics covered:

  1. Lesson 0: Introduction to regression.
  2. Lesson 1: Penalising weights to fit better (scikit learn intro)

Mathematics (optional)

  1. Lesson 2: Gradient Descent. Using basic optimisation methods.
  2. Lesson 3: Tensorflow intro: zero layer hidden networks (i.e. normal regression).
  3. Lesson 4: Tensorflow hidden layer introduction.

Deep Learning

  1. Lesson 5: Using Keras to simplify multi-layer neural nets.
  2. Lesson 6: Embeddings to deal with categorical data. (Keras)
  3. Lesson 7: Word2Vec. Embeddings to visualise words. (Tensorflow)
  4. Lesson 8: Application – Bike Sharing predictions
  5. Lesson 9: Choosing Number of Layers and more
  6. Lesson 10: XGBoost – A quick detour from Deep Learning
  7. Lesson 11: Convolutional Neural Nets (MNIST dataset)
  8. Lesson 12: CNNs and BatchNormalisation (CIFAR10 dataset)
  9. Lesson 13: Transfer Learning (Dogs vs Cats dataset)

Advanced Topics

  1. Lesson 14: LSTMs – Sentiment analysis.
  2. Lesson 15: LSTMs – Shakespeare.
  3. Lesson 16: LSTMs – Trump Tweets.
  4. Lesson 17: Trump – Stacking and Stateful LSTMs.
  5. Lesson 18: Fake News Classifier

You can read more here.

AlphaGo Zero, beats previous world champion winner AlphaGo, 100-0

You heard of DeepMinds AlphaGo that beat worlds best Go player in the game everyone said computers would still need ten years to beat humans in.

That version trained on millions of expert human gameplays and then trained on itself through reinforcement learning.

This version skips all human gameplay and learns by playing against itself through a novel reinforcement learning method. It only has the rules of the game and starts to play against itself, making adjustments and keeping the versions that improve.

Blog Post: https://deepmind.com/blog/alphago-zero-learning-scratch/
Research Page: https://deepmind.com/research/alphago/

If you would like to replicate the research, there is an open source project that is based on the paper https://github.com/gcp/leela-zero. However, in order to get the same results as AplhaGo Zero, you would need to have the same weights, and in order to achieve similar weights, you would need to have access to the same computing power as they. It would take 1700 years on commodity computers. The projects aim is to make a distributed effort to repeat the work.

Course 4 [deeplearning.ai] has been released!

The fourth course, Convolutional Neural Networks of Deeplearning.ai has now been released on coursera. People have been waiting for this one, but i think that the delay was to make the material very up to date with current research results. The four weeks of learning deals with:

  1. Foundations of Convolutional Neural Networks
  2. Deep convolutional models: case studies
  3. Object detection
  4. Special applications: Face recognition & Neural style transfer

You only look once

The “YOLO9000: Better, Faster, Stronger” paper describes the improvements to the YOLO, You only look once, architecture that enables realtime object detection and classification. It can classify over 9000 object categories and outperforms Faster RCNN with ResNet and SSD while being significantly faster. They train on both COCO dataset for detection simultaneously with ImageNet for Classification and combine it with a wordtree so that they can also fallback to “dog” if they cannot classify for instance a specific dog breed.

The first version, and architecture can be seen in this paper.

Here is a video presentation: https://www.youtube.com/watch?v=GBu2jofRJtk