news
The AI beat Q* Bert’s game in an incredible way
Source: WWW.THEVERGE.COM
Links:
https://www.theverge.com/tldr/2018/2/28/17062338/ai-agent-atari-q-bert-cracked-bug-cheat?utm_campaign=Revue%20newsletter &utm_medium=Newsletter&utm_source=The%20Wild%20Week%20in%20AI
A recent experiment showed an artificial intelligence agent beating the 1980s classic Q* Bert in an unprecedented way. It found a bug that crashed the game at certain points. Training AI agents to spot unexpected errors in games could be a new way to maintain games.
Using machine learning to decode and enhance memory
Source: WWW.WIRED.COM
Links:
https://www.wired.com/story/ml-brain-boost/?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild% 20Week%20in%20AI
The researchers collected training data on electrodes implanted in patients’ brains while they solved memory tasks. The same electrodes were then used to stimulate the patient’s neural activity. Such stimulation improved the patients’ ability to remember words by an average of 15 percent.
Google’s AI chip smart camera is now available
Source: TECHCRUNCH.COM
Links:
https://techcrunch.com/2018/02/27/googles-ai-powered-clips-smart-camera-is-now-available/?utm_campaign=Revue%20newslette r&utm_medium=Newsletter&utm_source=The%20Wild%20Week%20in%20AI
Google calls the product “smart camera,” because machine learning is deployed on the device to make the experience as simple as possible. Google says that by using machine learning on the device, the device can capture and photograph the moment when the facial expressions of people and pets are most appropriate.
Articles & Tutorials
Crash Course machine Learning (Google)
Source: DEVELOPERS.GOOGLE.COM
This course series is Google’s 15 + hour crash course on machine learning, featuring a series of video lectures, practical case studies, and hands-on exercises. Use Tensorflow.
Research opportunities: NLP and transfer learning
Source: the RUDER. IO
Links:
http://ruder.io/requests-for-research/?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild%20Wee k%20in%20AI
This article provides inspiration and ideas for junior researchers and those trying to enter the research field for beginners. It collects a range of interesting research topics focusing on NLP and transfer learning.
Introduction to Variant autoencoders
Source: WWW.YOUTUBE.COM
This video will briefly introduce variant autoencoders, a class of neural networks that can learn to compress data in a completely unsupervised manner.
Can neural networks always do a good job in object recognition?
Source: AIWEIRDNESS.COM
Links:
http://aiweirdness.com/post/171451900302/do-neural-nets-dream-of-electric-sheep?utm_campaign=Revue%20newsletter&utm_medi um=Newsletter&utm_source=The%20Wild%20Week%20in%20AI
If everything is done according to the rules, image recognition will work well. But these algorithms show their weakness whenever people or sheep (objects to be identified) do something unexpected.
Can increasing depth speed up optimization?
Source: WWW.OFFCONVEX.ORG
Links:
http://www.offconvex.org/2018/03/02/acceleration-overparameterization/?utm_campaign=Revue%20newsletter&utm_medium=Newsle tter&utm_source=The%20Wild%20Week%20in%20AI
“What does depth mean for neural networks?” Is a basic problem in deep learning theory. The conventional wisdom, supported by theoretical studies, is that adding layers increases expressiveness, but optimization becomes more difficult. However, it turns out that adding depth can sometimes speed up optimization.
Code, project & data
Robotics Research Brief (OpenAI)
Source: BLOG.OPENAI.COM
Links:
https://blog.openai.com/ingredients-for-robotics-research/?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_sou rce=The%20Wild%20Week%20in%20AI
This article introduces eight open source simulated robot environments from OpenAI and the standard implementation of the Hindsight Experience Replay algorithm. The algorithm implementation can be found in the base library.
Lore: How do you model deep learning in 15 minutes
Source: TECH.INSTACART.COM
Lore is a Python framework that enables machine learning to be used by engineers and maintained by machine learning researchers. This article introduces examples of designing and deploying models using Lore.
TensorFlow has released version 1.6.0
Source: GITHUB.COM
Links:
https://github.com/tensorflow/tensorflow/releases/tag/v1.6.0?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_s ource=The%20Wild%20Week%20in%20AI
The new land version includes pre-built binaries for CUDA 9.0 and cuDNN 7, new optimizer internal apis for non-slot variables, improvements to SavedModels exports, added FFT support for XLA CPUS/Gpus, and more.
Keras implementation of GANs
Source: GITHUB.COM
Links:
https://github.com/eriklindernoren/Keras-GAN?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild %20Week%20in%20AI
This project performed Keras implementations on many GANs from various research papers. Some of these are simplified versions of the ones presented in the paper, but they are still great learning resources.
Hot style paper
Build agents with intrinsic motivation and self-awareness to learn to play
Source: ARXIV.ORG
Links:
https://arxiv.org/abs/1802.07442?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild%20Week%20in %20AI
Babies are experts at play, with an amazing ability to perform novel structured behaviors in an unstructured environment that lacks clear external reward signals. The authors investigate curiosity-driven intrinsic motivation and propose a network of “world models” that learn to predict the dynamic consequences of agent behavior. At the same time, they trained a separate “self model” that enabled the agent to track the error graph of its own world model and then use the self model against the developing world model.
Back to basics: Atari’s standard evolution strategy
Source: ARXIV.ORG
Links:
https://arxiv.org/abs/1802.08842?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild%20Week%20in %20AI
The authors qualitatively show that ES algorithms have very different performance characteristics than traditional RL algorithms: in some games, they learn to exploit the environment better, while other agents may fall into suboptimal local minima. Therefore, combining their advantages with those of traditional RL algorithms may lead to new advances in existing technologies.
Schmidhuber, Paul Schmidt
Source: ARXIV.ORG
Links:
https://arxiv.org/abs/1802.08864?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild%20Week%20in %20AI
Train the increasingly common problem solvers incrementally, constantly learning to solve new tasks without forgetting previous skills. Problem solvers are single recursive neural networks called ONE and can be trained in a variety of ways, for example, black box optimization, reinforcement learning, artificial evolution, and supervised and unsupervised learning.
The original post was published on March 6, 2018
Author: Abstract bacteria
This article is from the cloud community partner “Big Data Digest”. For relevant information, you can follow the wechat public account of “Big Data Digest”