|September 22 · Issue #58 · View online |
Happy reading and hacking,
As always if you like receiving this newsletter, you can help us by sharing it with your friends and colleagues.
| Introducing: Unity Machine Learning Agents |
Unity releases Machine Learning Agents SDK
allowing researchers and developers to transform games and simulations created using the Unity Editor into environments where intelligent agents can be trained using Deep Reinforcement Learning, Evolutionary Strategies, or other machine learning methods through a simple to use Python API.
| Expanding Facebook AI Research to Montreal |
The Montreal lab will house research scientists and engineers working on a wide range of ambitious AI research projects, but it will also have a special focus on reinforcement learning and dialog systems.
| Ever Wonder What A Rugged, Self-Contained AI Camera System Would Look Like? Here It Is. (sponsored) |
Most AI camera systems require a dry, dust-free operating environment and a high speed local network. A new device called DNNCam™ is different. With on-board processing and data storage, DNNCam™ can function as a stand-alone AI system. With a dustproof and waterproof case, DNNCam™ operates in nearly any environment. Learn more here.
| Why we should be Deeply Suspicious of BackPropagation |
An insightful elaboration of Geoffrey Hinton’s recent remarks that he was “deeply suspicious” of back-propagation. In short, there is no neurological analogy to learning through backpropagation nor are there true unsupervised learning techniques which do not rely on backpropagation and hence impractically long training time. This excellent quora
answer also expands on these points.
| Neural Artistic Style Transfer: A Comprehensive Look |
Great overview of the fundamentals of neural artistic style transfer including code snippets and a list of literature for a deeper dive.
| Learning to Optimize with Reinforcement Learning |
This post explores the potential of learning the algorithms that power machine learning instead of laboriously hand crafting them. An important advantage in the case of optimization algorithms might be that while they are currently devised under the assumption of convexity they are applied to non-convex objective functions; learning the optimization algorithm under the same setting as it will actually be used in practice could significantly boost performance.
| Build your own Machine Learning Visualizations with the new TensorBoard API |
Google updated the existing dashboards (tabs) in TensorBoard to use the new API, so they serve as examples for their new TensorBoard visualization API.
| Learning to Model Other Minds |
OpenAI is released an algorithm which accounts for the fact that other agents are learning too, and discovers self-interested yet collaborative strategies like tit-for-tat in the iterated prisoner’s dilemma.
| FAIRfairseq-py: Facebook AI Research Sequence-to-Sequence Toolkit written in Python. |
Great reference implementation in PyTorch of fairseq, a sequence-to-sequence learning toolkit from Facebook AI Research.
| vrn: Torch7/MATLAB code for "Large Pose 3D Face Reconstruction from a Single Image via Direct Volumetric CNN Regression" |
A Torch7/MATLAB implementation for “Large Pose 3D Face Reconstruction from a Single Image via Direct Volumetric CNN Regression” linked to above.
| europilot: A toolkit for controlling Euro Truck Simulator 2 with python to develop self-driving algorithms. |
Very cool project which allows you to capture the euro truck game screen input, and programmatically control the truck inside the simulator. Europilot captures the screen input and outputs a numpy array in realtime. The mapping between the relevant screenshot and the joystick values is written to a CSV file.
| A TensorFlow Implementation of Representation Learning by Learning to Count |
A TensorFlow implementation of Representation Learning by Learning to Count
. This paper proposes a novel framework for representation learning, where we are interested in learning good representations of visual content, by utilizing the concept of counting visual primitives.
| 3D Face Reconstruction from a Single Image |
The authors of this paper
are able to achieve astonishing 3D face reconstruction by training a Convolutional Neural Network (CNN) on an appropriate dataset consisting of 2D images and 3D facial models or scans.
| ImageNet Training in 24 Minutes |
The authors detail how they were able to train 100-epoch ImageNet training with AlexNet in 24 minutes on a supercomputer while finishing 90-epoch ImageNet-1k training with ResNet-50 on a NVIDIA M40 GPU takes 14 days!
| Real-valued (Medical) Time Series Generation with Recurrent Conditional GANs |
This paper demonstrates the use of GANs to create synthetic medical data. This could be a vital method for data augmentation in domains where its availability is limited. Moreover medical data can often suffers from insufficiently anonymization, this synthetic data could be shared and published without privacy concerns, or even used to augment or enrich similar datasets collected in different or smaller cohorts of patients. The authors also present novel methods for evaluating GANs by, for instance training a supervised model on synthetic data and evaluating it on real data and vice versa. Really interesting paper, well worth the read.