Self-driving cars may not have steering wheels in future, dev preview for PyTorch 1.0 is here, etc
What Monday morning isn't complete without some, er, AI code to play with
Roundup How about we kickstart the week with artificial intelligence goodies?
PyTorch 1.0 is (almost) out: Details about PyTorch 1.0, the latest version of the popular AI framework developed by Facebook, were teased out at the inaugural PyTorch Developer Conference this week in San Francisco.
PyTorch 1.0 comes with new updates that make it faster to train AI models across Python and C++, easier to deploy systems with better support from Caffe 2 and offers integration with different cloud platforms and hardware.
For example, Amazon Sagemaker and AWS Deep Learning Amazon Machine Images, both platforms that make it easier for developers to build and deploy models on Amazon’s Cloud, now comes with PyTorch built-in.
“With the pre-configured PyTorch environment within Amazon SageMaker, developers and data scientists can specify their scripts using a single API call to train locally or to submit a distributed training job,” Amazon said.
"With a second API call, developers can also now deploy PyTorch-trained models to a managed, highly available, online endpoint that can be automatically scaled up or down as demand requires. PyTorch developers can take advantage of Amazon SageMaker features like automatic model tuning, Amazon CloudWatch integration, Amazon VPC support, and more."
Google and Microsoft are also doing the same for its own cloud platforms. Google is now going one step further and is also working to support the new library on its TPU chips. Other hardware vendors like IBM, ARM, Intel, Nvidia, and Qualcomm also announced software tools to speed up compile and inference runtimes for PyTorch on their chips too.
Fast.ai, a startup offering deep learning courses, has built its own software library designed to work on top of PyTorch 1.0. We wrote about that in more detail here.
If you want to download the preview of PyTorch 1.0, you can do it here.
Safety regulations for self-driving cars: The US Department of Transportation is planning to update its rules to pave the way for fully self-driving cars.
In a report titled “Automated Vehicles 3.0,” it states that current guidelines require cars to have specific features such as steering wheels, brakes, accelerator pedals, and other control features, as well as the visibility for a human driver and vehicle status indicators.
Although the rules don’t hinder the development and testing phase of current technologies for autonomous vehicles, the DoT does say that it doesn’t accommodate for Level 4 and Level 5 cars that may or may not have drivers present.
The DoT hinted that future changes, made together with recommendations from the US National Traffic Safety Administration (NHTSA), might allow exceptions for fully driverless cars. If there is no need for steering wheels or other controls traditionally operated by a human driver, then future self-driving cars won’t be required to have them.
Relaxing regulations will make it easier for companies to test self-driving cars that deviate from more traditional designs. You can find the rest of the 80-page report here [PDF].
Create an AI game with a cute puppy: Folks over at Unity, designers of game engines, have created 'Puppo, The Corgi' for its ML-Agents Toolkit.
In games, non-player characters have predictable and scripted behaviours. They have to be varied to keep the game interesting, so over time it can be tedious to keep writing these scripts to get them to perform specific actions.
The ML-Agents Toolkit aims to help developers experiment if these scripts can be replaced with reinforcement learning (RL). Puppo is a demonstration that RL can be useful. Instead of explicitly hard-coding the cartoon corgi’s movements, Puppo learns to walk, run, jump and fetch a stick after being trained with RL.
When Puppo gets closer to its goal of picking up the stick, it gets a positive reward. If it strays from its target, takes too long, or spins around too much, it gets a score of zero. The game is a good demonstration of how RL works, and there is even a short video that shows how several simulated versions of Puppo are trained in parallel.
You can watch it action and play around with the code here. ®
We'll be examining machine learning, artificial intelligence, and data analytics, and what they mean for you, at Minds Mastering Machines in London, between October 15 and 17. Head to the website for the full agenda and ticket information.