We are big fans of posts and videos that try to give you a gut-level intuition on technical topics. While [vas3k’s] post “Machine Learning for Everyone” fits the bill, we knew we’d like it from the opening sentences:
Machine Learning is like sex in high school. Everyone is talking about it, a few know what to do, and only your teacher is doing it.”
That sets the tone. What follows is a very comprehensive exposition of machine learning fundamentals. There is no focus on a particular tool, instead this is all the underpinnings. The original post was in Russian, but the English version is easy to read and doesn’t come off as a poor machine translation.
I’ve always been fascinated by AI and machine learning. Google TensorFlow offers tutorials and has been on my ‘to-learn’ list since it was first released, although I always seem to neglect it in favor of the shiniest new embedded platform.
Last July, I took note when Intel released the Neural Compute Stick. It looked like an oversized USB stick, and acted as an accelerator for local AI applications, especially machine vision. I thought it was a pretty neat idea: it allowed me to test out AI applications on embedded systems at a power cost of about 1W. It requires pre-trained models, but there are enough of them available now to do some interesting things.
You can add a few of them in a hub for parallel tasks. Image credit Intel Corporation.
I wasn’t convinced I would get great performance out of it, and forgot about it until last November when they released an improved version. Unambiguously named the ‘Neural Compute Stick 2’ (NCS2), it was reasonably priced and promised a 6-8x performance increase over the last model, so I decided to give it a try to see how well it worked.
I took a few days off work around Christmas to set up Intel’s OpenVino Toolkit on my laptop. The installation script provided by Intel wasn’t particularly user-friendly, but it worked well enough and included several example applications I could use to test performance. I found that face detection was possible with my webcam in near real-time (something like 19 FPS), and pose detection at about 3 FPS. So in accordance with the holiday spirit, it knows when I am sleeping, and knows when I’m awake.
That was promising, but the NCS2 was marketed as allowing AI processing on edge computing devices. I set about installing it on the Raspberry Pi 3 Model B+ and compiling the application samples to see if it worked better than previous methods. This turned out to be more difficult than I expected, and the main goal of this article is to share the process I followed and save some of you a little frustration.
This is an interesting development for media users and machine learning hackers: [doe300] has implemented OpenCL on the Raspberry Pi 3 Model B+called VCFCL That’s big news because the Pi 3+ has a Graphics Processing Unit (GPU) built into the processor that has been generally underutilized. The VideoCore IV GPU is built into the Broadcom BCM2837B0 and is surprisingly capable for a low-power chip. Although this GPU is well documented, it hasn’t been used that widely because you have to code specifically for this class of GPU. Adding in support for a high-level framework like OpenCL will make it much easier to run and adapt existing packages.
Most people are familiar with the idea that machine learning can be used to detect things like objects or people, but for anyone who’s not clear on how that process actually works should check out [Kurokesu]’s example project for detecting pedestrians. It goes into detail on exactly what software is used, how it is configured, and how to train with a dataset.
The application uses a USB camera and the back end work is done with Darknet, which is an open source framework for neural networks. Running on that framework is the YOLO (You Only Look Once) real-time object detection system. To get useful results, the system must be trained on large amounts of sample data. [Kurokesu] explains that while pre-trained networks can be used, it is still necessary to fine-tune the system by adding a dataset which more closely models the intended application. Training is itself a bit of a balancing act. A system that has been overly trained on a model dataset (or trained on too small of a dataset) will suffer from overfitting, a condition in which the system ends up being too picky and unable to usefully generalize. In terms of pedestrian detection, this results in false negatives — pedestrians that don’t get flagged because the system has too strict of an idea about what a pedestrian should look like.
[Kurokesu]’s walkthrough on pedestrian detection is great, but for those interested in taking a step further back and rolling their own projects, this fork of Darknet contains YOLO for Linux and Windows and includes practical notes and guides on installing, using, and training from a more general perspective. Interested in learning more about machine learning basics? Don’t forget Google has a free online crash course to get you up to speed.
Prosthetic arms can range from inarticulate pirate-style hooks to motorized five-digit hands. Control of any of them is difficult and carries a steep learning curve, rarely does their operation measure up to a human arm. Enhancements such as freely rotating wrist might be convenient, but progress in the field has a long way to go. Prosthetics with machine learning hold the promise of a huge step to making them easier to use, and work from Imperial College London and the University of Göttingen has made great progress.
The video below explains itself with a time-trial where a man must move clips from a horizontal bar to a nearby vertical bar. The task requires a pincer grasp and release on the handles, and rotation from the wrist. The old hardware does not perform the two operations simultaneously which seems clunky in comparison to the fluid motion of the learning model. User input to the arm is through electromyography (EMG), so it does not require brain surgery or even skin penetration.
We look forward to seeing this type of control emerging integrated with homemade prosthetics, but we do not expect them to be easy.
Machine learning is starting to come online in all kinds of arenas lately, and the trend is likely to continue for the forseeable future. What was once only available for operators of supercomputers has found use among anyone with a reasonably powerful desktop computer. The downsizing isn’t stopping there, though, as Microsoft is pushing development of machine learning for embedded systems now.
The Embedded Learning Library (ELL) is a set of tools for allowing Arduinos, Raspberry Pis, and the like to take advantage of machine learning algorithms despite their small size and reduced capability. Microsoft intended this library to be useful for anyone, and has examples available for things like computer vision, audio keyword recognition, and a small handful of other implementations. The library should be expandable to any application where machine learning would be beneficial for a small embedded system, though, so it’s not limited to these example applications.
There is one small speed bump to running a machine learning algorithm on your Raspberry Pi, though. The high processor load tends to cause small SoCs to overheat. But adding a heatsink and fan is something we’ve certainly seen before. Don’t let your lack of a supercomputer keep you from exploring machine learning if you see a benefit to it, and if you need more power than just one Raspberry Pi you can always build a cluster to get your task done just a little bit faster, too.
Neural networks are a key technology in the field of machine learning. A common technique is training them with sample data, and then asking them to create something new in the same vein. AI researcher [Janelle Shane] decided to task a neural network with a fun task – inventing new kinds of pie.
Using the char-rnn library, the network was initially trained on a sample of 2237 pie recipe titles, sourced from around the internet. Early iterations struggled to even spell “pie”, but as the network improved, so did the results. Where we can’t imagine how one would even make a “Sweesh Pie Ipple Pie”, later results, such as the “Impossible Maple Spinach Apple Pie” seem far more cromulent by comparison.
At this point, [Janelle] decided to mix things up, stirring in a further sample consisting of the names of various cookies and apples. The data were carefully sorted such that the network still prioritized pies, but this additional data gave the model a richer library to draw from. This led to such home-baked classics as Flangerson’s Blusty Tart and Chicken Pineapple Cream Pie.
On the surface, it’s a fun project with whimsical output, but fundamentally it highlights how much can be accomplished these days by standing on the shoulders of giants, so to speak. We’ve seen [Janelle]’s output before, too – naming tomatoes, no less.
