Google on Wednesday emitted a TensorFlow preview, finally put its Edge TPU hardware on sale, and rolled out its dreaded robo-caller Duplex – all amid its annual Tensorflow developer conference in Silicon Valley.
Here’s a quick summary of what you should know.
TensorFlow 2.0 teaser: If you’re waiting for TensorFlow 2.0, Google’s popular machine-learning framework used to build neural networks, to land, you’ll have to wait longer. But it’s not all bad news: you can play with an early preview.
Known as TensorFlow 2.0 Alpha, this preview includes a few exercises for beginners and experts to work through to get a feel for the code. One of the top complaints about TensorFlow is that it’s clunky and not easy to grasp compared to other frameworks, such as the Python-based PyTorch.
TensorFlow 2.0 promises to be simpler, making it easier for developers to build, run, and deploy machine-learning models and in fewer lines of code. It’ll do this by integrating Keras, a higher level API library written in Python, more tightly to TensorFlow. Users do not need to download Keras separately, and can switch to the API directly in TensorFlow using a simple command.
TensorFlow for mobile phones: TensorFlow Federated is another version of the framework, this time suited for training machine-learning models using data obtained from a large set of devices – such as a big bunch of smartphones running your mobile app.
The name of the game here is decentralized learning. Typically, a neural network is schooled from a single dataset. When you're teaching a model using data obtained from a lot of individual personal devices, though, it’s not always possible to do this in a privacy-conscious way: you may at some point have to upload people's data from the endpoints to this central AI system to train its neural net.
A better way is the TensorFlow Federated way: some training occurs on each device, and information generated as a result of that process is uploaded to the central neural network to learn from.
Crucially, that uploaded info should not compromise an individual's privacy: it will be effectively anonymized adjustments to apply to the primary neural network, while personally identifying stuff stays on the phone. Fingers crossed.
“There are an estimated 3 billion smartphones in the world, and 7 billion connected devices," Googlers Alex Ingerman and Krzys Ostrowski waxed lyrically. "These phones and devices are constantly generating new data ... Wouldn’t it be better if we could run the data analysis and machine learning right on the devices where that data is generated, and still be able to aggregate together what’s been learned?”
You can try it out for yourself, here.
Google Coral Edge TPU: The web giant's Edge TPU, a low-power version of its homegrown AI math accelerator technology teased in July last year, is now on sale as a RaspberryPi-like single-board computer, and as a USB dongle, under the Google Coral brand.
It's supposed to be useful for programmers and engineers who want to play around with TensorFlow on silicon specifically designed to accelerate the framework.
In its dev board form, the hardware features an NXP i.MX 8M system-on-chip with four Arm Cortex-A53 CPU cores, and a Cortex-M4F core, for running code, an integrated GPU, and an Edge TPU coprocessor to perform AI tasks in fast specialist hardware. It also comes decked with Bluetooth, Wi-Fi, multiple USB ports, a microSD slot, an audio jack, a microphone, a gigabit Ethernet interface, a HDMI port, and a camera. It supports models developed using TensorFlow Lite, runs Debian-based Mendel Linux, and can be yours for $150.
If that doesn’t float your boat, there’s a USB accelerator stick that you can plug into any Linux-based host machine to speed up inference workloads using the dedicated hardware. Again, it can be used with TensorFlow Lite, and it costs $75.
Get me a table at the Introvert Seafood and Grill: Google Duplex, the unnerving robo-caller shown off last year, is now available to Pixel device owners in 43 US states. Basically, it makes phone calls for you to book appointments and restaurant reservations if you're too busy or awkward to do it yourself. Here's how it works:
“All it takes is a few seconds to tell your Assistant where you'd like to go,” Team Google explained. “Just ask the Assistant on your phone, 'Book a table for four people at [restaurant name] tomorrow night.' The Assistant will then call the restaurant to see if it can accommodate your request. Once your reservation is successfully made, you’ll receive a notification on your phone, an email update and a calendar invite so you don’t forget.”
Great. Now more people can be harassed by automated callers. As far as we're aware, Duplex won’t reveal it’s not actually a real human being on the other side of the line when speaking to businesses. Many believe that’s a bad idea, and so Google has provided a way for company owners to opt out of the thing. ®