Huawei and Intel hype up AI hardware, TensorFlow tidbits, and more
A light snack of machine-learning news
Roundup Hello, welcome back to the AI roundup. Here’s a short list of what’s been happening so far since the Christmas and New Year break.
TensorFlow updates: Google has released new code for developers interested in training machine learning models more privately as well as a sneak peak of TensorFlow 2.0.
TensorFlow Privacy is a Python library that contains TensorFlow algorithms to train models with differential privacy for anonymizing data sets. It’s good for handling sensitive data like medical records, where you want to scrub the data of any characteristics that could potentially identify a patient.
Next, is the preview of the upcoming TensorFlow 2.0 updates. You can play around with it here - but be warned it’s not fully complete so it might be buggy and probably won’t work for advanced projects.
Here’s the link to download TensorFlow Privacy.
Sigh, more Intel AI chip fluff: Intel, never one to miss out on an opportunity to tout its promised Neural Network Processor chip, divulged a few details during the Consumer Electronics Show this week.
First, it announced that Chipzilla’s engineers were working with Facebook to finish the NNP chip for later this year. The announcement, however, isn’t actually new. Back in October 2017, Naveen Rao, veep of Intel's artificial intelligence products group, said he was “thrilled to have Facebook in close collaboration sharing their technical insights as we bring this new generation of AI hardware to market.”
Fast forward over a year later, and there's still no hardware to be seen. What is new, however, is that the chip will contain Intel’s 10nm Ice Lake cores, according to Rao.
Ok, guys...some news for you. NNP-i is 10nm Intel process. It will also include IceLake cores to handle general operations as well as the NN acceleration. It'll be a great product :) You're welcome.— Naveen Rao (@NaveenGRao) January 8, 2019
Intel hopes to ship the chip in laptops and servers by 2020 to handle the inference side of AI. We’ll believe it when we see it...
Huawei AI switch: Here’s more chip news. Huawei is also trying to nab itself a share of the AI cloud market and has released CloudEngine 16800, a data center switch for AI.
“The data center switch built for the AI era has three characteristics,” said Kevin Hu, president of Huawei network product line. It contains an “embedded AI chip, 48-port 400GE line card per slot, and the capability to evolve to the autonomous driving network.”
Details are all a little bit wishy washy, but the CloudEngine 16800 is essentially a server that Huawei hopes will eventually be used to process data generated from autonomous cars to do things like monitoring weather and traffic conditions.
Did a deepfake make it to TV? Donald Trump was transformed to look even more orange than normal and pulled funny faces during a speech about his Mexican border wall, broadcast this week by Q13, a Seattle TV station.
Here’s a side-by-side comparison with the original video feed, with the doctored version on the right. It’s pretty funny.
“Hopefully we can rise above partisan politics and in order to support security,” Trump said. He then briefly sticks out his tongue and continues his speech. The version shown by Q13, however, holds that face for longer and the fake Trump wags his tongue again. His face is also more zoomed in.
It’s not entirely clear which tools were used to create the doctored footage, but some, including the Seattle Times, fear it's a deepfake. A deepfake is typically any false material produced using deep-learning systems, such as generative adversarial networks, though in this case it's probably something simple like Adobe After Effects at play.
An employee at Q13, which is owned by Fox, was sacked over the debacle. It’s unknown if the staffer created the video or just allowed the clip to go to air. Regardless of whether or not this was a true deepfake – and it's obvious this one has been tampered with – people should be alert to subtly doctored video, which is easy to create using publicly available machine-learning tools. This could be the start of something more. ®