This article is more than 1 year old
Facebook doesn't know its onions: Seeds ad banned after machine-learning algo found vegetable pic 'overtly sexual'
Plus: Waymo risks driverless car trial, ML papers now include code, and AI will write your dev CV for you
In brief Facebook's computer-vision algorithm flagged an innocuous advert for onions posted on its social media platform for being "overtly sexual."
The Seed Company by EW Gaze, a Canadian garden store, wanted to advertise its Walla Walla onion seeds on the social network, but its attempt was shot down after the image used in the advert was deemed too naughty by Facebook's software.
Although most would argue that there's nothing remotely sexy about a close-up shot of onions in a basket, Facebook's automated software clearly disagreed. A screenshot posted by the company said the image was classified as having "overtly sexual positioning".
In experiments, Vinay Prabhu, a researcher at a privacy-focused startup, revealed that the image seems to contain one offending onion in the middle that is often mistaken for porn in object-recognition models. Perhaps it somewhat resembles a breast if you squint really hard. It's possible that's what Facebook's algorithms have zeroed in on too.
One again, 5 lines of Python code + amazing @GradioML
— Vinay Prabhu (@vinayprabhu) October 9, 2020
results in a more nuanced understanding on the brittleness of a real-world deployed ML model.
Colab or didn't happen? Here: https://t.co/acTCmHyMQ5 pic.twitter.com/9olqMQlSc3
Waymo ramping up self-driving taxi experiments
Waymo has gone public with "fully driverless" cars in Phoenix, Arizona.
The self-driving taxi fleet service known as Waymo One is still under trial and it was previously only available to a select number of riders who had agreed to sign non-disclosure agreements. Now the Alphabet project has opened up the driver-free service to all of its Waymo One riders.
"We'll start with those who are already a part of Waymo One and, over the next several weeks, welcome more people directly into the service through our app (available on Google Play and the App Store)," it said.
Again, people will be vetted and chosen to expand the Waymo One early rider program – it's not available to the general public yet. Also only about 20 per cent of the rides for Waymo One customers are fully autonomous and driverless right now, although that's something Waymo hopes to offer in the "near-term".
You've got the arXiv paper, now here's the code
There's now an extra tab at the bottom of machine learning research on arXiv to host links to the corresponding AI code described in the paper.
Machine learning research can be difficult to understand and replicate. The models described in papers are often quite vague, and not many developers bother sharing their code.
Papers With Code, an open-source project set up by a team of machine learning engineers, have attempted to encourage more people to publish their code for everyone to see with a new feature embedded into arXiv. Authors can directly submit links to their code, and community-wide efforts with other engineers attempting to reproduce the software can be added too.
"Having code on arXiv makes it much easier for researchers and practitioners to build on the latest machine learning research," Papers With Code explained in a Medium post.
"We also hope this change has ripple effects on broader computational science beyond machine learning. Science is cumulative. Open science, including making available key artefacts such as code, helps to accelerate progress by making research easier to build upon."
Fake machine learning-generated resumes
We've already seen a site that hosts a never-ending stream of images of imaginary people generated by AI software. Now, there's one for, erm, fake software developer resumes.
A new example of a CV spat out by a neural network will emerge every time you hit refresh. The "Almost Real Resume" website was described on GitHub by its creators as "just for a bit of fun". They trained a model to mimic the text format and display in about 6,000 resumes.
"We did not include the real data the models are currently trained off for privacy purposes," they added. "Note: The models/output obviously suck. We didn't clean the data enough. And we didn't train them for long enough. (Who can afford that)"
Still, it's fun to flick through the fake examples, and see what names and job titles it comes up with. The model isn't perfect. The grammar in the text is often incorrect, and it's unlikely that employers looking for new engineers will be impressed. ®