🍊 The Juice: Mislabeled Data
Zumo Labs presents The Juice, a weekly newsletter focused on computer vision problems (and sometimes just regular problems). Get it while it’s fresh.
Week of April 5–April 9, 2021
____
Researchers from MIT published a paper last week finding “label errors in the test sets of 10 of the most commonly-used computer vision, natural language, and audio datasets.” While not terribly surprising given that ImageNet alone has previously been found to include slurs, it’s disappointing and somewhat alarming considering just how many AI systems are trained on these public datasets.
It’s one thing if a cat is mislabeled as a frog, as cited in the paper, but it’d be another thing altogether if the labels of say, street signs, were misidentified in a dataset. Is this a consequence of outsourcing labeling tasks to the lowest bidder? At what point does dataset label QA become as vital a step as model development? And what if instead of labeling the data, we generated data to match the exact labels the model expected?
____
#DefectDetection
John Deere is welding steel in 52 different factories around the world. Their defect detection has traditionally been a manual process executed by highly skilled technicians. Now, to bolster their Industry 4.0 cred, they’ve partnered with Intel on a computer vision-based solution. The system uses a special welding camera and custom software to log defects in real time and automatically stops the welding process if necessary.
At John Deere, ‘Hard Iron Meets Artificial Intelligence’, via Intel Newsroom.
#SkinDeepLearning
Suspicious pigmented lesions (SPLs) can be a sign of skin cancer and early detection is critical to potentially life-saving treatment. To help scale the review process that dermatologists typically perform, a team of researchers have created a tool that uses computer vision to analyze and categorize pigmented lesions from a wide-field image. This should make it possible to have skin checkups at your primary care doctor — or even at home — in the future.
An artificial intelligence tool that can help detect melanoma, via MIT News.
#Cornhole
We were so busy worrying about robots taking our jobs that we failed to mind our leisure. This is Susan, a $30,000 industrial robot arm that — thanks to a Jetson AGX and OpenCV — can drain buckets on cornhole boards all day long.
This Robot Arm Can Throw a Perfect Cornhole Game, via YouTube.
#EmotionRecognition
With remote schooling and remote work at all-time highs during the pandemic, the computer vision-aided emotion recognition space is seeing tremendous growth. But should we be skeptical of the promises of this type of software? As Kate Crawford says in this piece, “Countries around the world have regulations to enforce scientific rigour in developing medicines that treat the body. Tools that make claims about our minds should be afforded at least the same protection.”
Time to regulate AI that interprets human emotions, via Nature.
Back in December, when Dr. Timnit Gebru was fired from Google, Gebru’s manager and co-founder of Google Brain, Samy Bengio, wrote on Facebook that he was stunned she was removed without him being consulted. Now, Bengio is out as well, having sent an internal email this week announcing his resignation. The article below cites folks familiar with the matter, calling it a huge loss for Google that’s directly related to Gebru’s (and more recently, Margaret Mitchell’s) dismissal.
Google AI scientist Bengio resigns after colleagues’ firings: email, via Reuters.
____
📄 Paper of the Week
StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery
Manipulating the latent spaces of GANs can now be done with text! Up until now, messing with GAN latent spaces meant sliding along a specific degree of freedom identified as head rotation, age, or opened/closed mouth. This work uses CLIP models to convert a text defined transformation into an actual step in latent space. This work is another example of how powerful the combination of language and images can be in ML. As with any GAN paper, be sure to check out the cool pictures.
____
Think The Juice was worth the squeeze? Sign up here to receive it weekly.
