Click the above “Beginner Learning Vision” to choose to add “Star” or “Top”
Important content delivered immediately
【Guide】How big is the gap between AI and human vision? Researchers from UC Berkeley and other universities created a dataset of 7,500 “natural adversarial examples”. After testing many machine vision systems, they found that AI’s accuracy dropped by 90%! In some cases, the software could only recognize 2%-3% of the images. The consequences of such AI being used in self-driving cars are unimaginable!
In recent years, there have been significant improvements in computer vision, but serious errors can still occur. The errors are so frequent that there is a research field dedicated to studying images that AI often misrecognizes, called “adversarial images”. These can be seen as optical illusions for computers; when you see a cat in a tree, the AI sees a squirrel.
AI misidentifies a cat climbing a tree as a squirrel
It is essential to study these images. When we place machine vision systems at the core of new technologies such as AI security cameras and self-driving cars, we believe that computers see the world the same way we do. Adversarial images prove this is not the case.
However, while much of the focus in this field is on images specifically designed to fool AI (such as Google’s algorithm mistaking a 3D printed turtle for a gun), these misleading images can also occur naturally. Such images are more concerning because they indicate that even images we did not intentionally create can confuse visual systems.
Google AI mistook this turtle for a gun
To demonstrate this, a group of researchers from UC Berkeley, the University of Washington, and the University of Chicago created a dataset of 7,500 “natural adversarial examples” and tested many machine vision systems on these data, finding their accuracy dropped by 90%, and in some cases, the software could only recognize 2%-3% of the images.
Below are some examples from the “natural adversarial examples” dataset:
AI sees a “shipwreck” when it’s actually a bug crawling on a dead leaf
AI sees a “torch”
AI sees a “ladybug”
AI sees a “sundial”
AI sees a “baseball player”
AI sees a “person driving a go-kart”
In their paper, researchers state that these data are expected to help cultivate more robust visual systems. They explain that these images exploit “deep flaws” that arise from the software’s “overreliance on color, texture, and background cues” to recognize what it sees.
For example, in the image below, AI wrongly identifies the image on the left as a nail, possibly due to the wood grain background of the image. In the image on the right, they only notice the hummingbird feeder but miss the fact that there is no actual hummingbird present.
The four dragonfly photos below are analyzed by AI based on color and texture, and from left to right, they are identified as a skunk, banana, sea lion, and glove. We can see why AI makes these mistakes from each image.
It is no news that AI systems make these mistakes. For years, researchers have warned that visual systems created using deep learning are “shallow” and “fragile”, and they do not understand the world’s subtle nuances as humans do.
These AI systems have been trained on thousands of sample images, but we often do not know which exact elements in the images the AI uses to make its judgments.
Some studies suggest that, considering overall shapes and content, algorithms do not view images holistically but focus on specific textures and details. The results given in this dataset seem to support this explanation; for example, images showing clear shadows on bright surfaces are misidentified as sundials.
But does this mean that these machine vision systems are hopeless? Not at all. Generally, the mistakes these systems make are minor errors, such as misidentifying a drainage cover as a manhole or mistaking a truck for a luxury car.
While researchers say these “natural adversarial examples” can fool a variety of visual systems, this does not mean they can fool all systems. Many machine vision systems are highly specialized, such as those designed to identify diseases in medical scan images. Although these systems have their own shortcomings and may not understand the world and humans, this does not prevent them from detecting and diagnosing cancer.
Machine vision systems may sometimes be quick and flawed, but they typically produce results. Such research exposes the blind spots and gaps in machine imaging research, and our next task is to figure out how to fill these gaps.
Paper link:
https://arxiv.org/pdf/1907.07174.pdf
Code and dataset:
https://github.com/hendrycks/natural-adv-examples
Original link:
https://www.theverge.com/2019/7/19/20700481/ai-machine-learning-vision-system-naturally-occuring-adversarial-examples
Discussion group
Welcome to join the reader group of the WeChat public account to communicate with peers. Currently, there are WeChat groups for SLAM、three-dimensional vision、sensors、self-driving, computational photography、detection, segmentation, recognition, medical imaging, GAN、algorithm competitions and other topics (which will be gradually subdivided in the future). Please scan the WeChat ID below to join the group, and note: “nickname + school/company + research direction”, for example: “Zhang San + Shanghai Jiao Tong University + Visual SLAM”. Please do not send advertisements in the group, or you will be removed from the group. Thank you for your understanding~
