The artificial intelligence giant, describing the weakness as a “typographic attack”, made the following statements in a statement made on March 4:
We believe this weakness is not a simple learning disability. Taking advantage of CLIP’s ability to read texts robustly, we found that even photographs of handwritten text can often fool the system.
These types of attacks still work. It does not require any technology other than pen and paper.
OpenAI pointed out that the typographic attack resembles adversarial images used to fool vision systems. Experts had discovered that such images could be used to mislead Tesla vehicles’ software and change lanes, for example, by sticking labels on the road.
However, according to the company, there is nothing to worry about yet. Because the aforementioned software is not used in any commercial product.
However, OpenAI thinks the weakness reflects some of the strengths of the experimental system CLIP.
Unlike other artificial intelligence systems, CLIP was found to have the ability to think about objects not only at the visual level, but also at the “conceptual” level. Experts had seen that the system contained “neurons that respond to the same concept, regardless of whether it is expressed symbolically or conceptually.”
One of them was the Spider-Man neuron. The neuron in question is based on Spider-Man’s photo and drawing; it was responding to the spider image and word. The firm noted that CLIP may at times fail to identify significant differences between them.
Independent Türkçe, The Guardian, The Verge
Compiled by: Uğurcan Yıldız
Pen and Paper Beat Artificial Intelligence: Elon Musk’s OpenAI Deceived By Handwriting
OpenAI, Founded by Elon Musk, Discovered an Unseen Neurone Outside the Human Brain in an Artificial Intelligence System
Artificial intelligence researchers have discovered neurons previously only seen in the human brain in an artificial intelligence system.