placeholder

Multimodal Neurons in Artificial Neural Networks

We’ve discovered neurons in CLIP that respond to the same concept whether presented literally, symbolically, or conceptually.

Click to view the original at openai.com

Hasnain says:

This attack is amazing - need to read the paper more in depth later. The tool recognizes an apple properly, but if you write “iPod” on it the model suddenly thinks it’s an iPod.

“We refer to these attacks as typographic attacks. We believe attacks such as those described above are far from simply an academic concern. By exploiting the model’s ability to read text robustly, we find that even photographs of hand-written text can often fool the model. Like the Adversarial Patch,22 this attack works in the wild; but unlike such attacks, it requires no more technology than pen and paper.”

Posted on 2021-03-04T23:19:46+0000