"Models learn the identities of prime ministers, presidents and even actors who played popular characters in TV series. Such models stop updating after their training process. The team developed an algorithm that makes it possible to do this without the need for new and expensive training * In addition, the researchers overcame a bias to prefer men when the request is casual
Image-generating models - systems that produce new images based on a textual description - have become a common phenomenon known to the general public in the last year. Their constant improvement, which relies to a large extent on developments in the world of artificial intelligence, makes them an important resource in the various areas of life.
To achieve good performance, these models are trained on huge amounts of image-caption pairs - for example pairing the caption "photo of a dog" with a photo of a dog, thousands and millions of times. From this training, the model learns to produce original images of dogs.
However, as noted Doctoral student Hadas Orgad and Dr. Bhajat Kawar, PhD graduate from the Taub Faculty of Computer Sciences, "Since these models are trained on a lot of data from the real world, they acquire and internalize assumptions about the world during the training process. Some of these assumptions are useful, like for example - "the sky is blue" - and they are what allow us to get beautiful pictures even based on short and simple descriptions. On the other hand, the model also encodes incorrect, or irrelevant, assumptions about the world, as well as social biases. For example, if we ask Stable Diffusion (a very well-known image generator) for a photo of a CEO, we will only receive photos of women in 4% of the cases."
Another problem that such models face is the large amount of changes that occur in the world around us, while the models do not change after the training process. Dana Arad, also a PhD student at the Taub Faculty of Computer Science, explains that "in their training process, models also learn a lot of factual knowledge about the world. For example, models learn the identities of prime ministers, presidents and even actors who played popular characters in TV series. Such models stop being updated after their training process, so if we ask a model today to produce us a picture of the President of the United States, it is still likely that we will get a picture of Donald Trump, who of course has not been the president in recent years. We wanted to develop an efficient way to update the information without relying on expensive operations."
The "traditional" solution to these problems is constant correction of the data by the user, re-training or fine-tuning; However, these repairs have a high cost both financially, both in terms of the duration of the work, both in terms of the quality of the results and in environmental aspects (due to longer operation of computer servers). In addition, using these methods does not guarantee control over the unwanted assumptions or over new assumptions that can be created. "Therefore," the researchers explain, "we would like a method that would allow precise and controlled control of those assumptions that the model encodes."
The methods developed by the doctoral students under the guidance of Dr. Yonatan Blinkov makes the aforementioned need redundant. The first method, developed by Orgad and Kawar and Kroya TEAM, allows quick and efficient correction of biases and assumptions. The reason for this is that the same correction does not require tuning, or re-training, but only the re-editing of a small part of the parameters - only about 1.95% of the model's parameters. Furthermore, the same process of re-editing is done quickly - in less than a second. In addition, in a follow-up work based on TIME, called UCE, developed in collaboration with Northeastern and MIT universities in Boston, the researchers proposed a way to control the many unwanted ethical behaviors of the model - such as copyright infringement or social biases - by deleting unwanted associations from the model such as Offensive content, or artistic styles of various artists.
Another method, later developed by Arad and Orgad and called ReFACT, offers a different algorithm for editing the parameters and thus succeeds in achieving higher quality and more accurate performance. ReFACT edits an even smaller percentage of the model's parameters - only 0.25% - and manages to make more varied edits, even in cases where previous methods failed, while maintaining the quality of the images and the facts and assumptions of the model that we would like to preserve.
The methods receive inputs from the user regarding the fact or assumption that we would like to edit. For example, in the case of unstated assumptions, the method receives a "source" on the basis of which the model builds unstated assumptions ("a bouquet of roses" for example, for which the model usually assumes red roses) and a "target" that describes the same circumstances but with the desired attributes (e.g. "A bouquet of blue roses", in order to edit the model so that it assumes that roses are blue from now on). When we want to use the method to edit positions, the method will receive the edit request (for example "President of the United States") followed by "source" and "target" ("Donald Trump" and "Joe Biden", respectively). The researchers collected about 200 facts and assumptions, tested the editing methods on them and showed that these are effective methods for updating information and correcting biases.
TEAM It was presented last October at a conference ICCV, one of the most important conferences in the field of computer vision and machine learning, and the follow-up work UCE was recently presented at the WACV conference. ReFACTRecently presented at the NAACL conference, one of the leading conferences for research in the field of natural language processing.
The research was supported by the National Science Foundation (ISF), the Azrieli Foundation, Open Philanthropy, FTX Future Fund, the Crown Family Foundation, and MLA-VAT. Hadas Orgad is an Apple Fellow in Artificial Intelligence for doctoral students. Dana Arad is a fellow in the Ariane de Rothschild scholarship for doctoral students.
To the article click here
For pictures click here
In pictures:
- Dr. Yonatan Blinkov
- Woven myrtle
- Dr. Bhajat Kawar
- Dana Arad
- In the picture: the correction of the gender bias when the input is "A developer". Left: before editing with TIME (implicit assumption: A developer is a man). Right: after editing.
- In the picture, a knowledge update in the model made with the help of ReFACT. On the left, the original images produced by the model. On the right, after editing. The edits also successfully generalize to close formulations, and show that the method succeeds in performing a significant edit in the knowledge encoded in the model.
Credit: Technion spokespeople
For more details: Doron Shaham, Technion spokesperson - 050-3109088
Doron Shaham
Technion Spokesperson
Division of Public Affairs and Resource Development
Tel: 077-8871992
Cell: 972-50-3109088
Email: spokesperson@technion.ac.il
...
[message cut off] Display the entire message
One attachment • Scanned by Gmail
