Month: July 2024

Prof. Brik was born in the village of Abu Snan in the Western Galilee and completed his undergraduate degree at Ben-Gurion University of the Negev and his master’s degree and PhD at the Technion. After completing his PhD, he continued to a postdoctoral fellowship at the Scripps Research Institute in San Diego, California. Upon his return to Israel, Prof. Brik was appointed as a senior lecturer at Ben-Gurion University, and within five years, he was promoted to full professor. In 2015, he was recruited to the Technion and has since been a faculty member in the Schulich Faculty of Chemistry.

Irith Rappaport, daughter of the founders of the Bruce and Ruth Rappaport Foundation said: “The vision behind the prizes established by my late parents is to encourage excellence, whether in the category for biomedical research or in the categories for art and for women generating change. The research prize is a professional award that has gained a reputation as one of the prestigious and important awards in Israel. Moreover, as someone who was born in Haifa, the most diverse city in Israel, the model of a shared society was and remains self-evident to me. Therefore, I am proud to award the prize each year to the winners who represent Israeli society as a whole. I am confident that my parents would have been proud that the prizes they established sanctify the values ​​they instilled in us: excellence, equality, and mutual respect regardless of religion, race, and gender.”

Prof. Brik focuses on biological chemistry and develops innovative methods for the synthesis of proteins with unique properties, such as those that undergo changes after translation. These proteins are used in structural, biochemical, biophysical, and functional analyses, as well as in understanding their roles in various diseases and in developing innovative treatments for these diseases. He has won numerous awards and grants, including the Humboldt Prize (Germany), the Hirata Award (Japan), the Tetrahedron Young Investigator Award, the Teva Award for Excellence in memory of Eli Hurvitz (Israel), the Israel Chemical Society Prize for Outstanding Young Chemist, and the ERC Advanced Grant, awarded to leading researchers with outstanding achievements in research. In 2019, he was elected as a member of the Israel Young Academy.

The Rappaport Prize for Outstanding Doctoral Students was awarded to Dr. Alia Ghrayeb, who completed her medical studies at the Rappaport Faculty of Medicine and continued her doctoral studies at the same faculty under the supervision of Prof. Eyal Gottlieb. She is currently under the supervision of Prof. Zaid Abbasi. Dr. Ghrayeb researches key metabolic changes in fatty liver disease, the most common cause of chronic liver disease and the leading cause of liver transplants worldwide. Using metabolic tools, she revealed not only changes in lipid profiles but also significant changes in amino acid metabolism, particularly in the glycine pathway. She extended these findings to demonstrate that such metabolic changes are also present in cardiovascular diseases, the main cause of mortality among people with fatty liver disease. Although the therapeutic potential of glycine has been previously studied, the mechanism leading to a decrease in glycine levels and the metabolic implications of this deficiency have not yet been thoroughly investigated. Using metabolic tools combined with unique pharmacological and genetic approaches in an experimental model of fatty liver in mice, Dr. Ghrayeb demonstrated that increased synthesis of serine from glycine through the reverse activity of the mitochondrial enzyme 2SHMT is the main cause of decreased glycine levels in fatty liver. Additionally, genetically engineered mice that do not express 2SHMT in the liver exhibit a preferred antioxidant capacity, which protects them from oxidative stress damage. Alia’s deep understanding of glycine metabolism in fatty liver holds potential for therapeutic applications in the not-too-distant future.

 For the video that was screened in honor of Prof. Brik at the award ceremony.

Image generator models – systems that produce new images based on textual descriptions – have become a common and well-known phenomenon in the past year. Their continuous improvement, largely relying on developments in the field of artificial intelligence, makes them an important resource in various fields.

To achieve good results, these models are trained on vast amounts of image-text pairs – for example, matching the text “picture of a dog” to a picture of a dog, repeated millions of times. Through this training, the model learns to generate original images of dogs.

 

However, as noted by Hadas Orgad, a doctoral student from the Henry and Marilyn Taub Faculty of Computer Science, and Bahjat Kawar a graduate of the same Faculty, “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, for example, ‘the sky is blue,’ and they allow us to obtain beautiful images even with short and simple descriptions. On the other hand, the model also encodes incorrect or irrelevant assumptions about the world, as well as societal biases. For example, if we ask Stable Diffusion (a very popular image generator) for a picture of a CEO, we will only get pictures of women in 4% of cases.”

Another problem these models face is the significant number of changes occurring in the world around us. The models cannot adapt to the changes after the training process. As Dana Arad, also a doctoral student at the Taub Faculty of Computer Science, explains, “during their training process, models also learn a lot of factual knowledge about the world. For example, models learn the identities of heads of state, presidents, and even actors who portrayed popular characters in TV series. Such models are no longer updated after their training process, so if we ask a model today to generate a picture of the President of the United States, we might still reasonably receive 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 actions.”

The “traditional” solution to these problems is constant data correction by the user, retraining, or fine-tuning. However, these fixes incur high costs financially, in terms of workload, in terms of result quality, and in environmental aspects (due to the longer operation of computer servers). Additionally, implementing these methods does not guarantee control over unwanted assumptions or new assumptions that may arise. “Therefore,” they explain, “we would like a precise method to control the assumptions that the model encodes.”

The methods developed by the doctoral students under the guidance of Dr. Yonatan Belinkov address this need. The first method, developed by Orgad and Kawar and called TIME (Text-to-Image Model Editing), allows for the quick and efficient correction of biases and assumptions. The reason for this is that the correction does not require fine-tuning, retraining, or changing the language model and altering the text interpretation tools, but only a partial re-editing of around 1.95% of the model’s parameters. Moreover, the same editing process is performed in less than a second. In ongoing research based on TIME, called UCE, which has been developed in collaboration with Northeastern and MIT universities, they proposed a way to control a variety of undesirable ethical behaviors of the model – such as copyright infringement or social biases – by removing unwanted associations from the model such as offensive content or artistic styles of different artists.

The methods receive inputs from the user regarding a fact or assumption they want to edit. For example, in cases of implicit assumptions, the method receives a “source” on which the model bases implicit assumptions (e.g., “red roses” by default the model assumes red roses) and a “target” that describes the same circumstances but with the desired features (e.g., “blue roses”). When wanting to use the method for role editing, the method receives an editing request (e.g., “President of the United States”) and then a “source” and “target” (“Donald Trump” and “Joe Biden,” respectively). The researchers collected about 200 works and assumptions on which they tested the editing methods and showed that these are efficient methods for updating information and correcting biases.

TIME was presented in October 2023 at the ICCV conference, one of the important conferences in the field of computer vision and machine learning. UCE was recently presented at the WACV conference. ReFACT was presented in Mexico at the NAACL conference, one of the leading conferences in natural language processing research.

The research was supported by the Israel Science Foundation (ISF), the Azrieli Foundation, Open Philanthropy, FTX Future Fund, the Crown Family Foundation, and the Council for Higher Education. Hadas Orgad is an Apple AI doctoral fellow.

Click here for the project website and papers:

https://aclanthology.org/2024.naacl-long.140/

https://technion-cs-nlp.github.io/ReFACT/

https://time-diffusion.github.io/