Integrated Calendar

In my talk I will introduce a novel framework that applies a principle from information theory, that of Minimum Description Length (MDL), to understand how regulation of human gene expression across organs, tissues is shaped by regulatory architecture.Examination of expression patterns of human genes across the body reveals an intriguing duality: While many genes are expressed in only one tissue, others, known as “housekeeping genes”, are ubiquitously expressed in essentially every tissue. Yet, interestingly, a considerable portion of the genes are on the mid-range, deliberately expressed in many tissues but are also absent in many others.Intuitively, in human language terms, specifying the expression program of the genes on the two ends of the spectrum requires a short description – e.g. “expressed in all tissues”, or “expressed only in brain”. Yet specifying the expression of genes in the middle of the scale requires longer description, or a longer MDL, having to specify in each tissue if the gene is expressed or not, and at what level. We sought to measure regulatory complexity of each human gene and examine if the MDL principle predicts and explains regulatory complexity. Our findings lend support to the MDL principle’s prediction. Our measure of regulatory complexity of a gene’s expression pattern can be predicted by quantifying its regulatory information content. In the talk we shall discuss evolutionary implications to the development of multi-cellularity.FOR THE LATEST UPDATES AND CONTENT ON SOFT MATTER AND BIOLOGICAL PHYSICS AT THE WEIZMANN, VISIT OUR WEBSITE: https://www.biosoftweizmann.com/ 

This is the first of a two-part talk covering pages 19–30 of [Lub]

Olfaction plays a crucial role in human perception, yet individuals with isolated congenital anosmia (ICA) navigate the world without ever experiencing smells. I will present findings from my research investigating how ICA influences behavior, physiology, and social interaction, with a particular focus on responses to fear-related chemosignals. First, I will describe distinct respiratory patterns observed in anosmics, both in general and in response to specific social and environmental contexts. While anosmics exhibited reduced sniff modulation in response to social cues, they adapted their breathing patterns similarly to normosmics during environmental transitions, suggesting compensatory mechanisms independent of olfaction. Next, I will discuss behavioral and physiological responses to fear sweat, demonstrating that normosmics exhibit heightened emotional engagement, whereas anosmics show reversed or minimal responses. These findings emphasize the role of olfactory input in emotional contagion and align with research on sensory processing differences in other populations. Finally, I will present functional brain imaging data revealing distinct neural activation patterns in anosmics in response to fear-related cues. While anosmics lack olfactory perception, they recruit alternative neural pathways, with lateralized brain activity suggesting adaptive mechanisms for processing social and emotional information. Together, these findings provide novel insights into the interplay between olfaction, social behavior, and neural adaptation. They highlight how anosmics adjust to sensory deficits and raise important questions about the flexibility of human sensory processing.

While neural networks revolutionized fields like computer vision and natural language processing, training them is time consuming and computationally demanding. Now, imagine that instead of training, one could simply search online for the most accurate model for their specific task and use it directly.  With more than 1.5 million public models available on HuggingFace, this is becoming feasible. However, retrieving the right model for each task is non-trivial, and the key technical challenge is designing meaningful representations of models themselves. To address this, we introduce ProbeLog, a probing-based approach for representing classification models from their weights. Then, by projecting CLIP text embeddings into this representation space we can enable zero-shot model retrieval.
 

Bio:
Jonathan is a Computer Science PhD student at The Hebrew University of Jerusalem, under the supervision of Prof. Yedid Hoshen.  Currently, He is working on weight space learning, aiming to understand what information can be extracted from pre-trained neural networks. His recent research can be viewed at https://pages.cs.huji.ac.il/jonkahana.