לוח שנה משולב

Visual recognition of materials and their states is essential for understanding the world, from determining whether food is cooked, metal is rusted, or a chemical reaction has occurred. Collecting data that captures this vast variability is complex due to the scattered and gradual nature of material states. Manually annotating real-world images is constrained by cost and precision, while synthetic data, although accurate and inexpensive, lacks real-world diversity. This work aims to bridge this gap by infusing patterns automatically extracted from real-world images into synthetic data. We show that neural nets trained on this data outperform state-of-the-art zero-shot models like Clip and SAM on material recognition and segmentation tasks.

Reference:
Drehwald, Manuel S., et al. "One-shot recognition of any material anywhere using contrastive learning with physics-based rendering." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.
‏Eppel, Sagi, et al. "Infusing Synthetic Data with Real-World Patterns for Zero-Shot Material State Segmentation." The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track. 2024

Bio:
Researcher at the Weizmann Institute AI Hub with a Ph.D. in Chemistry and a postdoc in Materials Engineering. Research topics include:  solar cells,  crystallization,  computer vision, self-driving cars, and autonomous laboratories in academia and industry.

The map of synaptic connectivity between neurons shapes the computations that neural circuits carry - making the identification of the design principles of neural “connectomes” crucial for understanding brain development, learning, information processing, and behavior.We present a class of probabilistic generative models for the connectomes of different brain areas in zebrafish, worm, and mouse. Our models accurately replicate a wide range of circuit properties - synapse existence and strength, neuronal in-degree and out-degree, and sub-network motif frequencies -  using surprisingly small sets of biological and physical architectural features. We then show that simulated synthetic circuits generated by our models recapitulate the neural activity and computation performed by the real ones. We extend these generative models to study the development of connectomes over time, and show they accurately replicate the “developmental trajectory” of the connectome of C. elegans, revealing a simpler set of functional cell types than commonly assumed, and identifying distinct developmental epochs. We further study structure-function relationships in simulated spiking neural networks and learn a metric that predicts the similarity of networks based on a small set of architectural features. Our findings suggest that connectomes across species follow surprisingly simple design principles and offer a general computational framework for analyzing connectomes, linking their structure to function. FOR THE LATEST UPDATES AND CONTENT ON SOFT MATTER AND BIOLOGICAL PHYSICS AT THE WEIZMANN, VISIT OUR WEBSITE: https://www.biosoftweizmann.com/ 

While AI has achieved remarkable advancements in areas such as image recognition and natural language processing, its application in Earth and environmental sciences is still emerging. Unprecedented data from satellites, remote sensors, and in-situ measurements offers new opportunities to improve physics-based model forecasts of environmental systems with AI and to gain deeper insights. However, extreme systems as weather and climate events, pose distinct challenges for AI, such as limited sampling of rare events, non-trivial data augmentation, errors-in-variables, and complexities of transfer learning across diverse tasks. In this talk, I will explore these challenges and showcase AI architectures designed to address them. I will use specific examples of forecasting dust storms, precipitation extremes, and drought events in the Middle East.

Structured State Space Models (SSMs) are emerging as efficient alternatives to Transformers, forming the backbone of neural architectures such as S4 and Mamba. In this talk, I will present a series of works theoretically analyzing SSMs. I will begin with the implicit bias of Gradient Descent (GD) over SSMs, proving that it often leads to generalization, but is susceptible to clean-label poisoning attacks. I will then tackle the open question of the benefits of complex parameterizations for SSMs, proving formal separations between real and complex parameterizations: a real SSM can only match a complex SSM if either the dimension of the real SSM or the number of iterations required for its training is exponentially large. Taken together, the presented findings deepen our theoretical understanding of SSMs, and highlight their potential towards interpretable state-of-the-art AI systems.

Covered works were in collaboration with Yotam Alexander, Avichai Ben David, Edo Cohen-Karlik, Raja Giryes, Amir Globerson, Eden Lumbroso, Itamar Menuhin-Gruman, Yuval Ran-Milo, Noam Razin and Yonatan Slutzky.