Electrical Engineering Systems Seminar

Speaker: Sagi Ben Itzhak

M.Sc. student under the supervision of Prof. Nahum Kiryati

Wednesday, 26^th March 2025, at 15:00

Room 011, Kitot Building, Faculty of Engineering

Localization-Guided Supervision for Robust Medical Image Classification by Vision Transformers

Abstract

A major challenge in developing data-driven algorithms for medical imaging is the limited size of available datasets. Furthermore, these datasets often suffer from inter-site heterogeneity caused by the use of different scanners and scanning protocols. These factors may contribute to overfitting, which undermines the generalization ability and robustness of deep learning classification models in the medical domain, leading to inadequate performance in real-world applications. To address these challenges and mitigate overfitting, we propose a framework which incorporates explanation supervision during training of Vision Transformers (ViT) models for image classification. Our approach leverages foreground masks of the class object during training to regularize attribution maps extracted from ViT, encouraging the model to focus on relevant image regions and make predictions based on pertinent features. We introduce a new method for generating explanatory attribution maps from ViT-based models and construct a dual-loss function that combines a conventional classification loss with a term that regularizes attribution maps. Our approach demonstrates superior performance over existing methods on two challenging medical imaging datasets, highlighting its effectiveness in the medical domain and its potential for application in other fields.

השתתפות בסמינר תיתן קרדיט שמיעה לתלמידי תואר שני ושלישי = עפ"י רישום שם מלא + מספר ת.ז. בדף הנוכחות שיועבר באולם במהלך הסמינר

Electrical Engineering Systems Seminar

Speaker: Yanay Danan

M.Sc. student under the supervision of Dr. Jonatan Ostrometzky

Sunday, 23^rd March 2025, at 15:30

Room 011, Kitot Building, Faculty of Engineering

Short-Term Spatiotemporal Photo-Voltaic power generation forecasting based on Interpolated Video Modeling

Abstract

Accurate short-term solar-based power forecasting is becoming a crucial element for efficient real-time power grid management, particularly in smart-grid systems, in which advanced technologies such as dynamic storage solutions are being implemented (e.g., for cost planning and vehicle to grid (V2G) applications). Cloud cover significantly impact photovoltaic (PV) power generation, often causing unexpected fluctuations. Since cloud dynamics are based on physical properties, their movement, which affect the PV power generation directly, can be effectively captured - and their movement predicted. Inspired by opportunistic sensing techniques used in weather monitoring via wireless communication channels, in this study I propose a novel data-driven approach that leverages the available PV power measurements to directly predict future disturbances in the expected power generation. Our approach first constructs a disturbance field from existing PV power snapshots and generates a video-like input using those snapshots, which is then processed by spatio-temporal video forecasting methods. Specifically, we utilize PredRNN++ as a recurrent-based model and SimVPv2 as a recurrent-free model to estimate the future evolution of these disturbances. Different from past works, here, we estimate the full field of the cloud-based power disturbances in a selected area of interest, rather than for pre-determined specific location. We achieve an nRMSE of 7.2% for forcasting horizons of 30 minutes, 9.1% for 60 minutes, and 11.6% for 2 hours,  These results are close to the state-of-the-art prediction methodologies - but with the clear advantage of having the ability to forecast solar-generated power in locations from which no data is being collected, as well as to seamlessly adapt to any power-grid  changes (with respect to the installment or removal of PV power generation elements) without requiring model retraining nor additional or new sensor data collection.

השתתפות בסמינר תיתן קרדיט שמיעה לתלמידי תואר שני ושלישי = עפ"י רישום שם מלא + מספר ת.ז. בדף הנוכחות שיועבר באולם במהלך הסמינר

Electrical Engineering Systems Seminar

Speaker: Sergey Timinsky

M.Sc. student under the supervision of Prof. Hagit Messer Yaron and Dr. Jonatan Ostrometzky

Sunday, 16^th March 2025, at 15:30

Room 011, Kitot Building, Faculty of Engineering

Rain Estimation Over a Region Using CycleGan

Abstract

Accurately measuring rainfall is essential for weather forecasting, flood prediction, and water resource management. Traditional methods rely on rain gauges for direct measurements, radar systems for broader coverage and satellites. However, these methods face challenges due to sparse sensor distribution and data coverage.

A promising alternative is using wireless commercial microwave links (CMLs)—the infrastructure behind cellular networks. CMLs experience signal attenuation when it rains, allowing them to serve as cost-effective, high-resolution rainfall virtual sensors. However, current training machine learning models require paired CML-rain gauge data, which limits their applicability due to missing or misaligned measurements.

To overcome this limitation, we propose a CycleGAN-based framework that enables rainfall estimation without requiring paired datasets. Instead of relying on direct matches between CMLs and rain gauges, our method learns the relationship between the two through an unpaired training strategy.

We introduce two mapping functions:

• G:A→R (Converts attenuation to rain rate).
• F:R→A (Converts rain rate to attenuation).

By enforcing cycle consistency, the model ensures that translating between the two domains preserves data structure, even in the absence of direct pairing between a CML and a gauge.

Our method offers several key advantages:

• Works with missing or sparse data.
• Adapts to different regions without direct alignment.
• Enhances rain estimation accuracy with a built-in detector.

We evaluated our approach on real-world CML datasets and rain gauge data from Israel and the Netherlands, demonstrating high accuracy in estimating accumulated rainfall, especially in heavy rain events.

This framework expands the capabilities of deep learning for rainfall estimation by enabling models to learn from unpaired datasets. It provides a scalable and flexible solution that overcomes the limitations of traditional supervised approaches.

השתתפות בסמינר תיתן קרדיט שמיעה לתלמידי תואר שני ושלישי = עפ"י רישום שם מלא + מספר ת.ז. בדף הנוכחות שיועבר באולם במהלך הסמינר