EE Seminar: Learning-to-Bound: A Framework for Predicting from data Approximate Fundamental Bounds on Parameter Estimation Performance
הרישום לסמינר יבוצע בתחילת הסמינר באמצעות סריקת הברקוד למודל (יש להיכנס לפני כן למודל, לא באמצעות האפליקציה)
Registration to the seminar is done at the beginning of the seminar by scanning the barcode for the Moodle (Please enter ahead to the Moodle, NOT by application)
Electrical Engineering Systems Seminar
Speaker: Hai Victor Habi
Ph.D. student under the supervision of Prof. Hagit Messer-Yaron & Prof. Yoram Bresler
Wednesday, 21st January 2026, at 15:00
Room 011, Kitot Building, Faculty of Engineering
Learning-to-Bound: A Framework for Predicting from data Approximate Fundamental Bounds on Parameter Estimation Performance
Abstract
Bounds on the performance of any parameter estimator, such as the Cramér–Rao bound (CRB), are fundamental tools in statistical signal processing. They provide lower bounds on the performance of parameter estimators and have been widely used to study a broad range of problems and use cases, including understanding the fundamental limitations of a problem, system design (e.g., array configuration, waveform adesign), and benchmarking different estimators. However, obtaining these bounds typically requires a precise and explicit statistical model of the measurement. In many practical applications, such a model is unavailable.
To address this limitation, we built on the recent success of generative models (such as generative adversarial networks, normalizing flow, diffusion) in modeling complex, high-dimensional distributions and adapted them for a core tool in statistical signal processing. Specifically, this work introduces a novel data-driven approach for computing approximate estimation performance bounds. Instead, we use a training data set consisting of parameter and corresponding measurement (observation) pairs, to predict fundamental bounds on the performance of any estimator that uses data drawn from the same unknown distribution. We derive data-driven formulations for several well-known estimation performance bounds, including the CRB, the Bayesian Cramér–Rao bound (BCRB), the Barankin bound, and the misspecified Cramér–Rao bound (MCRB).
In the Learning-to-Bound framework, we first learn a generative model from data that approximate the distributions of the measurements and the prior. Using the generative model, we predict an approximation of fundamental bounds on estimation performance. We call this framework learning-to-bound. Specifically, in the non-Bayesian setting, we employ normalizing flows to learn the measurement distribution. This allows both sample generation and likelihood evaluation, which are then used to compute the CRB, the misspecified CRB (MCRB), and the Barankin bound. In the Bayesian setting, we derive a new type of score matching, referred to as Fisher score matching, and combine it with standard score matching to estimate the BCRB. We provide a theoretical analysis of the proposed methods, including bounds on both the learning error and on the approximation error of the learned bound.
Numerical experiments validate the proposed approach. In addition, experiments on several real-world problems, including image denoising and edge detection using a learned camera noise model, frequency estimation under underwater noise, and one-bit quantization scenarios, demonstrate the effectiveness and advantages of the Learning-to-Bound approach.
-סמינר זה ייחשב כסמינר שמיעה לתלמידי תואר שני ושלישי-
This Seminar Is Considered A Hearing Seminar For Msc/Phd Students-
