Portrait of Eldad Haber

Eldad Haber

Professor & NSERC Industrial Research Chair, University of British Columbia

Computational scientist working at the intersection of inverse problems, machine learning, and the geosciences.

Google Scholar GitHub UBC Profile LinkedIn Email

About

I am a Professor in the Department of Earth, Ocean & Atmospheric Sciences and the Institute of Applied Mathematics at UBC, where I hold the NSERC Industrial Research Chair in Computational Geoscience. My work develops computational methods for large-scale problems in science and engineering — from inverse problems and electromagnetic geophysics to modern machine learning. A central thread of my recent research is the view of deep neural networks as dynamical systems governed by ordinary and partial differential equations, which opens a principled path to architectures that are stable, reversible, and interpretable.

What draws me to the ODE/PDE view of deep networks is that it connects a new science — deep learning — to the classical mathematics we already trust. When a network is understood as a discretized dynamical system, its behaviour stops being a black box: that bridge to well-studied theory is, to me, where the explainability of deep networks comes from.

Selected Publications

  1. Stable Architectures for Deep Neural Networks
    E. Haber, L. Ruthotto · Inverse Problems, 2017
    Introduced the dynamical-systems (ODE) view of deep networks, giving a stability theory for training arbitrarily deep architectures.
    arXiv:1705.03341 · doi:10.1088/1361-6420/aa9a90
  2. Deep Neural Networks Motivated by Partial Differential Equations
    L. Ruthotto, E. Haber · Journal of Mathematical Imaging and Vision, 2020
    Derived families of convolutional architectures from parabolic and hyperbolic PDEs, connecting network design to well-understood physical models.
    arXiv:1804.04272 · doi:10.1007/s10851-019-00903-1
  3. Reversible Architectures for Arbitrarily Deep Residual Neural Networks
    B. Chang, L. Meng, E. Haber, L. Ruthotto, D. Begert, E. Holtham · AAAI, 2018
    Used the reversibility of the underlying ODE to build memory-efficient, stable residual networks of essentially unbounded depth.
    arXiv:1709.03698 · AAAI
  4. Joint Inversion: A Structural Approach
    E. Haber, D. Oldenburg · Inverse Problems, 1997
    A foundational framework for jointly inverting multiple geophysical data sets by coupling them through shared structure.
    doi:10.1088/0266-5611/13/1/006

Full list on Google Scholar (150+ papers, ~13,800 citations).

Research

Inverse Problems

Large-scale computational inverse problems and optimization for noisy, ill-posed systems.

Deep Learning & PDEs

Viewing neural networks as dynamical systems to design stable, reversible, explainable architectures.

Computational Geophysics

Forward modelling and inversion of electromagnetic and time-domain geophysical surveys.

Machine Learning Applications

Applied ML for imaging, mineral prospectivity mapping, and scientific discovery.

Highlights