Summer Interns 2019 presentations

Speaker: Andrea Malleo

Title: A Guru Interface for the FINUFFT library.

Abstract: We have built a new software interface to the Flatiron library for performing the "non-uniform FFT" - a generalization of the FFT to off-grid data. We optimize the case of repeated executions via explicit plan instantiation, and present timing results. Supervisor: Alex Barnett.

Speaker: Reese Pathak

Title: Evaluating deep learning-based denoisers for cryo-EM images

Abstract: In this talk, I discuss the cryo-EM denoising problem, (briefly) survey methods employed by practitioners, and present results for some deep-learning based alternatives. Cryo-EM, an imaging technique used to resolve biomolecular structure, produces noisy images that require denoising. Current methods may perform poorly in low signal-to-noise regimes and do not attempt to exploit rotational or translational structure ("equivariance") in imaging data. As a possible remedy, we propose and empirically evaluate deep-learning based denoisers. Our results demonstrate that in some cases these methods exhibit improved performance over currently used methods. Acknowledging limitations of our study, I conclude with current and future directions for this work. Supervisors: Joakim Andén and Eftychios Pnevmatikakis.

Speaker: Kris Pan

Title: One-Photon Background Model with Masked Convolutional Neural Network

Abstract: One-photon calcium imaging experiments generate data with multiple overlapping background sources. The constrained non-negative matrix factorization(CNMF) approach demonstrated that it could accurately separate the background, improving detection of neural signals. However, this approach require computation of a large sparse weight matrix to reconstruct the background. We show that this step could be replaced by a masked convolutional neural network that is more time and space efficient. Supervisor: Eftychios Pnevmatikakis.

Speaker: Dan Fortunato

Title: Building the ultimate fully adaptive fast Poisson solver in domains with smooth boundaries

Abstract: We present a new framework for fast solution of elliptic BVPs with inhomogeneous volume data and boundary conditions on smooth boundaries. FFT or box-code based adaptive spectral solvers are efficient for the volume term, but require special correction near the boundary to allow them to be combined with a boundary integral solve for the boundary conditions. We avoid function extension and cut-cell quadratures near the boundary by using a smooth partition of unity defined by a variable-width annular region inside the domain conforming with the boundary. We show initial convergence results of our scheme. Supervisors: David Stein and Alex Barnett.

Speaker: Leo Simpson

Title: TBA. Optimization package for robust statistics using proximal operators.

Abstract: TBA. Supervisor: Christian Müller.