Software & Code
Projects
Check out our Github Group Repo
Implementation for the experiment in our pre-print on paper on Fast Hierarchical Games for Image Explanations, which provides a fast and accurate method for black-box interpretability particular suited for multiple instance learning problems.
Implementation for the experiment in our NeurIPS paper on Adversarial Robustness of Supervised Sparse Coding. We provide synthetic examples as well as training of supervised training of models for image classification.
LPCNN provides a solution to the ill-posed dipole deconvolution problem in Quantitative Susceptibility Mapping (QSM) as detailed in our MICCAI paper. By integrating proximal gradient descent with deep learning, it is the first deep learning based QSM method that can handle an arbitrary number of phase input measurements. Our implementation repo includes QSM training datasets (n=8, with local phase data acquired at 7T and 4-5 orientations COSMOS) and PyTorch implementations of LPCNN.
This is a PyTorch implementation for constructing and training principled residual networks from feed-forward ones based on the unfolding of a Multi-Layer Iterative Soft Thesholding Algorithm (ML-ISTA). These networks are designed based on the minimization of a Multi Layer Basis Pursuit problem, serving signals in the multi-layer convolutional sparse coding model.
This is a general purpose online dictionary learning algorithm, written in pytorch with GPU functionality - which I basically put together to run little experiments and tests.
The Sliced Based approach allows for a fast and light solution to the convolutional sparse coding problem. This algorithm leverages the decomposition of the global image into local slices: simple small dimensional elements that make up the global content of the high dimensional signals.
OSDL is a fast dictionary learning algorithm particularly design to work on high dimensional data. It exploits a sparse combination of separable and cropped wavelets in order to reduce the number of free parameters, and its online learning scheme allows it to train on million of examples. This work was presented and demonstrated in Trainlets: Dictionary Learning in High Dimensions and Large Inpainting of Face Images with Trainlets. The inpainting script and trained faces dictionary for the latter paper can be found here.
This is a demo on developing a classification scheme for person identification from ECG signals, based on an embedding with Diffusion Maps. This code exploits the Scattering Transform to make deformation invariant observations (and needs the ScantNet package). This script was applied to the ECG-ID dataset from Physionet, and just a couple of signals are included for demonstration.
The concept of Expected Patch Log Likelihood, originally proposed in the concept of Gaussian Mixture Models by Zoran and Weiss, can be deployed while exploiting a sparse prior model for image restoration. This demo implements the resulting denoising algorithm, resulting in comparable performance to that of EPLL-GMM but with automatically determined parameters, and cleaner dictionary.
Implementation of the Joint K-SVD image denoising algorithm that deploys the popular K-SVD approach on the wavelet coefficients of images, resulting with effectively larger and multi-scale denoised patches. This significantly improves the resulting PSNR (but perhaps more importantly the SSIM) of the denoised output for large images with respect to the single-scale version.
