 |
Image Completion via Inference in Deep Generative Models
, Saeid Naderiparizi,
arXiv preprint, 2021
Image completion task from the perpective of amortized inference in image generative models, particularly, variational auto-encoders.
It leads to diverse, realistic image completions even when the vast majority of an image is missing.
|
 |
Uncertainty in Neural Processes
Saeid Naderiparizi, , ,
arXiv preprint, 2020
Achieving calibrated uncertainty in Neural Processes, especially in low data regime, via different architectures and training objectives.
|
 |
Planning as Inference in Epidemiological Models
, , Saeid Naderiparizi, , , ,
, ,
arXiv preprint, 2020
Automating parts of infectious disease-control policy-making via performing inference in existing epidemiological dynamics models.
|
|
Amortized rejection sampling in universal probabilistic programming
Saeid Naderiparizi, , , ,
, , , ,
, , ,
arXiv preprint, 2019
Identifying and addressing the problems caused by rejection sampling loops in inference in universal probabilistic programming.
|
 |
Coping With Simulators That Don't Always Return
, Saeid Naderiparizi,
International Conference on Artificial Intelligence and Statistics (AISTATS), 2020
Model learning in models partially defined through "brittle" deterministic simulators.
A "brittle" simulator is one which would fail to produce results for invalid inputs that might occur over the course of training.
An ODE solver not converging to the required tolerance in the allocated time or solid bodies intersecting in robot simulators are examples of simulator failure.
|
|
Efficient Probabilistic Inference in the Quest for Physics Beyond the Standard Model
, , , ,
Saeid Naderiparizi, , , ,
, , , ,
, ,
Neural Information Processing Systems (NeurIPS), 2019
A novel probabilistic programming framework that couples to the existing large-scale simulators via a cross-platform protocol, allowing automatic inference in such simulators.
|
|
Etalumis: bringing probabilistic programming to scientific simulators at scale
, , , ,
, , ,
Saeid Naderiparizi, , ,
, , ,
, , ,
Proceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis, 2019
|
 |
LaTeq: Converting equation images to LaTeX
Under the supervision of Frank Wood
We trained a generative model for LaTeX equation codes, which will be used as a part of a probabilistic program to perform Bayesian inference for converting equation images to LaTeX codes.
|
|
A bayesian approach to Visual Question Answering
, Saeid Naderiparizi,
Probabilistic Programming course, 2018
Using probabilistic programming and inference compilation to solve visual question answering task.
|
 |
RegNet: Regularizing Deep Networks
(code)
, Saeid Naderiparizi
Multimodal Learning with Vision, Language and Sound course, 2018
We can define regularizers to improve model's performance in the case of not having enough data for few classes.
We tried different regularizers and tested it on image classification task.
|
 |
A Multi-Plane Block-Coordinate Frank-Wolfe Algorithm for Training Latent Structural SVMs
Under the supervision of Christoph H. Lampert
We designed and implemented a method for training Latent Structural SVMs with costly max-oracles using Frank-Wolfe algorithm on the dual objective. The method does not need specifying a learning rate and it gives a definite result.
|
 |
Sequential Bound Optimization
Under the supervision of Sobhan Naderi Parizi
We proposed an iterative procedure for optimizing non-convex objective functions that works by optimizing a sequence of convex bounds on the true objective.
|
 |
Learning to ride a bike
(code)
, Saeid Naderiparizi
Computer Animation course, 2017
We implemented a simulator for bicycle motion and used Fitted Value Iteration algorithm for learning to keep the bike balanced. The next goal is to add a target point which the bike should learn to get there and meanwhile, keep the bike balanced.
|
 |
Learning Influence Probabilities under ICIDT model
(code)
Saeid Naderiparizi, ,
Topics in Data Management - Social and Information Networks course, 2017
We came up with a model for influence propagation in social networks which captures decay in adoption probabilities over time and implemented maximum likelihood estimation in C++ to learn the probabilities under this model. To my knowledge, the problem of learning probabilities in this time decaying model had not been studied before.
|
|
Implementation of a 3D reconstruction system
Saeid Naderiparizi,
Fundamentals of 3D Computer Vision course, 2016
Implemented an algortithm to generate the point cloud of a 3D object using a few photos taken of it. It uses SURF features to find and match keypoints, find position of cameras and then aggregates them to generate the point cloud. This project is done in Matlab.
|
|