I am a Postdoc in Machine Learning and Computer Vision at MultixLab, University of Amsterdam. Previously I did my PhD at Bosch Delta Lab and VIS Lab , University of Amsterdam under the supervision of Prof. Dr. Arnold Smeulders. Before starting my PhD I completed my masters in Electrical Engineering from KAIST, South Korea under the supervision of Prof. Dr. Jong-Hwan Kim at Robotics Intelligence Technology lab. I completed my bachelors in Electrical and Computer Engineering from COMSATS Institute of Information and Technology, Pakistan.
In order to improve transparency, interpretability, and trust in AI systems, as well as to enable users to understand the reasoning behind the network's decisions I am working on providing explanations for deep neural networks in the context of videos.
XAI, CVPR 2023 spotlight presentation is available here.
Our work Hierarchical Explanations for Video Action Recognition will be presented at spotlight session in XAI, CVPR 2023 proceedings.
Our work Impact of Imperfection in Medical Imaging Data on Deep Learning Models published in Medical Physics journal.
Our work Hierarchical Explanations for Video Action Recognition is accepted for presentation at WICV, CVPR 2023 .
Research Publications and Preprints
Sadaf Gulshad, Teng Long, Nanne van Noord. Hierarchical Explanations for Video Action Recognition, Preprint, 2023
Inspired by human cognition system, we leverage hierarchal information and propose HIerarchical Prototype Explainer (HIPE). HIPE enables a reasoning process for video action classification by dissecting the input video frames on multiple levels of the class hierarchy.
Examples of various types of data imperfection : incorrect segmentation ,low contrast, and artifacts.
Ayetullah Mehdi Günes, Ward van Rooij, Sadaf Gulshad, Ben Slotman, Max Dahele, Wilko Verbakel, Impact of imperfection in medical imaging data on deep learning-based segmentation performance:An experimental study using synthesized data, Medical Physics Journal, 2023
This study investigates the influence of data imperfections on the performance of deep learning models for parotid gland segmentation. This was done in a controlled manner by using synthesized data. The insights this study provides may be used to make deep learning models better and more reliable.
Sadaf Gulshad. Explainable Robustness for Visual Classification , Phd Thesis, 2022
In this thesis, we explore the explainable robustness of neural networks for visual classification. We study an essential question for making neural networks deployable in real-world applications: “how to make neural networks explainably robust?”
Sadaf Gulshad, Ivan Sosnovik and Arnold Smeulders. Wiggling Weights to Improve the Robustness of Classifiers, Preprint 2021
While many approaches for robustness train the network by providing augmented data to the network, we aim to integrate perturbations in the network architecture to achieve improved and more general robustness.
Sadaf Gulshad, Ivan Sosnovik and Arnold Smeulders. Built-in Elastic Transformations for Improved Robustness. Preprint 2021
We present elastically-augmented convolutions (EAConv) by parameterizing filters as a combination of fixed elasticallyperturbed bases functions and trainable weights for the purpose of integrating unseen viewpoints in the CNN.
Jeroen F Vranken, Rutger R van de Leur, Deepak K Gupta, Luis E Juarez Orozco, Rutger J Hassink, Pim van der Harst, Pieter A Doevendans, Sadaf Gulshad, René van Es Uncertainty estimation for deep learning-based automated analysis of 12-lead electrocardiograms, European Heart Journal-Digital Health.
This study aims to systematically investigate uncertainty estimation techniques for automated classification of ECGs using DNNs and to gain insight into its utility through a clinical simulation.
Sadaf Gulshad and Arnold Smeulders. Counterfactual Attribute-based Visual Explanations for Classification. International Journal of Multimedia Information Retrieval (IJMIR2021)
In this paper, our aim is to provide human understandable intuitive factual and counterfactual explanations for the decisions of neural networks.
Sadaf Gulshad and Arnold Smeulders. Explaining with Counter Visual Attributes and Examples. International Conference on Multimedia Retrieval (ICMR 2020), ACM
Different from previous work on interpreting decisions using saliency maps, text, or visual patches we propose to use attributes and counter-attributes, and examples and counter-examples as part of the visual explanations. When humans explain visual decisions they tend to do so by providing attributes and examples.
Sadaf Gulshad, Jan Hendrik Metzen, Arnold Smeulders, and Zeynep Akata. Interpreting adversarial examples with attributes. preprint 2019
We propose to enable black-box neural networks to justify their reasoning both for clean and for adversarial examples by leveraging attributes, i.e. visually discriminative properties of objects.
Sadaf Gulshad, Dick Sigmund, and Jong-Hwan Kim. Learning to reproduce stochastic time series using stochastic LSTM. International Joint Conference on Neural Networks (IJCNN), pp. 859-866. IEEE, 2017. Paper
Sadaf Gulshad, and Jong-Hwan Kim. Deep convolutional and recurrent writer. International Joint Conference on Neural Networks (IJCNN), pp. 2836-2842. IEEE, 2017 Paper
Teaching and Graduate Student Supervision
I was teaching assistant for "Applied Machine Learning" course in 2017, and "Machine Learning" in 2018 and 2019 at the University of Amsterdam.
I was also a teaching assistant for "Multimedia Analytics" course in 2022, and 2023 at the University of Amsterdam.
Supervised Andrew Pizzuto on “Goal Detection using Spatio-Temporal Data.” in collaboration with Beyond Sports B.V.
Supervised Ella Cordus on “Analysing the distribution of green roofs in Amsterdam: The relationship between the distribution of green roofs and spacial injustice.” in collaboration with the municipality of Amsterdam.
Supervised Arend van Dormalen for his masters thesis entitled "Image-Level Supervised Semantic Segmentation with Network Attention and Saliency Priors." in collaboration with the municipality of Amsterdam.
Supervised Jeroen Vranken for his masters thesis entitled "Systematic Comparison of Uncertainty Estimation Methods for Diagnosing Heart Disease in Electrocardiograms using Deep Learning." in collaboration with the The University Medical Center Utrecht.
Supervised Mehdi Güneş for his masters thesis entitled "Exploring the effect of data imperfections in clinical data on model performance." in collaboration with VU University Medical Center Amsterdam.
Supervised Bella Nicholson for her masters thesis entitled "Interpretable Representation Learning for Relational Data." in collaboration with Crunchr.