Harleen Hanspal

PhD student (Nov 2021 - Present)

Imperial College London

Biography

I am a PhD candidate in the Verification of Autonomous Systems group led by Prof. Alessio Lomuscio in the Department of Computing, Imperial College London. I am working on verifying neural networks against semantic changes in image domain using generative modelling. I am also interested in defining and verifying semantic properties in language domain, multiagent social settings, etc.

Before joining Imperial, I was working at Daedalean AG, the coolest place to be at for working towards developing certifiable autonomous autopilot for manned aerial vehicles.

Interests

Verification
Generative modelling and Representation learning
Distributed systems

Education (Overview of past projects)

Master in Robotics, Systems and Controls, 2017 - 2020

ETH Zurich

Theses:
- Provable robustness and analysis of Neural Networks focussing on scaling applications of inexact verification methods
- Distributed stochastic optimization under data-driven setting
Bachelor in Electrical Engineering, 2013 - 2017

Indian Institute of Technology Bhubaneswar

Thesis: Manipulators on Micro Aerial Vehicles

Publications

Efficient verification of NNs for LVMs-based Specifications, CVPR 2023 (Aug 2022-Mar 2023, PhD research, SAIL ICL)
- developed an approach that uses Latent Variable Models to verify NNs' robustness against diverse, data-dependent semantic perturbation sets with improved efficiency and effectiveness (paper link, explainer link),
- involved conditional training for interpretable latent representation learning and controlled generation.

Notes

I like making consolidated diagrams on the topics I study, below are some of them (will add sources & improve them over time, till then I hope one finds some flow in them!)

Transformers, RNN & things in between... (YT video)

ML Architectures

OS

Sockets

Terminals

UNIX

Docker

Orchestration

Pytorch

ML frameworks

Useful resources

PCB design: Rick Hartley's lectures, Robert Feranec's Youtube channel
General data science: Brian Keng's blog (I personally loved his posts on Generative Modelling, Variational methods, Sampling, EM algorithm)
Optimization: Optimization Algorithms playlist by Prof. Constantine Caramanis
ML implementation:
- CUDA: Introduction to CUDA blog series by Nvidia
- PyTorch: Deep learning with PyTorch (Chapter 3 has useful information on tensor operations, which ones involve a data copy, which are inplace, etc. which could with microefficiency)