Cifre Docotoral Scholar

Department of Mathematics, Centrale-Supélec, Université Paris-Saclay

INRIA OPIS & Schneider Electric

I am working under Dr. Jean-Christophe Pesquet on stability and verification of neural networks used in open-loop and closed-loop control of dynamical systems. As part of my Ph.D., I am also working on modeling the dynamics of heavy electrical motors using neural networks under the guidance of Dr. Marc Castella, Nicolas Henwood, and Al-Kassem Jebai. I am also associated with Granular AI where I work on leveraging satellite imagery and geospatial algorithms for socio-economic problems. In INRIA OPIS, I work on AI for Radiology with Dr. Émilie Chouzenoux and Dr. Nathalie Lassau from Institut Gustave Roussy.

I was a Junior Research Fellow at Computer Vision and Machine Learning Lab (IIIT Delhi) under Dr. Chetan Arora where I worked on egocentric video understanding. I have worked with Dr. Angshul Majumdar on applications of generative models to collaborative filtering and energy management systems. I have also worked with Dr. P.B. Sujit on reinforcement learning for control problems. I completed M. Tech. in 2017 from IIIT-Delhi.

- One paper accepted in IECON 2020
- Recent work on COVID19 detection from chest CTs and clinical data mentioned in Le Monde.
- One paper accepted in AAAI 2020
- Organised two in-class kaggle challenges: ML-F19 for ML-DSBA-AI: Machine Learning and 2EL1730-Machine Learning for Machine Learning courses at CentraleSupelec.
- Granular-ai got selected in Hyperspace Challenge organised by US Air Force and business accelerator ABQid.
- Presented a poster on
*Urban Change Detection*at PRAIRIE AI Summer School 2019, Paris, France. [Poster] - Attending PRAIRIE AI Summer School 2019, Paris, France.
- I started my Ph.D. at École Centrale-Supélec Paris. on 1st October 2019.
- Thanks to the PAISS industrial partners for providing me financial support to attend PRAIRIE AI Summer School 2019, Paris, France.
- I presented our paper
*Detecting Urban Changes with Recurrent Neural Networks from Multitemporal Sentinel-2 Data*at IGARSS 2019, Yokohama, Japan. [PPT] - Volunteering at SIGIR 2019, Paris, France.
- I chaired the
*Deep Reinforcement Learning for Games*session at IJCNN 2019. - I presented our paper
*MAPEL: Multi-Agent Pursuer-Evader Learning using Situation Report*at IJCNN 2019, Budapest, Hungary. [PPT] - Hosted AI Challenge in Centrale-Supélec AI Summer School with Dr. Maria Vakalopoulou and Dr. Fragkiskos D. Malliaros. [PPT]
- Gave presentation on
*Modeling Electrical Motor Dynamics using Neural Networks*at Schneider Toshiba Inverter Europe, Pacy-sur-Eure, France. [PPT]

S. Verma, N. Henwood, M. Castella, AK Jebai, and JC Pesquet. Neural Networks based Speed-Torque Estimators for Induction Motors and Performance Metrics, (Accepted in IECON 2020) [Project Page, PDF, PPT, Video].

Lassau et. al. AI-Based Multi-Modal Integration of Clinical Characteristics, Lab Tests and Chest CTs Improves COVID19 Outcome Prediction of Hospitalized Patients, (Submitted to Nature Communications) [Submitted Version, Code, Article in Le Monde].

S. Verma, N. Henwood, M. Castella, F. Malrait, and JC Pesquet. Modeling Electrical Motor Dynamics using Encoder-Decoder with Recurrent Skip Connection, (Accepted in AAAI 2020) [Project Page, PDF, Poster].

M. Papadomanolaki, S. Verma, M. Vakalopoulou, S. Gupta, and K. Karantzalos. Detecting Urban Changes with Recurrent Neural Networks from Multitemporal Sentinel-2 Data, (Accepted in IGARSS 2019) [PDF, PAISS Poster, code].

S. Verma, R. Verma, and P.B. Sujit. MAPEL: Multi-Agent Pursuer-Evader Learning using Situation Report, (Accepted in IJCNN 2019) [PDF].

S. Verma, S. Singh, and A. Majumdar. Multi Label Restricted Boltzmann Machine for Non-Intrusive Load Monitoring, (Accepted in ICASSP 2019) [PDF].

S. Verma. A Survey on Machine Learning Applied to Dynamic Physical Systems [Project Page, PDF, code].

S. Verma. Action Recognition in Egocentric Videos, (M. Tech. Thesis 2017) [IIITD archive].

Note: I write blogs to learn and practice mathematical concepts, technical writing, and aesthetic presentation. These posts should never be taken as a primary source of information. For a clear and accurate understanding, please visit citations.

Why do bi-temporal data fail in modeling urban sprawl? Introducing a large multi-date change detection dataset. Deriving urban growth index from Sentinel-2 multi-date images. Exploring U-Net, RNN, QRNN to handle multi-date input.

ESA's SAFE and DigitalGlobe's FLAME provide solutions to atmospheric correction and base layer matching problems respectively. Both are proprietary products and require intensive computing resources and are not suitable for processing large maps in limited time. Can GANs solve atmospheric correction, tile blending, and cloud correction problem?

Unsupervised feature learning from satellite images using cross-channel encoder based on split-brain architecture. Feature learning using hyperspectral images, constructing thermal bands from RGB and vice-versa. Normalized cuts on learned features for semantic segmentation.

\begin{eqnarray}
minimize : & \sum_{i\epsilon\{1,...,N\}}T_i \\
s.t. : & T_i \epsilon [a_i,b_i] \\
& |T_i-T_j| \geq T_d \\
& where \quad i,j \epsilon \{1,...,N\} i\neq j \\
\end{eqnarray}
#### World of optimization

Optimal solution using MILP for ATC scheduling and multi-agent track-n-tag game. Using Branch and Bound method to select "equidistant" sentinel-2 capture dates for large scale inference.

\begin{eqnarray}
f(x;\theta) & = W_n\phi(W_{n-1}(...\phi(W_1x))) \\
\forall x,x' & : \frac{||f(x)-f(x')||_2}{||x-x'||_2} \leq ||f||_{Lip} \\
\underset{\theta}{\text{minimize}} : & \sum_{n=1}^{N} \frac{1}{N} ||y_n-f(x_n;\theta)||_2 \\
s.t. : & ||f||_{Lip} \leq 1 \\
\end{eqnarray}
#### Robustness of neural networks

Lipschitz constrained training of DNN for bounded control of electric motors.

Data-driven modeling of non-linear dynamics of electric motors using neural networks.

Activity and action recognition in first-person videos. Two datasets of short and long activities in Plumbing and PC Assembling domain respectively. Baselines of two-stream CNN, I3D, SlowFast, and C3D networks.

Article segmentation and word recognition in Indian regional newspapers.

A multi-agent pursuit-evasion game with partial observability. Situation reports for sparse communication and coordinated reinforcement learning.