Soham Palande

Hi, I’m Soham. I’m currently a research engineer in J.P. Morgan’s AI Research group.

Research: I study how abstract reasoning and decision-making emerge in large neural networks, and how this parallels the structure of human cognition. My work bridges cognitive science, neuroscience, psychology, and machine learning: I study tasks that demand belief formation, planning, and strategic adaptation, and mechanistically dissect how models—and by analogy, brains—solve them. I use tools from interpretability and ideas from probabilistic cognition to reverse-engineer internal circuits for reasoning, beyond perception. Language and vision form the inputs over which reasoning unfolds, but the core interest lies in how internal models are constructed, updated, and used. Studying these processes in artificial systems offers a new lens on the brain and the self—while insights from neuroscience and psychology can refine how we understand and design the models themselves. My work treats reasoning and higher-level cognition as mechanistic phenomena, and seeks to uncover the machinery behind them to bring us closer to understanding what it means to be human.

Previously: I graduated from Rutgers University-NB majoring in Computer Science and Mathematics. I was an R&D intern at L3Harris where I worked on building anomaly detection models.

news

*not updated

May 10, 2022	I was awarded the Chancellor-Provost’s Research Excellence Award
May 5, 2022	I was awarded the Nicholas Novielli Prize for academic excellence by the Rutgers CS Dept.

selected research

*not updated

Aresty

Understanding Correlated Error Events in Quantum Computers

Arpan Gupta, Michael Schleppy, Soham Palande, and 1 more author

Aresty Research Symposium, Apr 2022

Abs HTML PDF

Quantum Computing is at a critical juncture. Prototype quantum computers are now at a level of reliability and scale where researchers can run small scale algorithms. However, modeling errors in near-term Noisy Intermediate-Scale Quantum (NISQ) devices is necessary to harness their full potential. Existing frameworks such as IBM Qiskit are limited in their capacity to model and simulate complex noise events. We study the use of Probabilistic Graphical Models (PGMs) as a natural abstraction to efficiently simulate and model quantum circuits with no noise, uncorrelated noise (bit-flip and amplitude damping) and correlated noise using complex-valued Bayesian networks. By definition, Bayesian networks are a type of PGM that articulate conditional dependencies between events through the use of Directed Acyclic Graphs (DAGs). We study well-known quantum algorithms such as Deutsch- Jozsa and Simon’s algorithms, and transform their circuit representations into Bayesian network models by extending Python’s pgmpy library to allow for complex- valued networks. By using exact inference algorithms like Variable Elimination, we are able to create valuable inferences that can be converted into density matrices using our extension of pgmpy. We validate that our models produce correct density matrices for noisy Quantum circuits using IBM Qiskit, showing that Bayesian networks are a valid abstraction for Quantum correlated and uncorrelated noise events. The correctness of our results point to new languages and methods for representing Quantum Computations
Aresty

Solar Irradiance Nowcasting Using All-Sky Imager Data

Soham Palande, and Ahmed Aziz

Aresty Research Symposium, Apr 2021

Abs HTML PDF

As we continue to look for ways to accelerate our transition to renewable energy, one of the fundamental obstacles to achieving this remains the unpredictable nature of wind and solar energy. Clouds moving in front of the sun and rapid changes in weather conditions cause fluctuations in the measured solar irradiance and cause serious challenges in power grid operation. In this study, we address the unpredictable nature of solar irradiance by using past irradiance measurements and leverage images captured by an All-Sky imager using a Deep Learning Framework to forecast solar irradiance over the short term. We train the model on datasets of varying sizes and show that the models significantly outperform the smart persistence model and state of the art statistical approaches using optical flow techniques.