Kunal Ghosh

Office:

Room A322 in Computer Science T-Building,
Konemiehentie 2, Otaniemi campus area, Espoo

Postal Address:

Aalto University School of Science,
Department of Information and Computer Science,
P.O. Box 15400, FI-00076 Aalto, Finland

Email:

Kunal dot ghosh aat aalto dot fi

Other:

[Github], [LinkedIn], [Flickr], [CV]

About

Since summer 2018, I am a Doctoral candidate in the Department of Computer Science at Aalto University. I focus on deep generative models (focusing on molecular data). I recived my MSc (Honours) in the Machine Learning and Data Mining (MACADAMIA), also from Aalto University. .

Currently working on:

• Deep generative models for molecular data. Supervised by Prof. Patrick Rinke
• Black-box variational inference supervised by Prof. Aki Vehtari in the Probabilistic Machine Learning Group.

Previously I worked on:

• Predicting Molecular Electronic Properties using Deep Learning.
  For this project, we are collaborating with Prof. Patrick Rinke and members of his Computational Electronic Structure Theory (CEST) group, from the Department of Applied Physics at Aalto University on the European Union (EU) wide NOMAD project.
• Semi-Supervised Learning with extensions of the Variational Autoencoder (VAE).

Before returning to academia in 2015, I worked at Amazon India as a Software Development Engineer in Test.

Research Interests Deep Learning, Probabilistic Graphical Models, Generative Models

Education

Masters

(Honours Program) Machine Learning and Data Mining (Graduated with Honours, November 2017)
Aalto University, Finland.

Bachelors

Computer Science and Engineering (Graduated 2011)
Visvesvaraya Technological University, India

Thesis & Publications

• [New] Deep learning spectroscopy: neural networks for molecular excitation spectra.[url]
  Kunal Ghosh, Annika Stuke, Milica Todorović, Peter Bjørn Jørgensen, Mikkel N. Schmidt, Aki Vehtari, and Patrick Rinke
  Wiley Advanced Science, 2019.
• Chemical diversity in molecular orbital energy predictions with kernel ridge regression[arxiv]
  Annika Stuke, Milica Todorović, Matthias Rupp, Christian Kunkel, Kunal Ghosh, Lauri Himanen, Patrick Rinke
  Submitted, 2019.
• Deep Learning for Predicting Molecular Electronic Properties [Link]
  Kunal Ghosh, Supervisor : Prof. Aki Vehtari (Dept. of Computer Science), Advisor : Prof. Patrick Rinke (Dept. of Applied Physics)
  Masters Thesis, Aalto University, Finland, 2017.
• Comparison of the recognition accuracy of Eigen Faces, when applying different dimensionality reduction algorithms on the input images.
  Kunal Ghosh, M. GuruPrasad, S. Dharini, and J. L. KiranTej
  Bachelor Thesis, Visvesvaraya Technological University, India, 2011.

Teaching

• TA for CS-E4890 Deep Learning Autumn 2016, 2017, 2018 and 2019

Projects

• Implementing parallel coordinate descent for L1-Regularized Loss Minimization.
  In this project we implemented the parallel coordinate descent algorithm as Described in [Bradley et.al 2011] in Apache Spark, (Parallel) Matlab and NumPy. The Spark implementation was deployed the algorithem on Aalto University's internal compute cluster. A detailed report and code can be found on Github. [Code],[Report PDF]
• Experiments with the EM algorithm..
  In this project we derived the update equations of the EM algorithm for a gaussian mixture model and then Implemented the algorithm. The model was then compared with a simple linear model to fit various toy datasets. Finally a detailed report was written with the findings (and the derivations). [Code],[Report PDF]
• Approximating F0 and F2 of streaming data. The project was to implement probabilistic counting algorithms, specifically the Flajolet Martin Algorithm for approximate F0 (unique elements in a sequence) and to implement second moment F2 using the algorithm as described in Alon et.al 2002] useful for computing the Gini coefficient of variation for a data stream.[Code],[Report PDF]
• Comparing the effect on search relevance when using various parameter settings in a Information Retrieval system.
  We implemented an Information Retrieval system in Java using Apache Lucene and compared the Relevance of the Recommendations in a document corpus (relevance (or not) of a document was provided in the test corpus). We experimented with various parameters in the implementation such as the choice of stemmer, text preprocessing etc. We also compared the Vector Space Model (VSM) with TF-IDF weights and BM25 as Similarity measures. [Code],[Report PDF].
• Kah-Vis : Visualizing the effect of Bulk Purchase of Coffee on consumption.
  The objective of this project was to use principles of effective visualization design (using Munzner's Model and also from works of Edward Tufte). I visualized a dataset of coffee consumption over a period of time and tried to see if buying coffee in bulk has had any effect on (my) consumption. The dataset was from my own coffee purchase receipts over a period of 8 months. [Report PDF]