Aakanksha Bapna
Bio
I hold a B.Tech in IT and transitioned into research, leading me to pursue an MS by Research with a specialization in Machine Learning. This program was akin to a mini-PhD, during which I defended my thesis and published two research papers. Post-graduation, I joined TCS Research Labs as an ML Researcher for 2.5 years, publishing four additional papers. Seeking to create impactful products, I moved to Cybernetyx, where I developed end-to-end ML systems for Ed-tech, specifically visual intelligence-based classroom products. After a COVID-related layoff, I joined Razorpay as a Data Scientist, enhancing the checkout experience with solutions like address deduplication and RTO prediction services. I was promoted to Senior Data Scientist and, after maternity leave, resumed work in October 2023.
I like solving machine learning and computer vision problems. Currently I'm working as a Senior Data Scientist at Razorpay. Prior to that I waMy first role was a Research Engineer at TCS Research and Innovation labs. I finished my MS by research specialising in machine learning from IIIT, Bangalore in 2016. I also did summer internship in a deep learning based start-up named Artifacia. I did my Bachelor's in Information technology from IET DAVV, Indore.
Projects
Towards an Automated Home Interior Designer System (M.S. Thesis)
We developed an almost completely automated system that produces realistic and aesthetically appealing interior designs for homes. It generates multiple design options for an empty room. The relationships between different elements of a room and items placed in the room are represented as Bayesian networks. The causal relationships defining the network structure are derived from standard thumbrules of interior designing. The parameters for every node in the network are learnt from information extracted semi-automatically from the top view images of furnished living rooms. New layouts based on user inputs are generated upon inferencing from this learnt network.
Publication : Aakanksha Bapna, G. Srinivisaraghavan, “Towards an Automated Home Interior Designer System”. 18th International conference on Artificial Intelligence, Las Vegas, USA, 2016.
Learning parameters for Mixed Bayesian networks ( MS, IIITB)
This algorithm is an add-on over the famous Bayes Net Toolbox in Matlab. A rather non-trivial extension of present parameter learning algorithms which exhaustively models these properties: (i) implements Bayesian parameter learning paradigm, (ii) works for any configuration of discrete and continuous nodes, (iii) handles multivariate continuous nodes. More details can be found in our paper .
Publication : Aakanksha Bapna, G. Srinivisaraghavan, “On learning Mixed Bayesian Networks”. 15th International conference on Machine Learning and Cybernetics (IEEE), Jeju Island, South Korea, 2016.
An example of a mixed Bayesian network. Circles represent
continuous nodes and squares represent discrete nodes.
My system can learn parameters of the conditional distribution
of every node in such a mixed network.
Complete the Look (Summer Internship, Artifacia)
At Artifacia, a visual intelligence based start-up, I built a tool called ‘Complete the Look’ that automatically creates collections of fashion items (top-wear, bottom-wear, bags and footwear), which go well together. A user browsing top or bottom wear will be recommended all other complementary items. It was modeled using Bayesian Networks. I also created classifiers to identify type, fit and length of clothing items using high-level DL features and SVM.
Recognizing users of mobile based on accelerometer data (IIITB)
It was a competition on kaggle.com. Given 60 million unique samples of accelerometer data collected from 387 users, our goal was to use the accelerometer readings as a biometric to identify a person. We employed different machine learning algorithms like Logistic Regression, Neural Networks, SVM etc. to learn and predict the mobile phone user.
Performed Electrical Load Forecasting for office buildings. We used NARX neural networks and SVR for prediction of power consumption at 15 min granularity day ahead, month ahead and quarter ahead. Employed Dynamic factor graphs for day ahead hourly prediction. We attained more than 85% accuracy for all experiments. Paper submitted to ACM e-Engery.
I also worked out Electrical Load Disaggregation for REDD homes Dataset using Bayesian hidden semi-Markov models. Attained an accuracy of nearly 80% in disaggregating loads such as Refrigerator, Microwave and Lighting.
Machine Health Monitoring (TCS R&I)
Currently building a system which would automatically detect degradation in the health of CNC machines. We were provided the data for 4 different kinds of CNC machines which manufacture many different kinds of components. There exists a distinct set of operations for machining every kind of component. We employed different machine learning techniques like SVM, NN and Random Forest for classifying normal and abnormal operations. I used LSTM networks to detect wearing of tool used for a particular operation.
