Research/Blog
Meeting Minutes from Saturday 18 May AI Lab meetup in Bangalore
- May 21, 2019
- Posted by: CellStrat Editor
- Category: Uncategorized
CellStratAILab #disrupt4.0 #WeCreateAISuperstars
The CellStrat AI Lab meetup from last Saturday saw some rich presentations by Lab members.
First Vivek started with a detailed discussion on Linear Regression and Polynomial Regression. I explained how we train regression models using either the Normal Equation Method or the Gradient Descent algorithm. In case of Normal Equation method, one can set the slope of Loss Curve (wrt to weights) as zero and solve for ideal weights. In case of Gradient Descent, we adjust the weights iteratively till the loss minimum is found. Gradient Descent is one of the most critical algorithms and forms the basis for bulk of the ML algorithms as well as Artificial Neural Networks.
Next came a superb presentation by Shubha M. (Team Lead for Financial Svcs AI focus group) on Credit Card Fraud Detection using AutoEncoders and Restricted Boltzmann Machines (RBMs). Both AutoEncoders and RBMs are unsupervised learning techniques that help discover latent features in complex datasets. AutoEncoders work by a scheme of encoder-decoder architecture where the encoder encodes the input to a hidden context which is then decoded to reconstruct the input. In this process of reconstruction, the hidden layer is trained to discover the most important features of the input. In case of Credit Card Fraud, this hidden context would represent the most salient features of fraud data which can then be used for future fraud detection.
RBMs are an alternating set of visible and hidden neural layers which pass the signal back and forth, in the process learning the most salient features of data. Both AutoEncoders and RBMs do well in predicting Credit Card Fraud as measured via the AUC curve and Confusion Matrix.
Finally Shreyas J. (team lead for Driverless Cars and Reinforcement Learning focus groups) presented an advanced algorithm for image segmentation called the ESPNet. ESPNet does point-wise convolution after standard convolution layers. After this, it captures image context at varied spatial resolution scales using dilated convolutions and then concatenates them systematically. Such ESPNet modules provide better image segmentation results for CityScapes, Mapillary as well as ImageNet dataset.
Stay tuned for more rich presentations from CellStrat AI Lab members as well as advanced AI research and development projects in the coming weeks.