Vishal Krishna

• Used deep convolution network and few other algorithms to generate set of features
• Applied Random forests and SVM respectively creating hierarchical classifier
• Fine tuned custom ensemble model
• Secured rank 9 (out of 644) with algorithm accuracy of 96.4%

Segmentation of wound area from medical images
• Implemented Ng-Jordan-Weiss algorithm for spectral clustering. Created similarity matrix using gaussian distance function, computed unnormalized graph Laplacian and eigenvalues, applied k-means clustering to extract ROI, used morphology for post-processing.
• Proved higher segmentation accuracy of spectral clustering on chronic wound images than simple k-means, contour based segmentation, histogram based thresh-holding and snake based… Show more

Segmentation of wound area from medical images
• Implemented Ng-Jordan-Weiss algorithm for spectral clustering. Created similarity matrix using gaussian distance function, computed unnormalized graph Laplacian and eigenvalues, applied k-means clustering to extract ROI, used morphology for post-processing.
• Proved higher segmentation accuracy of spectral clustering on chronic wound images than simple k-means, contour based segmentation, histogram based thresh-holding and snake based segmentation Show less

Classifier to classify tweets from twitter.com into one of two categories - Sports | Politics
• Pre-processing text – Removed redundant information from tweets
• Feature extraction - Used combination of bag of words and 8F features, and converted it into feature vector
• Classification - Naïve Bayes using Weka toolkit, tuned threshold using cross-validation output
• Rank 15 (out of 118), algorithm accuracy of 92.077%

• Extracted 60 different textural features using various proven methods
• Used weka for visualization and selected best attributes
• Applied different machine learning algorithms for comparison of accuracy
• Adjusted parameters for optimal performance using grid search in Python.
• Plotted bias-variance trade-o?ff using cross-validation, got highest accuracy using support vector machines with polynomial kernel
• Classified with 90.7% accuracy on test set

Project category:- Machine Learning
Project Challenge:-
To help Financial Services organizations acquire more customers by cross-selling : Selling additional products to existing customers. Had to build a model to predict the likelihood of an existing customer accepting an offer of a Term Deposit.
Implementation:-
Trained using Neural Networks.
Software/ Programming Languages Used:- C++ ,Octave(similar to MATLAB)
Duration:- Mid July 2012 to Starting week of August… Show more

Project category:- Machine Learning
Project Challenge:-
To help Financial Services organizations acquire more customers by cross-selling : Selling additional products to existing customers. Had to build a model to predict the likelihood of an existing customer accepting an offer of a Term Deposit.
Implementation:-
Trained using Neural Networks.
Software/ Programming Languages Used:- C++ ,Octave(similar to MATLAB)
Duration:- Mid July 2012 to Starting week of August 2012
Result:- 23rd position out of 133 teams competing in the project competition. Show less

Project Category:- Machine learning and graphs
Project Challenge:- Recommend missing links in a social network. Participants were presented with an external anonymized, directed social graph from which some of edges had been deleted, and were asked to make ranked predictions for each user in the test set of which other users they would want to follow.
Implementation:-
Implemented the Page rank algorithm to get the highest probability of users that one may follow.
Software/… Show more

Project Category:- Machine learning and graphs
Project Challenge:- Recommend missing links in a social network. Participants were presented with an external anonymized, directed social graph from which some of edges had been deleted, and were asked to make ranked predictions for each user in the test set of which other users they would want to follow.
Implementation:-
Implemented the Page rank algorithm to get the highest probability of users that one may follow.
Software/ Languages Used:- C++ (Code::blocks IDE)
Duration:- Starting June 2012 to Mid July 2012
Result:- Ranked 132 out of 424 teams with accuracy of 0.69267.
Highest accuracy of 0.72985 was achieved in the project competition. Show less