Jale Dinler
ac0sga@r.postjobfree.com 608-***-**-** Palo Alto, CA
http://www.linkedin.com/in/jaledinler
jdinler (skype)
EDUCATION
UNIVERSITYOFWISCONSIN
MADISON
MA. IN MATHEMATICS
**** - **** *****on, WI
KOCUNIVERSITY
MSC. IN MATHEMATICS
2010 - 2013 Istanbul/Turkey
BOGAZICI UNIVERSITY
BSC. IN COMPUTER ENGINEERING
2004 - 2009 Istanbul/Turkey
LINKS
Github:// jaledinler
LinkedIn:// jaledinler
COURSEWORK
GRADUATE
Building DeepNeural Networks
Data Science
Machine Learning
Nonlinear Optimization
Integer Optimization
SKILLS
•Java •Python •Matlab
•C •Tensorflow •Weka
•Scikit •Google Cloud
•Magellan
AWARDS
2016 Honored Instructor Award
at UW-Madison
2013 Full Scholarship from
UW-Madison
2010 Full Scholarship from
Koc University
2004 337th/1.2million students in
National University Entrance
Exam
PROJECTS
EXTRACTING STRUCTURED INFORMATIONFROMRAWTEXT
DATA An information extractor to extract all person names from300 Pitchfork Reviews that is collected from a dataset on Kaggle was imple- mented. Scikit learn package was used to create different MLmodels by using classifiers (e.g. SVM, RandomForest, Decision Tree, Logistic Regression, Linear Regression). After applying these classifiers with cross validation to train data, the classifier that gives the highest preci- sion was selected to be applied to test data.
NEURAL NETWORK (NN) I appliedNNwith one layer of hidden units to predict protein secondary structure. One of the project require- ments was to implement NN library from scratch. I explored the effect of number of hidden units and number of epochs on accuracy and over- fitting. I also experimentedwithmomentum termand weight decay to see their effects on the convergence rate. To handle nominal features, after exploring possible encoding strategies, I decided to use one-of-k encoding.
ENTITYMATCHING (EM) Magellan (an EM tool developed at UW- Madison) is used for the task. Two tables are extracted fromPitch- fork Reviews Data (collected fromKaggle) andDiscogs albumdata. To obtain a set of candidate tuple pairs, several blockers were used. A small sample of these tuples is labeled (match, no-match) to train and test matchers like random forest, svm, naive bayes, logistic regression, linear regression and decision tree. The best matcher (99.4%preci- sion, 99.3% recall) was applied to find all possiblematches and these matches are combined into a single table.
NAIVE BAYES (NB) ANDTREE AUGMENTEDNAIVE BAYES (TAN) I appliedNB and TAN classifier on a lymphography domain. One of the project requirements was to implement these libraries from scratch. To find themaximal spanning tree in TAN, Prim’s algorithmwas implemen- ted. To analyze the statistical significance of the difference between NB and TAN, Stratified 10 fold cross validation and then a paired t-test was used.
DECISIONTREE (DT) I appliedDT to predict whether a person has diabetes or not. One of the project requirements was to implement an ID3-like decision tree from scratch. For numeric feature splits, a very similar algorithm to the one in C4.5was implemented. To avoid over- fitting, a threshold for the size of node is set before starting to grow the tree. Change in accuracy was analyzed for different values of this threshold.
K-NEARESTNEIGHBOR (K-NN) ANDK-NNSELECT I applied k-NN learner to analyze the quality of wine. One of the project requirements was to implement k-NN from scratch. For k-NN select, leave-one-out cross validationwas implemented to select the value of k. The one that results in theminimal cross-validated error was selected. EXPERIENCES
2017 - Research Engineer (Intern) at Docomo Innovations, Inc. 2013 - 2017 Teaching Assistant at University ofWisconsin - Madison 2011 - 2013 Teaching Assistant at Koc University