Ben Kizaric Masters Student, Eigenvector Enthusiast

I'm a masters student studying Electrical and Computer Engineering with a focus on Machine Learning.

I have full time experience as a software engineer and later as a data analyst at: Northwestern Mutual, the MKE Tech Hub Coaltion, and iit-SourceTech

My research interests focus on using upsupervised techniques in conjunction with supervised methods (both Neural and Classical) to improve performance under difficult constraints, like missing data and limited compute resources. My advisor is Daniel Pimentel-Alarcón.

Publications

January 2022 - May 2022

We propose an ensemble approach to classification with missing values in machine learning. The proposed approach does not rely on imputing missing values before performing inference on the filled dataset, but instead uses a novel subspace clustering algorithm to characterize an input space as a hierarchy of localized affine subspaces, each of which produces a low-dimensional embedding of samples with missing data. The experts are trained on the associated embeddings, and final predictions are a weighted sum of all expert predictions. The proposed method achieves state-of-the-art performance in experiments conducted on various reference datasets.


Taking the convolutions out of convolutional neural networks: Initialization is All You Need.
October 2023 - Awaiting Peer Review

We present a technique to accelerate the training of convolutional neural networks (CNNs) by cheaply introducing constrained supervision into layer-by-layer unsupervised initialization techniques. The proposed method uses a meta model to learn a linear transformation of filters initialized using a feature extraction technique, which is trained on a drastically downsized version of a layer's feature maps to enable fast learning. The method demonstrates large improvements in initialization performance, consistent across reference image classification datasets and feature extraction techniques. The proposed approach enables acceptable classification performance with only training the head of CNNs after initializing filters, enabling an order of magnitude faster training than the traditional approach.

Portfolio

January 2022 - May 2022

A collection of artistic works of static and interactive generative graphic design.

More...
ProcessingGraphic DesignTypographyAnimationData VisualizationDimensionality ReductionGreedy Optimization

October 2021 - December 2021

An exploration of a novel alternative to traditional CNN-based semantic segmentation as applied to self-driving cars with the CityScapes dataset.

More...
Computer VisionSemantic SegmentationPythonnumpyUnsupervised LearningDimensionality Reduction

February 2021 - May 2021

An analysis of past LSAT exams to determine the validity of popular strategies for guessing the correct multiple choice answer based on answer metadata like letter (A-E), length in words, and previous answer letter.

More...
Monte Carlo SimulationHypothesis TestingRdplyrggplot2R Markdown

September 2020 - December 2020

A project completed in partnership with the City of Madison to investigate the fairness of tax-assessed home values with respect to racial demographics.

More...
Python, PandasMatplotlibFeature EngineeringRegressionClusteringVirtual-Environments

March 2020 - January 2022

A new image compression format that uses both unsupervised and supervised machine learning techniques to (hopefully) outperform existing methods of image compression.

More...
PythonNumpysklearnMatplotlibCross-ValidationVirtual EnvironmentsKMeans ClusteringGaussian Mixture ModelsJupyter

June 2020 - June 2022

A showcase of metrics tracking Milwaukee's status as a Tech Hub.

More...
AWS CodePipelineAWS Elastic BeanstalkAWS S3NodeJSJavascriptTypescriptVueJSjQueryExpressSQLSQL ServerSelenuim

April 2020

My final project for an introductory Data Science class, an analysis of Horror Movie data

More...
PythonMatplotlibPandasSciKitLearnSelenuim

Swenden
May 2019

A full-stack NodeJS web app with basic social media functionality

More...
NodeJSJavascriptTypescriptVueJSExpressSQLSQL ServerAzureHeroku

Flexcards (Prototype)
November 2018 - March 2019

A Prototype of an advanced study tool for the 2019 innovation competition Transcend Madison

More...
NodeJSElectronTypescript

Lazy Math
May 2018

My final project for AP Calculus and AP CS Principles. For computation & visualization of Derivatives and Slope Feilds

More...
ElectronJavascriptjQuery