Atam Oguz Erkara | Personal Portfolio

Publications

Blind Confusion of Classification Networks

A black box evaluation under common and structured image corruptions

This work presents a black box robustness evaluation of image classification networks under common and structured image corruptions. It analyzes 37 models across 15 corruption types and introduces Accuracy Confidence Divergence as a diagnostic measure for comparing accuracy degradation with confidence behavior.

Atam O. Erkara, Markus Mayer, Blind confusion of classification networks: A black box evaluation under common and structured image corruptions, Neurocomputing, Volume 686, 2026, 133678, ISSN 0925-2312.

https://doi.org/10.1016/j.neucom.2026.133678

Read Paper →View Code →

Projects

PPO and DDPG Algorithms

PPO experiments in LunarLander with a focus on training speed and reward stability. Compared DummyVecEnv vs SubprocVecEnv, tested entropy tuning, gravity variations, and stacked observations to quantify what actually improves learning.

HandsON Music App

Real time hand gesture control for music playback using MediaPipe hand tracking. The app is complete and includes an in page usage guide for gesture driven control.

Full Stack Web Development

Full stack Next.js project built in TypeScript with React and MongoDB Atlas via Mongoose. Includes responsive UI aligned with TH Deggendorf styling, EN and DE content switching, and role based access for Guest, User, and Admin with admin content management.

Research and Engineering

ML Concept Breakdown

Support Vector Machines

Visual explanation of SVM for classification problems. Covers margin based separation, support vectors, soft and hard margins, kernel trick, and a penguin dataset example.

Open Insight →GitHub Notebook →