AI/ML Roadmap
Vision Statement
My objective is to build research-grade expertise in efficient deep learning and computer vision, grounded in an Electrical and Electronics Engineering foundation. The focus is on model compression, object detection architectures, and applied AI research — producing work that is both theoretically rigorous and deployment-ready.
Progress is measured through production-grade projects, ablation studies, and peer-reviewed quality technical deep dives.
Phase I: Foundational Mastery ✓
Goal: Solidify core ML principles and the mathematical foundations of deep learning.
| Focus Area | Core Concepts | Key Output |
|---|---|---|
| Applied Mathematics | Linear algebra for DL, multivariable calculus, optimization theory | Deep Dive: The Calculus of Loss Functions and Backpropagation |
| Core ML & Statistics | Supervised/unsupervised learning, probability, Bayesian methods | IBM AI Engineering Professional Certificate (Coursera) |
| Algorithms | Complexity, data structures, dynamic programming | Stanford Algorithms Specialization (Coursera) |
| Python & Tools | NumPy, PyTorch basics, Git | Version-controlled project repositories |
Phase II: Deep Learning Specialization ✓ ✓ 🔄
Goal: Master CV architectures and frameworks required for research-level work.
Computer Vision
| Focus Area | Architectures & Techniques | Key Output |
|---|---|---|
| CNN Mastery | Convolutional layers, transfer learning (ResNet, VGG) | Deep Dive: Weight Initialization Schemes in CNNs |
| Object Detection | YOLO, R-CNN, RT-DETR, DETR | Project: Real-Time Object Detection (GitHub) |
| Vision Transformers | ViT, self-attention, cross-attention | Ongoing — RT-DETR architecture study |
Systems & Acceleration
| Focus Area | Techniques | Key Output |
|---|---|---|
| CUDA Programming | Parallel computing, memory management, kernel optimization | NVIDIA: Getting Started with Accelerated Computing in CUDA C/C++ |
| Embedded AI | Real-time inference, edge constraints | Signal Processing Toolbox (C++ / Qt) |
Phase III: Applied AI Research 🔬 Current
Goal: Contribute original work at the intersection of model efficiency and computer vision.
| Focus Area | Research Direction | Key Output |
|---|---|---|
| Knowledge Distillation | Logit-KD vs Feature-KD for object detectors, teacher-student dynamics | CS229 Project: RT-DETR Distillation — 12-config ablation on COCO |
| Model Compression | Pruning, quantization, architecture search | Ongoing |
| Representation Learning | Feature alignment, intermediate supervision | CS229 theoretical analysis |
| Benchmarking & MLOps | FPS benchmarking, VRAM optimization, reproducibility | RTX 3050 → Colab → Kaggle GPU pipeline |
Phase IV: Research Publication & Positioning 📄 Next
Goal: Translate applied research into publishable work and establish a research identity.
| Focus Area | Direction | Key Output |
|---|---|---|
| Technical Writing | Conference-style papers, ablation reporting | CS229 final report → arXiv pre-print |
| Open Source | Well-documented repos, reproducible baselines | Public RT-DETR distillation codebase |
| Research Positioning | Applied AI Scientist roles, research engineer track | Portfolio: umutonuryasar.com + GitHub |
Last updated: March 2026