Shakib Hosseinzadeh
Machine Learning Researcher
ML researcher in training, passionate about solving real-world problems through AI. My work centers on self-supervised learning for medical imaging, particularly representation learning and generative diffusion-based methods in computational histopathology. I enjoy competitive deep learning and building things that matter.
Education
Amirkabir University of Technology — Tehran Polytechnic
Thesis Self-Supervised Learning of Robust Representations for Histopathology Images under Domain Shift
A three-stage framework that learns global semantic embeddings via contrastive objectives, conditions a latent diffusion process on those representations, then adapts toward segmentation through task-specific fine-tuning. Validated on MoNuSeg: AJI 0.6569 · F1 0.8005, surpassing several supervised baselines.
University of Science and Technology of Mazandaran
Thesis Improving Detection of Very Small Objects in Images
Analyzed the information bottleneck induced by high-to-low-dimensional projections in deep architectures and designed a network to preserve fine-grained spatial information critical for small-object representation. Achieved a +12% recall improvement in extremely small object regimes.
Experience
Teaching Assistant
Selected Awards & Competitions
Multi-stage challenge across 7 heterogeneous tasks including dental segmentation on high-noise imagery, zero-shot anomaly detection, and composed image retrieval — under strict VRAM and inference latency constraints.
Noise-robust dual-stream model using Mel-spectrograms and raw-waveform transformer latents for sleep-apnea segmentation from high-noise audio data.
Volunteering
Led workshops on edge-friendly vision models: CLIP-based encoders, lightweight diffusion variants, pruning, quantization, and distillation — end-to-end demos on limited hardware.
Technical operations — managed infrastructure, equipment, and session scheduling.
Projects
Medical Image Captioning
Fine-tuned BLIP2-OPT-6.7B on the ROCOv2 radiology dataset using Parameter-Efficient Fine-Tuning (PEFT) to generate medical captions under computational constraints.
ViT with Masked Autoencoders
Bone fracture detection using MAE (facebook/vit-mae-base) for image reconstruction, then fine-tuned for X-ray classification via transfer learning.
Satellite Image Analysis
Comparative study of YOLOv5–v9 on annotated satellite imagery using Ultralytics, exploring architectural modifications for improved multi-vehicle detection.
Technical Skills