Download as PDF

Preslav Aleksandrov

[email protected]
+44-7401-005612
linkedin.com/in/preslav-aleksandrov
https://github.com/Helios113
preslavaleksandrov.com
Google Scholar

Abstract

My research focuses on creating new foundational model architectures which are more parameter and data efficient, thus they require much less compute for training and inference The reduced parameter size leads to a higher throughput and lower latency and enables more capable edge model deployments.

I specifically work on dynamic iterative transformer models which allow test time compute scaling and circuit composition, as I believe that future of foundational models, similar to biological brains, will have a dynamic compute path.

Keywords: Machine Learning; Deep Learning; Natural Language Processing; Computer Vision; Reinforcement Learning; AI Safety; Large Language Models; Generative AI; Multimodal Learning

Education

University of Cambridge

2024 -- Current
PhD in Computer Science | Cambridge, UK
  • Novel Foundational Architectures: Deep Learning, Machine Learning, Natural Language Processing, Computer Vision, Reinforcement Learning, Probabilistic Graphical Models, Advanced Algorithms, Statistical Learning Theory
  • Foundational Model Training: Dean's List (all semesters), Research Excellence Award, AI Research Fellowship
  • Distributed Optimisation of Foundational models: Dean's List (all semesters), Research Excellence Award, AI Research Fellowship

University of Glasgow

2024 - Current
MSc in Computer Science, GPA: 3.9/4.0 | Cambridge, UK
  • Relevant Coursework: Deep Learning, Machine Learning, Natural Language Processing, Computer Vision, Reinforcement Learning, Probabilistic Graphical Models, Advanced Algorithms, Statistical Learning Theory

Research Experience

Research Institution/Lab Name

Month Year -- Present
Research Intern/Assistant | Advisor: Prof. Name | City, State/Country
  • Developed novel transformer architecture achieving 15\% improvement in BLEU score on WMT translation benchmark, reducing model parameters by 30\% while maintaining performance
  • Implemented and optimized distributed training pipeline using PyTorch and DeepSpeed, scaling to 128 GPUs with 85\% efficiency, reducing training time from 2 weeks to 3 days
  • Published first-author paper at NeurIPS/ICML/ICLR (specify conference) on efficient attention mechanisms for long-context language models with 100+ citations

Previous Research Lab/Company

Month Year -- Month Year
Machine Learning Research Intern | Mentor: Dr. Name | City, State/Country
  • Designed and trained multimodal vision-language model achieving state-of-the-art results on 3 benchmark datasets (VQA, COCO Captioning, Visual Reasoning)
  • Conducted ablation studies analyzing attention patterns in vision transformers, identifying key architectural improvements adopted by 50+ follow-up papers

Publications

  • Your Name, Co-author Names. ``Descriptive Title About Your Research Contribution.'' NeurIPS/ICML/ICLR/AAAI, Year. [arXiv] [code]

Technical Projects

Project Name: Descriptive AI/ML Focus

Month Year -- Month Year
  • Built end-to-end deep learning system for [specific task] using PyTorch/TensorFlow, achieving [specific metric] performance
  • Implemented novel [algorithm/architecture] based on recent research, demonstrating [percentage] improvement over baseline
  • Open-sourced implementation with comprehensive documentation, gaining 500+ GitHub stars and used by 10+ research groups

Another Significant Project Name

Month Year -- Month Year
  • Developed [specific AI system] processing [data scale] with [performance metric], deployed in production serving 10,000+ users
  • Optimized inference latency from [X]ms to [Y]ms through model quantization, pruning, and efficient deployment strategies

Technical Skills

Awards & Honors

  • Best Paper Award, [Conference Name], Year -- Recognized for outstanding research contribution
  • Research Fellowship, [Institution], Year -- Competitive fellowship awarded to top 5\% of applicants
  • Kaggle Competition Winner/Grandmaster, [Competition Name], Year -- Ranked 1st/Top 1\% among 5,000+ teams

Teaching & Mentorship

Teaching Assistant -- Course Name (ML/DL/AI)

Month Year -- Month Year
  • Led weekly discussion sections for 50+ students, achieving 4.8/5.0 teaching evaluation
  • Developed course materials and programming assignments on neural networks, CNNs, RNNs, and transformers

Service & Leadership

  • Reviewer for [Conference Names: NeurIPS, ICML, ICLR, etc.], Year
  • Organizer, AI Research Reading Group at University, Year -- Present