Sign In

Sahand Sharifzadeh

Research Scientist at Google DeepMind

Sahand Sharifzadeh is a research scientist at DeepMind with a focus on developing multimodal models. He holds a PhD in Artificial Intelligence from Ludwig-Maximilians-Universität München, where he graduated with the highest distinction. His areas of interest include Artificial General Intelligence (AGI) and the application of AI in fields such as medicine, sports analysis, and robotics. Sahand completed his Master's degree at the Technical University of Munich, specializing in Computer Vision, Artificial Intelligence, and Robotics. His master's thesis explored Deep Inverse Reinforcement Learning for Autonomous Driving under the guidance of Prof. Daniel Cremers. Prior to his PhD, he gained practical experience in various AI domains and pursued research in knowledge graph embedding models and human memory processes. Sahand collaborates extensively with professionals from diverse fields and has made significant contributions to Computer Vision, Knowledge Graph Modeling, and NLP. He has a rich academic lineage tracing back to notable historical figures like Nasir al-Din al-Tusi, Gauss, Euler, and Heisenberg.

Highlights

Jun 19 · linkedin

🚀 Excited to share our latest work at Google DeepMind: "Synth^2: Boosting Visual-Language Models with Synthetic Captions and Image Embeddings" together with Christos Kaplanis, Shreya Pathak, Dharshan Kumaran, Anastasija Ilic, Jovana Mitrovic, Charles Blundell & Andrea Banino!

#VLM #Synthetic #AI 1/ 🔍 Tackling the bottleneck of high-quality human-labeled datasets for Visual-Language Models (VLMs), we propose a novel approach using Large Language Models (LLMs) and image generation models to create synthetic image-text pairs.

2/ 🖼️ Our method employs a pretrained text-to-image model to generate image embeddings from LLM-generated captions. This approach expands beyond the original dataset, creating novel compositions that enrich VLM training data.

3/ 📊 Extensive experiments show our VLM, finetuned on synthetic data, performs comparably to models trained on human-annotated data, but with significantly less data! This demonstrates the power and efficiency of our synthetic approach.

4/ 🔄 We found that semantic diversity and balance in captions are crucial for better downstream performance. Our analysis provides new insights into optimizing synthetic data for VLM training.

5/ ⚡ Synthesizing images in the embedding space is 25% faster than in the pixel space, reducing memory overhead and resource consumption without compromising data quality.

6/ 🌍 Our work opens up promising avenues for developing self-improving multimodal models, addressing data scarcity, high curation costs, and noise in traditional datasets.

7/ 🔬 By pre-training both the text-to-image model and VLM on the same dataset, we isolate the true benefits of synthetic images, ensuring improvements are due to our method, not pre-existing knowledge.

8/ 🎓 Key Contributions:    •Fully synthetic high-quality text-image pairs for VLM training   •Efficient image embedding generation   •Controlled study for fair evaluation  •Significant performance gains in scene description, QA, and external knowledge QA

9/ 🌟 We believe this research marks a significant step toward the future of VLM training, with customized datasets, open-ended training, and self-improving models.

Read the full paper here:  https://lnkd.in/e-saTKEX #DeepMind #AIResearch

Jun 11 · linkedin

🎉 Excited to share our latest work published in Nature Communications! 🚀 We've developed a new approach pushing the boundaries of particle physics full detector simulation. Big thanks to Baran Hashemi & our amazing team at Ludwig-Maximilians-Universität München . Check it out here:

Sahand Sharifzadeh
Add to my network

Location

London Area, United Kingdom