Ibrahim Sobh - PhD’s Post

View profile for Ibrahim Sobh - PhD, graphic

🎓 Senior Expert of Artificial Intelligence, Valeo Group | LinkedIn Top Voice | Machine Learning | Deep Learning | Data Science | Computer Vision | NLP | Developer | Researcher | Lecturer

⚡ 𝗵𝗼𝘄 𝘄𝗲 𝗰𝗮𝗻 𝗺𝗮𝗸𝗲 𝗹𝗮𝗿𝗴𝗲 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗺𝗼𝗱𝗲𝗹𝘀 (𝗟𝗟𝗠𝘀) 𝗺𝗼𝗿𝗲 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗮𝗻𝗱 𝗰𝗼𝘀𝘁-𝗲𝗳𝗳𝗲𝗰𝘁𝗶𝘃𝗲? This recent study from Google DeepMind introduces an innovative approach called 𝗥𝗲𝗹𝗮𝘅𝗲𝗱 𝗥𝗲𝗰𝘂𝗿𝘀𝗶𝘃𝗲 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗲𝗿𝘀, which reduces the size of #LLMs through parameter sharing across layers while maintaining performance. 🟢 LLMs are expensive to deploy, but parameter sharing can reduce their size and cost. 🟢 Recursive Transformers share parameters across layers, with minimal performance loss. 🟢 The study introduces Relaxed Recursive Transformers, which add flexibility to parameter sharing via low-rank adaptation (#LoRA) modules. 🟢 These models outperform similar-sized vanilla models and can recover most of the performance of full-size models. 🟢 A new inference paradigm, Continuous Depth-wise Batching, can lead to significant gains in inference throughput. ✨ I find this research incredibly promising. By making LLMs more efficient, we can deploy them more widely and effectively, driving innovations across various industries. 👉 Paper https://lnkd.in/dvKZxgBR #AI #MachineLearning #Innovation #TechLeaders

  • diagram

To view or add a comment, sign in

Explore topics