Sign In

Mikhail Burtsev

Founder and Leader at DeepPavlov.ai

Mikhail Burtsev is a seasoned professional with a profound expertise as the Head of Deep Learning research initiatives in various research establishments.

He holds a Doctor of Philosophy (Ph.D.) with a specialization in Artificial Intelligence. Mikhail's skills encompass Neural Networks, Mathematical Modeling, Python programming, Computer Science, Cognitive Science, and Genetic Algorithms.

His educational background includes pursuing a Ph.D. in Computer Science from the Keldysh Institute of Applied Mathematics, an MSc in Microelectronics from Moscow Power Engineering Institute, and studying Complex Systems at the Santa Fe Institute.

Mikhail Burtsev is the Founder and Leader of the pioneering platform DeepPavlov.ai. Previously, he held prominent roles such as the Head of Neural Nets and Deep Learning Lab at Moscow Institute of Physics and Technology (MIPT).

With a proven track record in organizing significant events, Mikhail was the Lead Organizer at The Conversational Intelligence Challenge and the 1st NIPS Conversational Intelligence Challenge. He also conceptualized and led AI school-hackathons at RL.DeepHack.me and QA.DeepHack.me.

His professional journey includes serving as the Vice Chair of the Expert Panel at the European Commission, being a Senior Research Scientist at Kurchatov NBIC-Centre, and holding roles as a Senior Academic Advisor at Wolfson College Cambridge. Additionally, he contributed as a Senior Researcher at the Anokhin Institute of Normal Physiology RAMS and as a Researcher at the Keldysh Institute of Applied Mathematics.

Highlights

Aug 23 · linkedin

🔎📚🪡📚❓ BABILong updates on the new Phi-3.5 models!

The latest Phi-3.5 models have shown solid improvements on the BABILong benchmark, which tests long-context reasoning across distributed facts. Notably, the Phi-3.5-mini-instruct model performs exceptionally well on context lengths up to 32K, surpassing its predecessor, Phi-3-mini-128k-instruct. The Phi-3.5-MoE-instruct also delivers similar performance to Phi-3-medium-128k-instruct, with fewer active parameters (6.6B vs. 14B), and competes closely with LLama-3.1-8B-Instruct. https://lnkd.in/eT_S3wWY

Check out the full BABILong leaderboard for more insights! 🔗 https://lnkd.in/eXdg42N6

#AI #ML #LongContextReasoning #LLM #Benchmarking

Jul 29 · linkedin

🦙Llama vs. 🗿Goliath: Open-Source Leaps Over "Open" AI in long context reasoning!!!

🥇 Llama 3.1 70B is a new leader among LLMs in long-context multi-hop QA. It out-competes much lager GPT-4 on BABLong benchmark.

🌟 Llama 3.1 70B wins by getting advantage at the longer 10K+ context lengths.

🙏 A big thank you to Yann LeCun for his consistent support and promotion of open-source AI.

🔗 Leaderboard: https://lnkd.in/eXdg42N6

BABILong is a novel generative benchmark designed to evaluate NLP models' ability to process arbitrarily long documents with distributed facts. It challenges models to identify key information by "hiding" relevant and distracting facts within the text.

BABILong eval links:

Paper: https://lnkd.in/ezV5Xxzs Benchmark dataset: https://lnkd.in/e-Zxcyp2 GitHub: https://lnkd.in/d2Q5bgsq

#AI #MachineLearning #NLP #OpenSource #Meta #Llama3.1 #Benchmarking

Mikhail Burtsev
Add to my network

Location

Russia