[Tensor Learning Team Seminar] - Talk by Prof. Hai Zhao (SJTU) "A Mathematical Theory of AGI and its Brain-inspired LLM Exploration, A BriLLM-MetaPred Implementation" and Prof. Yan

20 ноября 2026, 07:00doorkeeper ↗
1

Япония

Япония

Date and Time: November 20th, 2026, 16:00 -- 18:00 (JST) Venue: Open Space, RIKEN Nihonbashi Office and Online *In-person attendance is limited to AIP researchers. Talk by Prof. Hai Zhao Title: A Mathematical Theory of AGI and its Brain-inspired LLM Exploration, A BriLLM-MetaPred Implementation Abstract: The current research and industry roadmap for general artificial intelligence (AGI) relies too heavily on unsupervised scaling of large language models (LLM), but this approach lacks a solid mathematical foundation. We argue that pure symbolic large language models face insurmountable barriers in achieving AGI. To overcome these limitations, we derive the necessary and sufficient conditions for computable AGI: first, an internally driven compression-based autonomous consistency loop that can converge independently without external input; second, an external physical anchoring mechanism that aligns internal fixed points with the real world through perception-action coupling. Through proof validated across three mathematical derivation paths, AGI is shown to result from the convergence of infinite self-cognitive ability. Additionally, we rigorously show that the MetaPred machine learning paradigm, which supports infinite recursive metacognition, exists and is unique in an isomorphic sense, and this mechanism directly corresponds to the semantic mapping and neural oscillation processes in the biological cortex. For practical application, we propose the Signal Fully Connected Flow (SiFu) non-representational learning paradigm and implement the BriLLM-MetaPred model. This architecture can achieve brain-inspired AGI capabilities on consumer-grade GPU hardware without relying on spiking neural network infrastructure. It can reach the same scale of pretraining and comparable performance as GPT-1 and 2. This study is the first to propose a mathematically falsifiable AGI development framework, steering the field away from heuristic scaling paths towards a physically ancho

  • Дата начала:20.11.2026 07:00
  • Дата завершения:20.11.2026 09:00

Похожие мероприятия

[Tensor Learning Team Seminar] - Talk by Dr. Rong-Yang Sun (The Chinese University of Hong Kong) “Autonomous Generation of High-Performance Tensor Network Applications with Agentic

[Tensor Learning Team Seminar] - Talk by Dr. Rong-Yang Sun (The Chinese University of Hong Kong) “Autonomous Generation of High-Performance Tensor Network Applications with Agentic

Date and time: Thursday, November 19, 2026, 14:00–...

Бесплатно Япония 13-го ноября, 3:00