Singapore
Hi, I am Fangyuan, a Senior AI Scientist at @Thoughtworks AI Lab. I work on pushing the boundary towards AGI. Rather than teaching neural networks to mimic human language, I study how they can grow their own internal language of abstractions. Following Feynman’s “what I cannot create, I do not understand,” I’m chasing a unification of symbolic reasoning and neural networks as a step toward AGI. Previously, I worked as Research Scientist on automonous driving perception system, as well as Senior AI Engineer on LLM agents. I hold a PhD in Markov model and Sequential Monte Carlo Sampling at KAUST. My Google Scholar link: https://scholar.google.pl/citations?user=GqZfs_IAAAAJ&hl=en
Pushing boundary of AGI - Emergence of abstraction (language of thought) within neural nets - Emergence of causality within neural nets - Neural consolidation & schema effect
Information bottleneck language modeling Pre-training LLM with vocabulary curriculum Annotation-free LLM & VLM Alignment
1. Enhance self-driving perception systems, focusing on object depth estimation and predictive modeling 2. Develop real-time 3D object detection systems; Reduced position estimate relative error from 10% to 3% by neural network structure refinement, geometric-based data augmentation, online position optimization & temporal smoothing. Deploy the vehicle perception system in C++ for mass production and commercial sale 3. Develop generative AI video prediction models for autonomy, generating realistic driving videos using multi-modal input. Train diffusion model on massive real-world driving videos, to predict subsequent frames in video sequence, achieving autoregressive prediction through self-supervised learning similar to LLMs 4. Develop point cloud registration system using state-of-the-art graph-based algorithm; Achieve 10-fold reduction in registration time, by heuristic down sampling which filters out moving objects and ground points
1. Coded Python & R packages to apply sequential Bayesian inference on robot localization problems (and other filtering problems) Analyzed sequential noisy sensor observation data, model uncertainty with Hidden Markov Process, and conduct inference to estimate the hidden state using the sequential Monte Carlo method (Particle Filter). 2. Proposed and studied theoretically Coupled strategy which achieves variance reduction effect on Particle Filter for the diffusion process. Reducing the cost order to O(e^2) from O(e^3) to achieve an accuracy level (MSE) of O(e^{-2}). 3. Published the result in an academic journal.