Hsi-wen (Eddie) Chu

Research Summary

My research interests include Complementary Learning Systems, continual learning and catastrophic forgetting, neuroevolution and speciation, nonlinear dynamical systems, and macroeconomic dynamics.

Current work falls into three independent research areas. They are related by a general interest in adaptive systems, but they are not presented as a single unified theory. Each has its own research question, methods, and evidence standard.

Research Areas

Continual Learning Architectures

Independent research project, 2025-present

Investigating biologically inspired continual learning architectures for mitigating catastrophic forgetting in neural systems. Current work explores Complementary Learning Systems-inspired memory consolidation mechanisms, adaptive retrieval strategies, and persistent long-term learning across sequential tasks.

The emphasis is on whether memory systems can support both stability and plasticity over repeated updates, rather than only improving performance on a static training distribution.

Current experiments use a Complementary Learning Systems framing with a slow cortex-like model, a fast surprise-gated episodic store, and replay-based consolidation. The live-ingest line of work studies whether a user-specific adapter can accumulate experience over many sequential updates while preserving retrieval behavior and avoiding destructive drift.

AlphaTrader

Independent research project, 2021-present

Designed and implemented a neuroevolutionary framework for neural-network-based technical trading systems incorporating endogenous speciation mechanisms. The project investigates evolutionary dynamics in which specialized trading agents emerge from an initially unified population, producing expanding differentiation across evolutionary search trajectories.

Current system performance exceeds benchmark BTC buy-and-hold returns over a 13-month evaluation period.

Capital-Consumption Theory

Independent research project, 2018-present

Developed an independent macroeconomic framework using nonlinear dynamical systems to model interactions between capital accumulation, labor compensation, and consumption feedback mechanisms. The project investigates how these interactions generate long-term macroeconomic dynamics affecting productivity, inequality, inflation, interest rates, and capacity utilization.

Manuscript released as a public preprint for ongoing revision and feedback. Preprint available on SSRN.

Continual Learning Notes

This research line investigates per-user continual learning for language models. The working architecture separates a shared base model from user-specific long-term state: a persistent low-rank adapter functions as the slow-learning cortex, while a fixed-size hippocampus-like retrieval bank stores high-surprise events for later replay and retrieval.

The current experimental program emphasizes sequential ingestion rather than static benchmark training. Nightly consolidation mixes replay from the episodic store with model-generated samples from the current cortex. This allows the model to use its own distribution as a stabilizing anchor while still receiving grounded learning signal from newly ingested events.

Recent runs evaluate retrieval accuracy, span-level cross-entropy with and without the memory bank, proxy perplexity, and recovery behavior after drift triggers. The main open question is whether the architecture can remain coherent at production-scale retrieval sizes while carrying user-specific learning forward over extended sequences of updates.

Working Papers and Software

Curriculum Vitae

Education

Rice University: Master of Computer Science, Specialization in Machine Learning, 2025
UC San Diego: Certificate in Machine Learning Engineering, 2023
Rice University: BA in Cognitive Sciences, 2004

Professional Experience

HotSpot Delivery LLC: Technical Lead, 2024-present. Adapted YOLO computer vision models to liquor product recognition and fill-level estimation for automated inventory management.
SG Wines: Co-Founder and Financial Lead, 2009-2019. Managed operations, finance, and logistics for a wine import company.

Technical Skills

Python, JavaScript, Java, OCAML, SQL; PyTorch, neural networks, neuroevolution, computer vision; PostgreSQL, Pandas, AWS, REST APIs; React Native, Git, Linux.

Contact

Houston, Texas

eddiechu@rice.edu