Molecular ML

Bio-Inspired Human-Level Machine Learning

Sep 2012 - Sep 2015 Air Force Research Laboratory (AFRL)

Overview

How can brain computation be so fast, flexible, and robust? What kinds of representational and organizational principles facilitate the biological brain to learn so efficiently and flexibly on the sub-second time scale and so reliably on the continuous lifetime scale? Understanding these principles and constructing computational models that implement them in a natural way can achieve scientific breakthroughs in computational architectures and algorithms that enable true human-level robust intelligence.

This project developed bio-inspired machine learning technology that is competitive with human learning in its performance (speed, flexibility, reliability, robustness) and style (online, incremental, predictive, self-teaching). The approach rests on two interlocking layers: a “human-like” machine learning model based on dynamic neural populations (neural assemblies), and a “molecular” implementation of that model in molecular populations (molecular assemblies) using in vitro DNA computing. The molecular machine learning model is validated on high-level cognitive information processing tasks involving language, vision, and decision-making.

The project targets innovation across computational intelligence, cognitive science, and engineering. The notion of bio-inspired human-level machine learning combined with molecular-computing implementation — grounded in population coding and dynamic coordination — offers a novel paradigm for flexible and reliable computing. In particular, the dynamic molecular assembly model of cognitive memory and learning provides a new tool for simulating dynamical cognitive systems.

Two-Layer Architecture

Human-like machine learning model: knowledge is encoded as weighted hypergraphs over dynamic neural assemblies, enabling online, incremental, and predictive learning from few examples.
Molecular implementation: in vitro DNA computing executes the learning model using DNA strands as computational units, realizing molecular pattern classification via DNA beacons and DNA-based language models.

Validated Cognitive Tasks

Language processing: sentence parsing and generation using a DNA language model encoding semantic and orthographic distances
Visual pattern recognition: concept learning from multimodal stimuli (including cartoon video) via hypernetworks
Decision-making under uncertainty: anagram solving through molecular computational simulation of cognitive processes

Research Team

Principal Investigator

Prof. Byoung-Tak Zhang (Seoul National University)

Researchers

Ji-Hoon Lee
Je-Hwan Ryu
Hyo-Sun Chun
Christina Baek
Sang-Woo Lee

Contact: Prof. Byoung-Tak Zhang — btzhang@snu.ac.kr

Publications

Zhang, B.-T. (2013). Communication as moving target tracking: Dynamic Bayesian inference with an action-perception-learning cycle. In: Wachsmuth, I. et al. (Eds.), Alignment in Communication: Towards a New Theory of Communication, Chapter 7, Simon & Shuster.
Zhang, B.-T. (2013). Information-theoretic objective functions for lifelong learning. AAAI 2013 Spring Symposium on Lifelong Machine Learning, March 25-27, Stanford University, AAAI Press.
Zhang, B.-T. (2012). Higher-order predictive information for learning an infinite stream of episodes. NIPS Workshop, Lake Tahoe.
Lee, J.-H., Lee, E.S., Ryu, J.-H., Chun, H.-S., Zhang, B.-T. (2013). Molecular computational simulation of cognitive processes for anagram solving. The 19th International Conference on DNA Computing and Molecular Programming (DNA19), Phoenix. (poster)
Ryu, J.-H., Lee, J.-H., & Zhang, B.-T. (2013). Integrated encoding of semantic and orthographic distances in a DNA language model. The 19th International Conference on DNA Computing and Molecular Programming (DNA19), Phoenix. (poster)
Lim, H.-W., Lee, S.H., Yang, K.-A., Yoo, S.-I., Park, T.H., & Zhang, B.-T. (2013). Biomolecular computation with molecular beacons for quantitative analysis of target nucleic acids. BioSystems, 111:11-17.
Lee, I.-H., Lee, S.H., Park, T.H., & Zhang, B.-T. (2013). Non-linear molecular pattern classification using molecular beacons with multiple targets. BioSystems, 114(3):206-213.
Lee, B.J., Ha, J.W., Kim, K.M., Zhang, B.T. (2013). Evolutionary concept learning from cartoon videos by multimodal hypernetworks. IEEE Congress on Evolutionary Computation (CEC 2013), pp. 1186-1192.