Review - “Towards an evolvable neuromolecular hardware”

Amanda Seipel Neurocomputing recently devoted an entire issue to evolutionary neural systems. This issue covered research on both the biological and the artificial in an attempt to promote a synthesis of the two that would lead, ostensibly, to the creation of an artificial brain. The lead article is reviewed here, but the reader is encouraged to explore the entire volume…

Amanda Seipel, April 21, 2002

A Review of "Towards an evolvable neuromolecular hardware: a hardware design for a multilevel artificial brain with digital circuits", Chen & Chen, Neurocomputing, 42, 9-24, (2002). www.elsevier.com/locate/neucom

Reviewer: Amanda Seipel, Contributing Editor

Neurocomputing recently devoted an entire issue to evolutionary neural systems. This issue covered research on both the biological and the artificial in an attempt to promote a synthesis of the two that would lead, ostensibly, to the creation of an artificial brain. The lead article is reviewed here, but the reader is encouraged to explore the entire volume.

In an effort to give digital machines more of the advantages that biological systems have with regards to information processing (specifically, malleability and the ability to learn new strategies for new environments), Chen and Chen created a design for digital hardware based on specific molecular-level mechanisms that occur within actual neurons (the artificial neuromolecular model).

Specifically, Chen and Chen model two molecular processes that are believed to be critical for information processing in real neurons. One layer of their model contains units (cytoskeletal neurons), each of which have an internal structure based on the cytoskeleton in a human neuron. The various components of the cytoskeleton are included in the model (microtubules, microfilaments, and intermediate or neuro- filaments). The different proteins that bind these components to each other are also reflected in the model. Each component and protein has unique functional attributes. The other two layers of the model are comprised of reference neurons, which allow for memory and coordinated activation of the cytoskeletal subgroups. The cytoskeletal neurons, however, are the key to the complexity and increased adaptability of this model.

The self-organizing structural changes that are the key to biological adaptation are reflected in the model in that the concentration of cytoskeleton components within one cytoskeletal neuron is allowed to change. Each cytoskeletal neuron is allocated 64 processing units, each of which can represent one of the three cytoskeletal components and/or associated enzymes or proteins. Additionally, each of the three major components has six possible degrees of activation. Between the different types of learning that occur within a cytoskeletal neuron and the interactions between cytoskeletal neurons and reference neurons, Chen and Chen have created six possible levels at which learning can occur, all within the confines of conventional digital circuitry. This surpasses earlier attempts to model human cognition, which restricted learning to changes in the connection strengths between the units representing neurons. And, while the idea of molecular-based learning is not new, this represents a unique translation of this concept into a design for hardware.

A software simulation of this design has successfully solved problems in a variety of domains (including maze navigation, character recognition, and chronic hepatitis B diagnosis) and has been effective at adapting to problem changes. This represents a significant advance in the level of adaptability typically seen in models of human information processing. Hardware with this ability to adapt would have a greatly expanded range of applications. Chen and Chen continue to refine and expand their design, hoping to be able to include higher cognitive functions in the model.