Andrea Loettgers

Computer simulations have become a universal instrument for knowledge production. They are of enormous importance for technological development. The project focuses on the practice of development and application of computer simulations in studying neural networks. Neural networks are the basic physiological structure of the human brain, which consists of billions of nerve cells (neurons), each connected to thousands of others by synaptic couplings.

In 1982, the American physicist John Hopfield published the results of investigations of a special property of neural networks (1): to retrieve a stored pattern of information from incomplete or noisy input. This property is called associative memory. Hopfield used computer-based simulations in his investigations in order to show that his model for describing the neural network indeed shows the property of associative memory. The model, which quickly became called Hopfield model, has its origin in solid-state physics where it was and is used to describe disordered magnetic systems, so-called spin-glasses. Hopfield's computer simulations give rise to the following historical and epistemological questions: How is knowledge produced about our brain by performing computer simulations? What were and what are the foregoing conditions of knowledge production by computer simulations? What is the relation between the knowledge produced by computer simulations and the simulated, real processes? What is it that the human brain and disordered magnetic systems have in common and that allows us to describe them by the same model?

A crucial point in answering these questions is the question of historical sources. A huge number of publications allow us, on the one hand, to trace back Hopfield's investigations and to reconstruct its roots in neurophysiology, computer science, psychology, mathematics and in contexts of scientific practice where models without computer simulations were used. On the other hand, those publications that are following up on Hopfield's work allow us to examine how the Hopfield model and the computer simulations were used, changed and improved, which kind of results were produced, and how investigations in different disciplines were connected. This analysis is being done by means of a bibliographical database that classifies and hyperlinks the various publications on the basis of an analysis of their contents, their formulation of and perspective on the questions, as well as by their respective citations. This bibliographical analysis will give an overview of the network in which Hopfield's work is to be placed historically. However, it does not provide any access to the practice of constructing, testing, using, changing, and improving of the computer simulations.

The claim that computer simulations are an instrument of knowledge production also suggests another, complementary approach. The method of replicating historical experiments has been shown to provide interesting new insights into historical experimental practice. (2) In our project, a replication of Hopfield's computer simulations will be done in the very same spirit as a replication of a historical experiment. Producing a replica version of Hopfield's computer programs of associative memory implies designing, testing, changing and optimizing a historical computer program. A working version of the historical computer code opens up the possibility that important aspects of the development and application become visible and can be explored by hands-on-experience with all the constraints and possibilities of historical programs. These aspects include: Were the model and the dynamics of its simulation tuned with respect to each other and were they connected in any way? How were the model and the dynamics translated into a computer program? What was the role of the programming language, the available hardware, their storage capacity, speed and architecture? How were the programs tested? What strategies were developed and applied in the testing of the programs? In what form was the output given and how was it further processed? And, finally, how were the results obtained and compared to the real world?

(1) See, e.g., Hopfield, John: 'Neural networks and physical systems with emergent collective computational abilities', Proc. Natl. Acad. Sci. USA 79, 1982.
(2) Peter Heering, Falk Riess, Christian Sichau: Im Labor der Physikgeschichte. Zur Untersuchung historischer Experimentalpraxis.
Bibliotheks- und Informationssystem der Universitaet Oldenburg, 2000.