# Research

Ruedi Stoop is Professor at the Institute of Neuroinformatics in Zürich (INI, part of ETH Zürich and University Zürich). He is originally mathematician and theoretical physicist. The Stoop Group itself is a highly interdisciplinary group, that unifies competences in mathematics, theoretical and applied physics, computer science, engineering, biology and philosophy.

The main research interest of the group is:

## Natural Biocomputation

Biology computes differently to computers. Computational processes run with cycle times far below those used in nowadays computers, and they are based on complex and recurrently connected architectures, a building principle that is generally not present in the design of artificial chips. Biological processes are also very robust, they work in permanently changing environments and compute reliably in the presence of strong noise.

Our main research question is: How can the difference between artificial devices and the complex systems found in biology be captured, understood, and be used to solve technological problems?

To answer this question, we derive measures of biological computation, examine and classify the dynamics that underlie the computational processes in biology, and model the strategies nature uses to perform computational tasks. We apply methods from nonlinear dynamics and statistical mechanics, where we define exact models and compare with experimental data. Finally, we transfer the knowledge gained into applications that implement the computational power of biological systems.

We work on seven specific research issues as listed below.

## A Hopf-system based hearing approach

The mammalian auditory system extracts from the arriving soundwaves rich portraits of the auditory environment. To a large extent, this ability is the consequence of the nonlinear signal processing in the cochlea, where the local amplification profiles have been shown to originate from systems close to Hopf bifurcations. Starting from biophysical principles, we have developed and implemented an analog electronic Hopf-cochlea displaying mammalian hearing characteristics (patented). Using this novel sensor, we aim at implementing in this project an autonomous hearing system, supported by an auditory cortico-cochlear feedback loop. In particular, we investigate the role of cochlear processing in assisting a listener to discriminate between different auditory objects in a cocktail-party environment and how much this task can be improved by an efferent tuning of the Hopf-amplifiers.

As a long-term goal, the sensor will be used as a next-generation cochlea implant. In this context, a hardware-bioware interfacing project was recently set up together with specialists in this field.

The Hopf-approach is further used to study how hearing impairment could be mended, to study and model human perception of complex sounds, and to better understand the hearing systems of insects. Finally, the Hopf-cochlea will be embedded within a general theory of Hopf-sensors.

Read an article in ETH-Life about this project here.

### Software implementation of the Hopf cochlea model

The executable implementation of the Hopf cochlea model can be downloaded here. The executable is called via the Matlab header script 'cochleainit.m'. The executable runs on a 64-bit Windows operating system.

## Behavior and language

Before mating, most animals engage in extensive courtship activities, the reason why is still subject to speculation. An obvious suggestion is that information about the genetic suitability of a prospective mate is exchanged. In natural languages, nontrivial grammars provide the mechanisms for expressing detailed relationships between fundamental symbols and introducing redundancy, by which the receiver can check the validity of the arriving information. The proposed role would thus hint at a variable courtship language with nontrivial grammatical rules as the central element.

In this project, we investigate courtship under this angle, where the emphasis is on courtship graphs and their properties.

## Information processing in the brain

The ways by which information is encoded and processed in the brain are still largely unknown. In 1997, we have proposed that the mechanism of phase-locking among coupled limit-cycle oscillators, a phenomenon widely observed in in vitro recordings, could serve as the coding scheme, unifying rate- with temporal coding in a noise-robust way. In this project, we investigate how computation by means of phase-locking interfaces with the underlying network architecture, how this could be realistically implemented in biology, and in hardware. We develop measures for the information content in spike trains, for cortical computations in general and for the efficency of phase-locked computation depending on its underlying architecture.

## Biological networks as optimized mathematical structures

We investigate and classify mathematical network structures based on biological data and constraints. The project includes the development of the relevant optimality classifiers in biological networks and an evaluation of biological time series from this point of view, using a statistical mechanics approach to networks.

## Sequential superparamagnetic clustering (SSC)

Clustering is defined as a task of grouping of similar objects. As the concept of similarity between objects is ambiguous, the clustering task is inherently ill-defined. The Sequential Super-paramagnetic Clustering (SSC) algorithm approaches the clustering task by providing a unique clustering hierarchy along with a measure of the naturalness of the cluster (where naturalness is defined as a group without any significant substructure). The SSC algorithm does not assume any a priori information about shape or internal distribution of the clusters, nor is the number of clusters predefined. Moreover, the SSC algorithm can easily deal with clusters of different shapes, densities and largely unequal distances between clusters.

The SC/SS Clustering Tool can be downloaded here.

## Natural computation

The classic notion of computation is based upon Turing computability. In neuroscience, however, the notion computation usually remains

undefined and is rather used as a metaphor to describe an "information processing" aspect of natural systems. Natural computation, that is computation in or by natural systems, is inter alia required to be a real-time, continuous and decentralised process, to be embodied in biochemical or physical processes and to show flexibility, adaptibility and robustness to external environments. We hence investigate various models of natural computation more adapted to the biological sciences than the classic Turing model. The main research focus is on cellular automata and the notion of computation put forward by the Stoop Group, that is embedded in the general framework of dynamic system theory. Related to the problem of what natural computation shall be is the quest for a general quantifying notion of information. We are investigating various measures of information and complexity regarding their applicability in neuroscience with a focus on Shannon information theory and the complexity measure developed in our group.

## Towards machine autonomy

In near future, machines that interact autonomously with their environment will enter our everyday live.

In this project, we investigate from a theoretical point of view the following questions:

- How can machine behavior be described in terms of grammar and syntax?
- How can machine behavior efficiently be implemented involving as few "intelligence" as possible?
- How can machine behavior and language be implemented by means of genetic programming?
- How can machines be equipped with means that enable them to participate autonomously in a complex environment?
- How can machine autonomy, in particular perception, be implemented in real-time?