Brains on Board: Neuromorphic Control of Flying Robots

Odor Objects: Odor-background segregation and source localization using fast olfactory processing

Neuromorphic Implementations of Multivariate Classification Inspired by the Olfactory System (Human Brain Project, EU)

Completed Research projects

BIOMACHINELEARNING (Marie Curie Fellow Dr. Michael Schmuker)

Enabling scientific computing with GPUs with domain specific languages and meta-compilers (Royal Academy of Engineering/ Leverhulme Trust)

Green Brain project (EPSRC)

Fast electronic noses through spiking neuromorphic networks

Feature selection for electronic noses

Bio-inspired Classification

Accurate Neuron Models

Spike Timing Dependent Plasticity of Inhibitory Synapses

PheroSys: Olfactory Coding in the Insect Pheromone Pathway

PostDoc work

Multifractal properties of Random Field Ising Models (PhD work)

Dimension theory of graphs (Diploma work)

BIOMACHINELEARNING (Marie Curie Fellow Dr. Michael Schmuker)

One of the main obstacles to a wider uptake of general purpose chemical sensing devices, so-called electronic noses (e-noses), are their generally slow response times. Yet, animals use their olfactory senses for essential tasks such as detecting threats and locating food or mating partners. The reason for the enormous success of biological olfactory systems is only partially due to faster sensors. Recent behavioural and physiological work has shown that animals make decisions, and that the response of olfactory brain structures is most informative, long before the receptors and the corresponding receiving neurons in the brain reach a steady state.

The three main objectives of this project are to
1. Overcome the response time problem in e-noses by developing fast, bio-inspired signal processing for chemical sensing that will be employed on powerful hardware platforms, including general purpose graphical processing unit platforms and genuinely neuromorphic hardware.
2. To bring together the expertise of the fellow, Schmuker, with the expertise of the host, Nowotny, to form a hotspot of research in bioinspired enoses within Europe.
3. To extend the training, research experience and scientific network of Dr. Schmuker to enable him to reach a position of professional maturity at the professorial level upon completion of this project.

We will achieve our objectives by combining existing bio-inspired olfaction algorithms of Nowotny with Schmuker's work on neuromorphic hardware and apply them to e-nose data sets of Nowotny's collaborators in Australia. Schmuke will also receive valuable professional training and experiences with the UK academic environment. He will form new, lasting links to collabprators of Nowotny. The proposal is highly relevant to the Work Programme in developing the fellow and generating European excellence in bio-inspired chemical sensing, with large potential future impacts on the quality of life and security applications.

Neuromorphic Implementations of Multivariate Classification Inspired by the Olfactory System (Human Brain Project, EU)

The overall goal of the proposed work is to implement a scalable spiking neuronal network for multivariate classification on the large-scale neuromorphic system developed in the Heidelberg group within the HBP. We will base our work on a study by Schmuker & Schneider (2007) in which a firing-rate model for multivariate classification inspired by insect olfaction has been devised. We will use the virtual receptor approach from this study to efficiently encode real-valued multivariate data sets into firing rate representations that are suitable for processing on spiking neuromorphic systems. Classification will be accomplished in a three-step process: Multivariate data is first encoded using broadly tuned virtual receptors. Their responses are then filtered by lateral inhibition in a decorrelation layer inspired by the insect antennal lobe. This layer projects onto a winner-take-all circuit and the weights of this projection are learnt in a supervised fashion using a perceptron learning rule. While a prototype that implements the described network on the Spikey hardware is already implemented (Schmuker et al., submitted), the challenge in the proposed novel work will lie in scaling the network up to a large number of input neurons in order to process very high-dimensional data sets. In particular, scaling up the number of virtual receptors that encode the input data to several hundreds to thousands, will allow to adequately encode data sets that live on high-dimensional manifolds and are embedded in even higher-dimensional spaces. The scaling will first be investigated in the framework of digital implementations on massively parallel graphical processing units (GPUs) for which we have appropriate expertise (GeNN simulator and related work). In a subsequent step, we will implement the large-scale network on the hardware system from the Heidelberg group. By using PyNN for network design an implementation on the SpiNNaker system can also be explored. The proposed work will directly contribute to the work in work package 113 (Future computing). We expect synergy effects from and to work packages 91, 92 and 93 (neuromorphic emulation of brain models, digital many-core implementations and software tools for neuromorphic computing). Our neuronal network designs will be tested on massively parallel digital graphical processing unit (GPU) super-computing platforms, complementing the work on digital many-core implementations of brain models undertaken in WP92. At the same time, the models will then provide a use case for the requirements on the neuromorphic computing systems developed in WP91, and provide requirements for the design of effective software tools for neuromorphic computing, which are developed in WP93. The original participants in these WPs will benefit from an additional point of view on design decisions and requirements from an independent research group like us, to make their tools and designs useful to the wider computational neuroscience and machine learning communities. Regarding the contribution of the proposed project to the ramp-up phase of the HBP and beyond, our network will provide a working proof-of-concept that the analysis of "Big Data" (in the sense of high-dimensional multivariate data sets) is feasible on large-scale neuromorphic platforms. Our work will furthermore expose the specific benefits of a massively parallel neuromorphic approach for Big Data processing, and identify challenges to be addressed in future projects. The main thrust of the proposed project aligns perfectly with HBP WP 113 "Future Computing", since it provides a bio-inspired design for a computing system that solves computational tasks outside the realm of biology.

Enabling scientific computing with GPUs with domain specific languages and meta-compilers (Royal Academy of Engineering/ Leverhulme Trust)

In scientific computing, our ability to build accurate models of natural and engineered structures is often impeded by a lack of computational speed. This limits research in two ways: problem size is limited and accuracy is constrained through resolution and extent of parameter exploration. Traditionally, expensive super-computers are used to achieve simulations of relevant size and accuracy. Here I propose to develop a new approach of using graphical processing units (GPUs) for super-computing to overcome speed limitations in scientific and engineering applications. With optimized parallel code, modern GPUs can be 10 to 100 times faster than a single core on a contemporary CPU. In 2011 I began developing a code-generation based framework for GPU simulations of neuronal networks and have built the GeNN (GPU enhanced neuronal networks) prototype (http://genn.sourceforge.net). I now propose to extend the underlying idea of using domain specific languages (DSLs) in conjunction with meta-compilers to generate optimized code for GPU platforms. This approach has decisive advantages in that code can be optimized both for each individual scientific model and the specific GPU detected at compile time. Furthermore, we can offer virtually unlimited model libraries while the generated code remains compact and efficient.

Green Brain Project

The Green Brain Project is an initiative in collaboration with the University of Sheffield to formulate models of the honeybee brain and simulate them in real time on masively parallel GPU hardware. The model will then be used to control an autonomous flying robot to solve benchmark congnitive tasks that bees are known to master.

Our partners in this project are

Dr. James Marshall, Professor of Theoretical and Computational Biology

Dr. Eleni Vasilaki, Senior Lecturer in Computer Science

Prof. Kevin Gurney, Professor of Psychology

Abstract:

The development of an "artificial brain" is one of the greatest challenges in artificial intelligence, and its success will have innumerable benefits in many and diverse fields from robotics to cognitive psychology. Most research effort is spent on modelling vertebrate brains. Yet, smaller brains can display comparable cognitive sophistication while being more experimentally accessible and amenable to modelling.

The "Green Brain Project" will combine computational neuroscience modelling, learning and decision theory, modern parallel computing methods, and robotics with data from state-of-the-art neurobiological experiments on cognition in the honeybee Apis mellifera to build and deploy a modular model of the honeybee brain describing detection, classification, and learning in the olfactory and optic pathways as well as multi-sensory integration across these sensory modalities.

Fast electronic noses through spiking neuromorphic networks

Chemosensing is an important technology for environmental monitoring, with great potential for applications in everyday life, e.g. demand-driven air conditioning, food quality assessment and even detection of illness. Existing chemosensors are inherently slow but it is known that their detection speed can be enhanced by smart processing [1,2]. Here we do this with a bio-inspired approach using adaptive neurons [3], operating on output of novel sensors [4]. The work sets the stage for online e-nose analysis using accelerated neuromorphic hardware [5]. Matching EPSRC strategic goals, this project brings together scientists from the University of Sussex, Queen Mary University and University of Leicester in a new interdisciplinary collaboration of leaders in the respective areas of GPU computing, chemical sensors and neuromorphic computing. The work is directly relevant to the EPSRC growth area of digital signal processing.

Feature selection for electronic noses

In this project we are investigating feature selection for electronic noses in real world applications. These include general classification of volatiles, breath analysis and the detection of pests in crops.

Our partners in this project are

Dr. Amalia Berna, Research Scientist

Dr. Stephen Trowell, theme leader Quality Biosensors

PheroSys: Olfactory Coding in the Insect Pheromone Pathway: Models and Experiments

The aim of our project is to investigate olfactory information processing in the first stages of the olfactory pathway. We will conduct experiments and modelling studies in the pheromone subsystem of the moth Spodoptera littoralis (the cotton leafworm).

Our partner groups in this project are

1.   Jean-Pierre Rospars

2.   Dr Dominique Martinez

Bioinspired classification algorithms

We have continued our work on the olfaction of insects. From our results we have developed a general classification system and have tested it on the MNIST database of handwritten digits.

Accurate conductance based models of identified cells

In the process of developing novel hybrid systems technology for online measurement and manipulation of biological systems we have become interested in the problem of accurate models of identified cells/ identified cell types. In many neurons the direct assembly of voltage clamp data to a Hodgkin-Huxley type neuron model seems to fail. The reasons for this are manifold. A trivial but unresolvable problem is that each current is characterized on a different preparation. Even with the greatest care and using identified cells, there will always be experimental and animal-specific variations in the properties. Other more fundamental reasons might be redundancies of paramters that lead to an appropriate paramter region which is non-convex, such that averaging, as done in the voltage clamp experiments, is inappropriate to determine adequate parameter values. On an even more principal level, identified cells might not be identifiable based on similar parameters but similar dynamics with possibly completely different underlying paramter sets as suggested recently (Prinz et al., Nature Neurosci. (2004)).
I have adapted data fitting techniques to overcome these problems and fit conductance based models to large amounts of data.

Inhibitory Plasticity in the Entorhinal Cortex

I have also started an investigation of the possible functions of inhibitory plasticity in the entorhinal cortex in collaboration with Julie Haas. This is a puzzeling question as potentiation of inhibitory synapses occurs for post-after-presynaptic spike pairings which subsequently suppresses this type of events. The plasticity is removing the spike patterns it is caused by rather than imprinting them into the system as STDP of excitatory synapses does. We are able to demonstrate in models that this type of inhibitory spike-timing dependent plasticity can be a powerful while subtle control for run-away or seizure-like activity in the entorhinal cortex.

PostDoc work

After my PhD I once more took a slight change in my field of work. Presently I investigate neural systems with several main questions in mind.

For one we try to understand the mechanisms of sequence learning in neural systems mediated by Spike Timing Dependent Plasticity (STDP). First encouraging results have been achieved and a manuscript is in preparation.

The second topic I'm presently working on is the information processing in the olfactory system in particular in the mushroom body of the locust. This information processing seems to have extremely interesting dynamical systems properties. The key question is how the spatio-temporal code produced by the winnerless competition principle believed to be implemented in the antennal lobe might be processed in the downstream areas (mushroom body, beta lobe).

The third topic arose in connection with recent work done by Valentin Zhigulin and Mish Rabinovich in our group. They showed that a synapse obeying an anti-Hebbian STDP learning rule can synchronize model neurons of various types in a very effective way which comes as a little bit of a surprise. As this inverse STDP is a rather rare effect in nature we decided to check whether the effect might be relevant for real biological neurons. To this end I adapted a dynamic clamp software of Reynaldo Pinto to include learning rules for the synapse emulated by the software. We now use this program to couple a simulated driving neuron with a real neuron from the abdominal ganglion of Aplysia.

PhD work

For my PhD work I investigated properties of multifractals in general and the multifractal properties of the probability distributions of effective fields in random field Ising models in particular.

The one-dimensional random field Ising model with dichotomous quenched random field can be reformulated as a random iterated function system (RIFS) of two functions. This RIFS is contractive and therefore has by a result of Hutchinson 1984 a unique invariant measure. This invariant measure turns out to be a multifractal for a generic choice of the physical parameters. Its multifractal properties change drastically if the physical parameters temperature and random field strength are changed. We were able to find and understand the mechanism underlying these drastic changes often called phase transitions in the spirit of the thermodynamic formalism for multifractals. The results are published in the paper "Orbits and phase transitions in the multifractal spectrum" cited below.

During my investigations I noticed that while not being a physical observable itself the effective field under investigation leads in a simple way to the local magnetization. The local magnetization in the one-dimensional random field Ising model is essentially just the sum of two effective fields. Its probability distribution in the thermodynamic limit therefore is essentially the convolution of the invariant measure of the effective field with itself.

This leads directly to the question whether one can deduce information about the multifractal properties of the convolution of two multifractals from the multifractal properties of the two measures being convoluted. It turns out that on can give lower and upper bounds on the Dq-spectrum of the convolution in terms of the Dq-spectra of the convoluted measures. The results are published in the paper "Convolution of multifractals and the local magnetization in a random field Ising chain". This paper also contains the application of the bounds to and numerical results on the multifractal properties of the local magnetization in the one-dimensional random field Ising model.

In a third part of my PhD work I addressed the question of phase transitions, now in the proper physical sense, in the random field Ising model on the Bethe lattice. We developped several numerical criteria for such phase transitions based on our iterated functions approach. The numerical analysis revealed interesting discrepancies with an early work of Bruinsma. The results and a discussion are published in the paper "Phase diagram of the random field Ising model on the Bethe lattice".

All results obtained during my PhD work are summarized in my PhD thesis (naturally). As the University of Leipzig thankfully allows thesis submission in English the thesis is in this language and therfore hopefully accessible to everyone. It can be downloaded from the link below.

Diploma work

The first subject which rendered concrete results was dimension theory of graphs. We, i.e. M.Requardt and I, developed two notions of dimension on graphs and investigated their properties. The results are summarized in a paper about dimension theory of graphs and networks.

The second paper on pregeometric concepts follows a similar spirit.

My diploma thesis sums up most of what I did on this topic as well as a brief survey of different approaches to define geometric concepts on graphs. Regulations of the University of Göttingen require any diploma thesis to be in German.

The unpublished notes sum up some of the results of numerical investigations carried out throughout the two years of work on my diploma thesis. They are slightly incomplete because of lack of time. I apologize that they are in German because they originally were not intended for publication.