Research Fellow
Institute of Neuroinformatics, ETH and University of Zurich
School of Informatics,
University of Edinburgh
I am fascinated by the organisation of brains and the situated behavioural repertoire they produce. My current focus is on how brains enable the impressive navigation capabilities of insects.
Institute of Neuroinformatics, ETH and University of Zurich
INP APPS
University of Essex
LOGOS Research, Greece
Ph.D. in BioRobotics
School of Informatics,
University of Edinburgh
Artificial Intelligence (Robotics)
Department of Artificial Intelligence,
University of Edinburgh
Electronics Engineering
Electronics Department, TEI of Piraeus, Greece
I am currently studying the brain circuits that enable the impressive navigation capabilities of insects. Specifically, my focus is on modelling the neuronal circuits of the Central Complex of the insect brain.
In the past, I developed neuro-symbolic learning methods for robots. These methods enabled robots to learn about their environment, make plans, and navigate autonomously.
Other research interests that I pursue when time allows are sensory modalities and signal processing, as well as space exploration applications of robotics.
Solitary foraging insects, such as ants, maintain an estimate of the direction and distance to their starting location as they move away from it, in a process known as path integration. This estimate, commonly known as the “home vector,” is updated continuously as the ant moves and is reset as soon as it enters its nest, yet ants prevented from returning to their nest can still use their home vector when released several hours later. This conjunction of fast update and long persistence of the home vector memory does not directly map to existing accounts of short-, mid-, and long-term memory; hence, the substrate of this memory remains unknown. Chill-coma anesthesia has previously been shown to affect associative memory retention in fruit flies and honeybees. We investigate the nature of path integration memory by anesthetizing ants after they have accumulated home vector information and testing if the memory persists on recovery. We show that after anesthesia the memory of the distance ants have traveled is degraded, but the memory of the direction is retained. We also show that this is consistent with models of path integration that maintain the memory in a redundant Cartesian coordinate system and with the hypothesis that chill-coma produces a proportional reduction of the memory, rather than a subtractive reduction or increase of noise. The observed effect is not compatible with a memory based on recurrent circuit activity and points toward an activity-dependent molecular process as the basis of path integration memory.
Many animal species are able to return to their nest after a foraging excursion without using familiar visual cues to guide them. They accomplish this by using a navigation competence known as path integration, which is vital in environments that do not have prominent visual features. To perform path integration, an animal maintains a running estimate of the distance and direction to its origin as it moves. This distance and direction estimate needs to be maintained in memory until the animal uses it to return to its nest. However, the neural substrate of this memory remains uncertain. A common hypothesis is that the information is maintained in a bump attractor’s state. We test the bump attractor hypothesis and find that its predictions are inconsistent with the path integration behaviour of ants, thus highlighting the need for alternative models of path integration memory.
Recent studies of the Central Complex in the brain of the fruit fly have identified neurons with activity that tracks the animal's heading direction. These neurons are part of a neuronal circuit with dynamics resembling those of a ring attractor. The homologous circuit in other insects has similar topographic structure but with significant structural and connectivity differences. We model the connectivity patterns of two insect species to investigate the effect of these differences on the dynamics of the circuit. We illustrate that the circuit found in locusts can also operate as a ring attractor but differences in the inhibition pattern enable the fruit fly circuit to respond faster to heading changes while additional recurrent connections render the locust circuit more tolerant to noise. Our findings demonstrate that subtle differences in neuronal projection patterns can have a significant effect on circuit performance and illustrate the need for a comparative approach in neuroscience.
Insects demonstrate a remarkable ability to navigate the world -- foraging for food and returning to their nests over large distances. Simple neural circuits have been shown to underlie insect navigation behaviours, encoding the animal's head direction and velocity vector as a sinusoidal activity pattern. We offer novel mathematical insights into the benefits of this sinusoidal encoding, suggesting that it might offer an evolutionary advantage in terms of noise resilience. Previous work has recognised that encoding vectors as sinusoidal population activity patterns enables vector addition by simple piecewise activity addition. However, we show that other activity patterns also fulfil this property, calling into question whether this is the only advantage of a sinusoidal activity pattern. In this work we use signal processing formalism to demonstrate that the sinusoid is the most noise-resilient activity shape that allows for stable path integration, given that it is coupled with a sinusoidal connectivity. We additionally show that the required sinusoidal connectivity will arise automatically in a network of neurons which have their weights updated with a Hebbian learning rule. This analytical result demonstrates that the weight matrix and the activity could naturally converge to sinusoidal shapes during development.
Many animals use path integration to maintain a running estimate of the direction and distance to their nest. Solitary foraging insects can return to their nest using path integration after excursions of several kilometres. As an insect moves, its path integration estimate is updated continuously and can be reset once inside the nest. However, ants prevented from returning to their nest remember their path integration state for several days. The conjunction of fast updating and long persistence of the path integration memory does not directly map to current short, mid, and long-term memory accounts, while no signs of consolidation phases are present. We used neuroanatomically constrained models of the insect path integration system to investigate the nature of the underlying memory substrate. Recent results using cold-induced anesthesia to manipulate the path integration memory of ants found that temperature reduction degrades the homing distance but not the homing direction. Our model predicts that the two components of path integration memory — the direction and distance to the nest — even though differentially affected by anesthesia, are not maintained in two separate memories. Instead, they are combined as two aspects of a sinusoidal population code pattern, with the sinusoid’s amplitude encoding the distance and its peak the direction. Furthermore, we investigated the different effects low temperature might have on memory and found that a memory loss proportional to its original value best reproduces animal homing behaviour. Our results explain how a population encoding with proportional memory loss results in degradation of the homing distance but not direction, as observed in ants. Moreover, using the degradation dynamics of the path integration memory of non-anesthetised animals as constraints, we show that, unlike models of rodent path integration, the substrate of ant path integration memory is incompatible with attractor network dynamics. Although the variance (error) of both animal homing distance and attractor network state increases with time, the animal homing distance only decreases, while the state of attractor networks may increase or decrease following a random walk. Our findings point towards a new type of underlying memory mechanism with properties that differ from current memory models and are adapted to the ecological needs of ants and other insects.
Solitary foraging insects use path integration to estimate the direction and distance to their nest. This estimate, the home vector, is updated continuously and is reset once an ant enters its nest. Ants prevented from returning to their nest remember their home vector for several hours. The conjunction of fast update and long persistence of the home vector does not directly map to current short, mid, and long-term memory accounts. We sought to investigate the nature of this memory using cold-induced anaesthesia (chill-coma). We hypothesised that chill-coma anaesthesia would disrupt any reverberating activity in recurrent circuitry, which has been hypothesised to underlie home vector memory. We captured Myrmecia croslandi foragers 11m away from their nest (full vector) and returning foragers outside their nest (zero vector). We kept all ants in darkness for 30min, half of them at ambient temperature and the other half at 0ºC, inducing chill-coma. Once the anaesthetised ants regained locomotion, we gave food to all ants and released them individually in an unfamiliar location. Zero vector ants searched for their nest while non-anaesthetised full vector ants ran towards the fictive nest location. By contrast, anaesthetised full vector ants walked in the fictive nest direction but did not walk as far as the non-anaesthetised full vector ants. Using modelling, we found that the best explanation for this phenomenon is that the home vector is stored in Cartesian coordinates and that chill-coma proportionally reduced the memory across all coordinate axes by around 85%, resulting in degradation of the distance but not the direction memory. Our results pose strict constraints on the plausible biophysical mechanisms that may underpin path integration memory.
Ants, bees and other insects are able to return to their nest using a navigation strategy known as path integration. To perform path integration, an animal maintains an estimate of the distance and direction to its nest as it travels. This information needs to be maintained in memory for the duration of the trip so the animal can return to its nest. However, the nature of this memory remains uncertain. It is often hypothesised that this information is maintained through sustained neural activity in a recurrent network that forms a bump attractor. We sought to test this hypothesis by comparing its predictions with actual behavioural data of path integrating Cataglyphis fortis ants.
Understanding neuronal circuits that have evolved over millions of years to control adaptive behavior may provide us with alternative solutions to problems in robotics. Recently developed genetic tools allow us to study the connectivity and function of the insect nervous system at the single neuron level. However, neuronal circuits are complex, so the question remains, can we unravel the complex neuronal connectivity to understand the principles of the computations it embodies? Here, I illustrate the plausibility of incorporating reverse engineering to analyze part of the central complex, an insect brain structure essential for navigation behaviors such as maintaining a specific compass heading and path integration. I demonstrate that the combination of reverse engineering with simulations allows the study of both the structure and function of the underlying circuit, an approach that augments our understanding of both the computation performed by the neuronal circuit and the role of its components.
We identify differences in the head direction tracking circuit of different insect species that result in faster response to turns for the fruit fly and higher robustness to noise for the locust. These differences in the dynamical response of the neural circuits have significant ecologically relevant repercussions on the navigational capabilities of each species. Our results highlight the advantages of a comparative approach, combining behavioural with neurobiological data, to the study of insect navigation.
Ants, bees and other insects have the ability to return to their nest or hive using a navigation strategy known as path integration. Similarly, fruit flies employ path integration to return to a previously visited food source. An important component of path integration is the ability of the insect to keep track of its heading relative to salient visual cues. A highly conserved brain region known as the central complex has been identified as being of key importance for the computations required for an insect to keep track of its heading. However, the similarities or differences of the underlying heading tracking circuit between species are not well understood. We sought to address this shortcoming by using reverse engineering techniques to derive the effective underlying neuronal circuits of two evolutionary distant species, the fruit fly and the locust. Our analysis revealed that regardless of the anatomical differences between the two species the essential circuit structure has not changed. Both effective neural circuits have the structural topology of a ring attractor with an eight-fold radial structure. However, despite the strong similarities between the two ring attractors, there remain differences. Using computational modelling we found that two apparently small anatomical differences have significant functional effect on the ability of the two circuits to track fast rotational movements and to maintain a stable heading signal. In particular, the fruit fly circuit responds faster to abrupt heading changes of the animal while the locust circuit maintains a heading signal that is more robust to inhomogeneities in cell membrane properties and synaptic weights. We suggest that the effects of these differences are consistent with the behavioural ecology of the two species. On the one hand, the faster response of the ring attractor circuit in the fruit fly accommodates the fast body saccades that fruit flies are known to perform. On the other hand, the locust is a migratory species, so its behaviour demands maintenance of a defined heading for a long period of time. Our results highlight that even seemingly small differences in the distribution of dendritic fibres can have a significant effect on the dynamics of the effective ring attractor circuit with consequences for the behavioural capabilities of each species. These differences, emerging from morphologically distinct single neurons highlight the importance of a comparative approach to neuroscience.
For a variety of navigation behaviours, an insect needs to be able to keep track of its heading relative to salient external objects. A highly conserved brain region known as the ‘central complex’ has been identified as being of key importance for an insect to keep track of its heading. We have modelled the heading tracking circuit found in two evolutionary distant insect species, the fruit fly and the locust, at the single neuron level. Our analysis shows that the neuronal circuitry has been preserved through evolution and functions as an internal compass in both species. However, two small neuroanatomical differences have significant functional effects enabling the fruit fly circuit to respond faster to heading changes while rendering the locust circuit more tolerant to noise. Our results highlight that even seemingly small differences in the distribution of dendritic fibres can have a significant effect on the dynamics of the circuit with consequences for the behavioural capabilities of each species. These findings highlight the importance of a comparative approach to neuroscience.
Ants, bees and other insects have the ability to return to their nest or hive using a navigation strategy known as path integration. Similarly, fruit flies employ path integration to return to a previously visited food source while moths use it to migrate. An important component of path integration is the ability of the animal to keep track of its heading relative to salient visual cues. A highly conserved brain region known as the central complex has been identified as being of key importance for the computations required for an insect to keep track of its heading. However, the details of the underlying heading tracking neural circuit as well as its operation are not well understood. We sought to address this shortcoming by deriving the effective underlying neural circuits of two species, the fruit fly and the locust. Our analysis revealed that regardless of the anatomical differences between the two species the essential circuit structure has not changed. Both effective neural circuits have the structural topology of a ring attractor with an eight-fold symmetry. However, despite the strong similarities between the ring attractors in the fruit fly and the locust, there remain differences. We found that two apparently small anatomical differences have significant functional effect on the ability of the two circuits to track fast rotational movements and to maintain a stable heading signal. Our results highlight that even seemingly small differences in the distribution of dendritic fibres can have a significant effect on the dynamics of the effective ring attractor circuit with consequences for the behavioural capabilities of the animal. These differences, emerging from morphologically distinct single neurons, call into question the validity of broad generalisations drawn from studies in any one species, such as Drosophila, and highlight the importance of a comparative approach to neuroscience.
Recent studies of the Central Complex brain area of the fruit fly Drosophila melanogaster have identified neurons with localised activity that tracks the animal’s heading direction, in a similar fashion to rat `head direction' cells. These neurons are part of a neural circuit with dynamics resembling those of a ring attractor. Other insects have a homologous circuit sharing a generally similar topographic structure but with some important differences. Most salient are the differences in the pattern of inhibitory connections and the shape of the Ellipsoid Body that forms a torus in Drosophila melanogaster but has open edges in other insects. In this study, we reverse engineer and compare the neural circuits found in the fruit fly Drosophila melanogaster and the desert locust Schistocerca gregaria. We demonstrate that once the neural circuitry is untangled the effective circuits in the two species are identical. Both effective circuits have an eight-fold symmetry with each class of neurons assuming the same roles in both species. That is, even though in Drosophila melanogaster the Protocerebral Bridge has nine glomeruli on each hemisphere instead of the eight glomeruli found in Schistocerca gregaria, the effective circuit has eight-fold symmetry as that of the locust Schistocerca gregaria. The ensemble of Delta7 neurons provides global, uniform, inhibition in Drosophila melanogaster but non-uniform focused inhibition in the locust Schistocerca gregaria. In both species, the P-EN, E-PG and P-EG neurons form a ring with eight-fold symmetry in which the P-EG neurons form local feedback loops improving the stability of the activity `bump'. These findings reveal that even though the neural anatomy differs significantly between these two evolutionary remote species the effective circuit is preserved.
Understanding neural circuits which have evolved over millions of years to control adaptive behaviour may provide alternative solutions for robotics. Recently developed genetic tools and methods allow us to study the connectivity and function of the insect nervous system at the single neuron level, but can we unravel this complex spaghetti to understand the principles of computation it embodies? We here illustrate the plausibility of such an approach by reverse engineering part of the Central Complex circuit in the insect brain, which is known to be involved in navigational behaviours such as maintaining a specific compass heading and path integration. We demonstrate that analysis of the effective structure results in an orderly circuit forming a ring attractor with an eight-fold symmetry, capable of tracking the current heading of the animal.
The contribution of this paper is that illustrates the use of funneling actions in combination with local deictic reference frames for forming consistent and useful large scale maps. These maps do not rely on any geodetic sensors. Indications for the feasibility of such representations in humans, and other species, can be found in studies of spatial cognition. However, such implementations or applications in robotics have not been illustrated until now.
Robots operating in the real world should be able to make decisions and plan ahead their actions. We argue that learning using generalized representations of the robot's experience can assist such a ability. Design/methodology/approach: We present results from our research on methods for enabling mobile robots to plan their actions using generalized representations of their experience. Such generalized representations are acquired through a learning phase during which the robot explores its environment and builds subsymbolic (connectionist) representations of the result that its actions have to its sensory perception. Then these representations are employed by the robot for autonomously determining task-achieving sequences of actions (plans),for attaining assigned tasks. Findings: Such subsymbolic mechanisms can employ generalization techniques in order to pursue plans through unexplored regions of the robot's environment. Originality/value: Subsymbolic motion planning can autonomously determine task-achieving sequences of actions in real environments, without using presupplied symbolic knowledge, but instead generating novel plans using previously acquired subsymbolic representations.
The ability of a robot to estimate its location in the environment is crucial for variety of tasks. Several approaches have been developed offering either low cost solutions with low accuracy or high accuracy solutions at high cost. This paper presents a displacement sensor, for mobile robots, which exploits the frequency shift caused to a signal due to the robot's motion. This method offers a low cost solution combined with higher accuracy than other approaches of similar cost. The achieved accuracy level in combination with the relatively low power consumption and weight make this sensor a reasonable choice for mobile robotics applications.
In this paper, we investigate the feasibility of building action-planning mechanisms capable of autonomously determining task-achieving sequences of actions (plans), using previously acquired subsymbolic representations. These subsymbolic representations are acquired by the robot autonomously during an exploration phase. Furthermore, we investigate whether such subsymbolic mechanisms can employ generalisation techniques in order to pursue plans through unexplored regions of the robot's environment. Performance comparison of three subsymbolic action-planning mechanisms on different tasks conclude the paper.
The ability to determine a sequence of actions in order to reach a particular goal is of utmost importance to mobile robots. One major problem with symbolic planning approaches regards assumptions made by the designer while introducing externally specified world models, preconditions and postconditions. To bypass this problem, it would be desirable to develop mechanisms for action planning that are based on the agent's perceptual and behavioural space, rather than on externally supplied symbolic representations. We present a subsymbolic planning mechanism that uses a non-symbolic representation of sensor-action space, learned through the agent's autonomous interaction with the environment. In this paper, we present experiments with two autonomous mobile robots, which use autonomously learned subsymbolic representations of perceptual and behavioural space to determine goal-achieving sequences of actions. The experimental results we present illustrate that such an approach results in an embodied robot planner that produces plans which are grounded in the robot's perception-action space.
Space and planetary exploration and construction can be significantly facilitated by the use of robotic technology, which can provide low risk and low cost means. Especially for remote missions, it seems necessary to use robots able to autonomously pursue mission goals specified beforehand by humans. One key competence that such an autonomous robotic agent should possess is the ability to plan sequences of activities for achieving mission goals. Current technologies base such planning skills on human intervention and/or a priori obtained maps of the operation environment. However, missions to remote unexplored planetary objects will not always have these privileges of pre-supplied maps of the landing and operation terrain and of frequent communication with human operators on earth. On the contrary we need robots able, after landing, to adapt to their environment and learn how to successfully operate in it autonomously for achieving pre-specified mission goals. Achieving such degrees of autonomy requires two capabilities of the robots: first, to be able to learn how to operate in a new environment and second, to be able to generalize over their experience in order to behave properly in novel situations by exploiting previous experience.
The ability to determine a sequence of actions in order to reach a particular goal is of utmost importance to mobile robots. One major problem with symbolic planning approaches regards assumptions made by the designer while introducing externally specified world models, preconditions and postconditions. To bypass this problem, it would be desirable to develop mechanisms for action planning that are based on the agent's perceptual and behavioural space, rather than on externally supplied symbolic representations. We present a subsymbolic planning mechanism that uses a non-symbolic representation of sensor-action space, learned through the agent's autonomous interaction with the environment. In this paper, we present experiments with two autonomous mobile robots, which use autonomously learned subsymbolic representations of perceptual and behavioural space to determine goal-achieving sequences of actions. The experimental results we present illustrate that such an approach results in an embodied robot planner that produces plans which are grounded in the robot's perception-action space.
You can find me at my office located at the Informatics Forum of the University of Edinburgh.
I am at my office every day from 9:00 until 18:00, but you may consider emailing me to fix an appointment.