AVÊÓƵ

Showcase of recent publications

Learning with reinforcement prediction errors in a model of the Drosophila mushroom body

James E. M. Bennett, Andrew Philippides, Thomas Nowotny, Learning with reinforcement prediction errors in a model of the Drosophila mushroom body. Nature Communications 12: 2569.

Effective decision making in a changing environment demands that accurate predictions are learned about decision outcomes. In Drosophila, such learning is orchestrated in part by the mushroom body, where dopamine neurons signal reinforcing stimuli to modulate plasticity presynaptic to mushroom body output neurons. Building on previous mushroom body models, in which dopamine neurons signal absolute reinforcement, we propose instead that dopamine neurons signal reinforcement prediction errors by utilising feedback reinforcement predictions from output neurons. We formulate plasticity rules that minimise prediction errors, verify that output neurons learn accurate reinforcement predictions in simulations, and postulate connectivity that explains more physiological observations than an experimentally constrained model. The constrained and augmented models reproduce a broad range of conditioning and blocking experiments, and we demonstrate that the absence of blocking does not imply the absence of prediction error dependent learning. Our results provide five predictions that can be tested using established experimental methods.

Larger GPU-accelerated brain simulations with procedural connectivity

Knight, J.C., Nowotny, T. Larger GPU-accelerated brain simulations with procedural connectivity. Nat Comput Sci (2021).

Simulations are an important tool for investigating brain function but large models are needed to faithfully reproduce the statistics and dynamics of brain activity. Simulating large spiking neural network models has, until now, needed so much memory for storing synaptic connections that it required high performance computer systems. Here, we present an alternative simulation method we call ‘procedural connectivity’ where connectivity and synaptic weights are generated ‘on the fly’ instead of stored and retrieved from memory. This method is particularly well suited for use on graphical processing units (GPUs)—which are a common fixture in many workstations. Using procedural connectivity and an additional GPU code generation optimization, we can simulate a recent model of the macaque visual cortex with 4.13 × 106 neurons and 24.2 × 109 synapses on a single GPU—a significant step forward in making large-scale brain modeling accessible to more researchers.

,

Evolved transistor array robot controllers

Garvie, Michael, Flascher, Ittai, Philippides, Andrew, Thompson, Adrian and Husbands, Phil (2020) Evolved transistor array robot controllers. Evolutionary Computation. pp. 1-32. ISSN 1063-6560

For the first time a field programmable transistor array (FPTA) was used to evolve robot control circuits directly in analog hardware. Controllers were successfully incrementally evolved for a physical robot engaged in a series of visually guided behaviours, including finding a target in a complex environment where the goal was hidden from most locations. Circuits for recognising spoken commands were also evolved and these were used in conjunction with the controllers to enable voice control of the robot, triggering behavioural switching. Poor quality visual sensors were deliberately used to test the ability of evolved analog circuits to deal with noisy uncertain data in realtime. Visual features were coevolved with the controllers to automatically achieve dimensionality reduction and feature extraction and selection in an integrated way. An efficient new method was developed for simulating the robot in its visual environment. This allowed controllers to be evaluated in a simulation connected to the FPTA. The controllers then transferred seamlessly to the real world. The circuit replication issue was also addressed in experiments where circuits were evolved to be able to function correctly in multiple areas of the FPTA. A methodology was developed to analyse the evolved circuits which provided insights into their operation. Comparative experiments demonstrated the superior evolvability of the transistor array medium.

Trait phenomenological control predicts experience of mirror synaesthesia and the rubber hand illusion

Lush, P, Botan, V, Scott, R B, Seth, A K, Ward, J and Dienes, Z (2020) Trait phenomenological control predicts experience of mirror synaesthesia and the rubber hand illusion. Nature, 11. a4853 1-10. ISSN 0028-0836

In hypnotic responding, expectancies arising from imaginative suggestion drive striking experiential changes (e.g., hallucinations) — which are experienced as involuntary — according to a normally distributed and stable trait ability (hypnotisability). Such experiences can be triggered by implicit suggestion and occur outside the hypnotic context. In large sample studies (of 156, 404 and 353 participants), we report substantial relationships between hypnotisability and experimental measures of experiential change in mirror-sensory synaesthesia and the rubber hand illusion comparable to relationships between hypnotisability and individual hypnosis scale items. The control of phenomenology to meet expectancies arising from perceived task requirements can account for experiential change in psychological experiments.

Learning action-oriented models through active inference

Tschantz, Alexander, Seth, Anil K and Buckley, Christopher L (2020) Learning action-oriented models through active inference. PLoS Computational Biology, 16 (4). pp. 1-30. ISSN 1553-734X

Converging theories suggest that organisms learn and exploit probabilistic models of their environment. However, it remains unclear how such models can be learned in practice. The open-ended complexity of natural environments means that it is generally infeasible for organisms to model their environment comprehensively. Alternatively, action-oriented models attempt to encode a parsimonious representation of adaptive agent-environment interactions. One approach to learning action-oriented models is to learn online in the presence of goal-directed behaviours. This constrains an agent to behaviourally relevant trajectories, reducing the diversity of the data a model need account for. Unfortunately, this approach can cause models to prematurely converge to sub-optimal solutions, through a process we refer to as a bad-bootstrap. Here, we exploit the normative framework of active inference to show that efficient action-oriented models can be learned by balancing goal-oriented and epistemic (information-seeking) behaviours in a principled manner. We illustrate our approach using a simple agent-based model of bacterial chemotaxis. We first demonstrate that learning via goal-directed behaviour indeed constrains models to behaviorally relevant aspects of the environment, but that this approach is prone to sub-optimal convergence. We then demonstrate that epistemic behaviours facilitate the construction of accurate and comprehensive models, but that these models are not tailored to any specific behavioural niche and are therefore less efficient in their use of data. Finally, we show that active inference agents learn models that are parsimonious, tailored to action, and which avoid bad bootstraps and sub-optimal convergence. Critically, our results indicate that models learned through active inference can support adaptive behaviour in spite of, and indeed because of, their departure from veridical representations of the environment. Our approach provides a principled method for learning adaptive models from limited interactions with an environment, highlighting a route to sample efficient learning algorithms.

Mushroom bodies are required for learnt visual navigation, but not for innate visual behaviour, in ants

Buehlmann, Cornelia, Wozniak, Beata, Goulard, Roman, Webb, Barbara, Graham, Paul and Niven, Jeremy E (2020) Mushroom bodies are required for learnt visual navigation, but not for innate visual behaviour, in ants. Current Biology, 30. pp. 1-6. ISSN 0960-9822

Visual navigation in ants has long been a focus of experimental study [1, 2, 3], but only recently have explicit hypotheses about the underlying neural circuitry been proposed [4]. Indirect evidence suggests the mushroom bodies (MBs) may be the substrate for visual memory in navigation tasks [5, 6, 7], while computational modeling shows that MB neural architecture could support this function [8, 9]. There is, however, no direct evidence that ants require MBs for visual navigation. Here we show that lesions of MB calyces impair ants’ visual navigation to a remembered food location yet leave their innate responses to visual cues unaffected. Wood ants are innately attracted to large visual cues, but we trained them to locate a food source at a specific angle away from such a cue. Subsequent lesioning of the MB calyces using procaine hydrochloride injection caused ants to revert toward their innate cue attraction. Handling and saline injection control ants still approached the feeder. Path straightness of lesioned and control ants did not differ from each other but was lower than during training. Reversion toward the cue direction occurred irrespective of whether the visual cue was ipsi- or contralateral to the lesion site, showing this is not due simply to an induced motor bias. Monocular occlusion did not diminish ants’ ability to locate the feeder, suggesting that MB lesions are not merely interrupting visual input to the calyx. The demonstrated dissociation between innate and learned visual responses provides direct evidence for a specific role of the MB in navigational memory.

Individual differences in the tendency to see the expected

Andermane, Nora, Bosten, Jenny M, Seth, Anil K and Ward, Jamie (2020) Individual differences in the tendency to see the expected. Consciousness and Cognition, 85. a102989 1-20. ISSN 1053-8100

Research has established that prior knowledge of visual stimuli facilitates their entry into awareness. We adopted an individual differences approach to explore whether a tendency to ‘see the expected’ is general or method-specific. We administered a binocular rivalry task and manipulated selective attention, as well as induced expectations via predictive context, self-generated imagery, expectancy cues, and perceptual priming. Most prior manipulations led to a facilitated awareness of the biased percept in binocular rivalry, whereas strong signal primes led to a suppressed awareness, i.e., adaptation. Correlations and factor analysis revealed that the facilitatory effect of priors on visual awareness is closely related to attentional control. We also investigated whether expectation-based biases predict perceptual abilities. Adaptation to strong primes predicted improved naturalistic change detection and the facilitatory effect of weak primes predicted the experience of perceptual anomalies. Taken together, our results indicate that the facilitatory effect of priors may be underpinned by an attentional mechanism but the tendency to ‘see the expected’ is method-specific.

Temporal ordering of input modulates connectivity formation in a developmental neuronal network model of the cortex

Hartley, Caroline, Farmer, Simon and Berthouze, Luc (2020) Temporal ordering of input modulates connectivity formation in a developmental neuronal network model of the cortex. PLoS ONE, 15 (1). e0226772. ISSN 1932-6203

Preterm infant brain activity is discontinuous; bursts of activity recorded using EEG (electroencephalography), thought to be driven by subcortical regions, display scale free properties and exhibit a complex temporal ordering known as long-range temporal correlations (LRTCs). During brain development, activity-dependent mechanisms are essential for synaptic connectivity formation, and abolishing burst activity in animal models leads to weak disorganised synaptic connectivity. Moreover, synaptic pruning shares similar mechanisms to spike-timing dependent plasticity (STDP), suggesting that the timing of activity may play a critical role in connectivity formation. We investigated, in a computational model of leaky integrate-and-fire neurones, whether the temporal ordering of burst activity within an external driving input could modulate connectivity formation in the network. Connectivity evolved across the course of simulations using an approach analogous to STDP, from networks with initial random connectivity. Small-world connectivity and hub neurones emerged in the network structure—characteristic properties of mature brain networks. Notably, driving the network with an external input which exhibited LRTCs in the temporal ordering of burst activity facilitated the emergence of these network properties, increasing the speed with which they emerged compared with when the network was driven by the same input with the bursts randomly ordered in time. Moreover, the emergence of small-world properties was dependent on the strength of the LRTCs. These results suggest that the temporal ordering of burst activity could play an important role in synaptic connectivity formation and the emergence of small-world topology in the developing brain.

Brian2GeNN: accelerating spiking neural network simulations with graphics hardware

Stimberg, M., Goodman, D.F.M. & Nowotny, T. Brian2GeNN: accelerating spiking neural network simulations with graphics hardware. Sci Rep 10, 410 (2020).

“Brian” is a popular Python-based simulator for spiking neural networks, commonly used in computational neuroscience. GeNN is a C++-based meta-compiler for accelerating spiking neural network simulations using consumer or high performance grade graphics processing units (GPUs). Here we introduce a new software package, Brian2GeNN, that connects the two systems so that users can make use of GeNN GPU acceleration when developing their models in Brian, without requiring any technical knowledge about GPUs, C++ or GeNN. The new Brian2GeNN software uses a pipeline of code generation to translate Brian scripts into C++ code that can be used as input to GeNN, and subsequently can be run on suitable NVIDIA GPU accelerators. From the user’s perspective, the entire pipeline is invoked by adding two simple lines to their Brian scripts. We have shown that using Brian2GeNN, two non-trivial models from the literature can run tens to hundreds of times faster than on CPU.

Multimodal interactions in insect navigation

Buehlmann, Cornelia, Mangan, Michael and Graham, Paul (2020) Multimodal interactions in insect navigation. Animal Cognition. ISSN 1435-9448

Animals travelling through the world receive input from multiple sensory modalities that could be important for the guidance of their journeys. Given the availability of a rich array of cues, from idiothetic information to input from sky compasses and visual information through to olfactory and other cues (e.g. gustatory, magnetic, anemotactic or thermal) it is no surprise to see multimodality in most aspects of navigation. In this review, we present the current knowledge of multimodal cue use during orientation and navigation in insects. Multimodal cue use is adapted to a species’ sensory ecology and shapes navigation behaviour both during the learning of environmental cues and when performing complex foraging journeys. The simultaneous use of multiple cues is beneficial because it provides redundant navigational information, and in general, multimodality increases robustness, accuracy and overall foraging success. We use examples from sensorimotor behaviours in mosquitoes and flies as well as from large scale navigation in ants, bees and insects that migrate seasonally over large distances, asking at each stage how multiple cues are combined behaviourally and what insects gain from using different modalities.

Network inference from population-level observation of epidemics

Di Lauro, F, Croix, J -C, Dashti, M, Berthouze, L and Kiss, I Z (2020) Network inference from population-level observation of epidemics. Scientific Reports, 10. a18779 1-14. ISSN 2045-2322

Using the continuous-time susceptible-infected-susceptible (SIS) model on networks, we investigate the problem of inferring the class of the underlying network when epidemic data is only available at population-level (i.e., the number of infected individuals at a finite set of discrete times of a single realisation of the epidemic), the only information likely to be available in real world settings. To tackle this, epidemics on networks are approximated by a Birth-and-Death process which keeps track of the number of infected nodes at population level. The rates of this surrogate model encode both the structure of the underlying network and disease dynamics. We use extensive simulations over Regular, ErdÅ‘s–Rényi and Barabási–Albert networks to build network class-specific priors for these rates. We then use Bayesian model selection to recover the most likely underlying network class, based only on a single realisation of the epidemic. We show that the proposed methodology yields good results on both synthetic and real-world networks.

Odor Stimuli: Not Just Chemical Identity

Pannunzi, Mario and Nowotny, Thomas (2019) Odor stimuli: not just chemical identity. Frontiers in Physiology, 10. ISSN 1664-042X

In most sensory modalities the underlying physical phenomena are well understood, and stimulus properties can be precisely controlled. In olfaction, the situation is different. The presence of specific chemical compounds in the air (or water) is the root cause for perceived odors, but it remains unknown what organizing principles, equivalent to wavelength for light, determine the dimensions of odor space. Equally important, but less in the spotlight, odor stimuli are also complex with respect to their physical properties, including concentration and time-varying spatio-temporal distribution. We still lack a complete understanding or control over these properties, in either experiments or theory. In this review, we will concentrate on two important aspects of the physical properties of odor stimuli beyond the chemical identity of the odorants: (1) The amplitude of odor stimuli and their temporal dynamics. (2) The spatio-temporal structure of odor plumes in a natural environment. Concerning these issues, we ask the following questions: (1) Given any particular experimental protocol for odor stimulation, do we have a realistic estimate of the odorant concentration in the air, and at the olfactory receptor neurons? Can we control, or at least know, the dynamics of odorant concentration at olfactory receptor neurons? (2) What do we know of the spatio-temporal structure of odor stimuli in a natural environment both from a theoretical and experimental perspective? And how does this change if we consider mixtures of odorants? For both topics, we will briefly summarize the underlying principles of physics and review the experimental and theoretical Neuroscience literature, focusing on the aspects that are relevant to animals’ physiology and behavior. We hope that by bringing the physical principles behind odor plume landscapes to the fore we can contribute to promoting a new generation of experiments and models.