Search

We live in a technological generation surrounded by interconnected sensors that can collect and distribute immense amounts of data on a daily basis. These data would have a better connotation and would have been more practical if sensor-based networks allowed us to capture and monitor the characteristics of physical objects from a highly dynamic environment. At this point, sensor-based networks could substantially enhance their applicability if machines process and interpret vast amounts of data correctly, an essential characteristic of scalable and interoperable wireless sensor network architectures. Through this research project, a) We will identify and evaluate wireless sensor network architectures enabling ap- plications from a highly dynamic environment. Then a data architecture will be proposed to enhance machine-to-machine (M2M) communication and human understanding, considering the issues and challenges of sensor networks. The future proposed data architecture will overcome the existing data frameworks' limitations identified in the literature review. b) The significant contribution of the research is to propose energy-efficient data collection models (the first layer of data architecture) that will reduce data transmissions using prediction models between nodes in sensor networks. The proposed models intend to predict values at the sink node using coefficients built and transmitted by sensor platforms. The goal is to build models that improve the energy of battery-powered sensory devices by reducing data transmissions and recovering values at sink nodes using the same coefficients of models while ensuring data integrity. Furthermore, the models are evaluated using real data sets from real sensor networks with the following metrics; RMSE, MAE, MSE, data reduction percentage, and energy savings.

Animals adapt their systems to optimise for different competing goals at the same time. Ideally, they will reach an optimal state of equilibrium where the outcome from any goal cannot get better without at the same time making another worse off, similar to the state of Pareto optimaility (Mock 2011). Animals can seek different goals like, to maintain their systems' stability and robustness, or improving their performances in a given computational task, which is reflected in their memory capacity and ability to make more rewarding decisions. Many species are capable of forming associative memories, they can learn to contextualise sensory stimuli as good, bad or neutral, when they are associated by a shortly upcoming salient outcome and bias their behaviours to approach or avoid these cues in the future. In this work I will focus on modelling the associative learning in the mushroom body circuit of the fruit fly, its center of olfactory associative learning. Flies can learn to associate an odor (sensory experience) with an appetitive or aversive outcome. They do so by modifying the connections between the mushroom body intrinsic neurons, called Kenyon cells (KCs), and their downstream mushroom body output neurons (MBONs). The fly motor behaviour was found to be biased by the activity of the MBONs to either approach or avoid an odor (Aso et al. 2014). Although many studies uncovered the molecular mechanisms and the neurons underpinning associative learning in different species, there has been no work done to answer some specific questions: (a) Why do the neurons in the same circuit within the same animal exhibit variability among each others in their intrinsic properties? It is unknown how variability among the same types of neurons in the same circuit and animal would eventually affect the animal's optimal behaviour in a computational task. Even previous studies that tackled inter-neuronal variability were trying to study its effect on circuits stability and were dealing with inter-neuronal variability across animals and not within an individual circuit (Marder and Goaillard 2006; Golowasch et al. 2002; Schulz, Goaillard, and Marder 2006; Schulz, Goaillard, and Marder 2007). Can the observed inter-neuronal variability be a result of some optimisation protocol that enhances the circuit computational performance, for example, memory or data performance? Or has it just happened at random? (b) Learning in the cerebellum (and its alike structures in other animals like the fruit fly mushroom body) happen by long term depression (weakening) between its intrinsic neurons -encoding the sensory input- and the downstream neurons that guide the animal's motor behaviour (Ito 1989). Like in (a), I ask if this learning rule has been conserved across species for optimising some computational aspects of learning In this 3 Chapters thesis, I will present a computational model of associative learning in the fruit fly mushroom body using realistic input odors statistics, as well as putting some constraints on the model network that were observed experimentally in the real mushroom body (e.g. the level of KCs sparse coding, the level of KCs sparse coding when their inhibitory inputs are silenced). In Chapter 2, I will answer the first question, the first aim, of this thesis and show that random variability between the KCs in their intrinsic parameters will impair the fly's memory performance. I find that the random inter-KCs variability will result in a high variability among the neurons in their sparsity values, which results in very few neurons being specifically active for some odors whilst the vast majority are activated by all incoming odors, that reduces the fly's ability to distinguish between odors and their identity as 'good' rewarded or 'bad' punished odors. However, I show that compensatory variability mechanisms will rescue the memory performance. I present 4 different models (activity-independent and activity-dependent rules) for how this compensatory variability can take place in real neurons. Last but not least, I show that the data from the newly released fly connectome actually reveal compensatory variability in the KCs which agree with my models' predictions. In Chapter 3, I will answer the second question in this thesis and show that, under some conditions, learning by depression can be more optimal than by potentiation. I will show that if the fly's decision making policy integrates the information from the MBONs in a divisive normalisation like manner (I explain more about divisive normalisation in Chapter 3), then learning by depression will lead to a higher memory performance. I also suggest a biologically plausible implementation for this normalisation decision policy using a winner-take-all (WTA) circuit model. I predict that in a WTA circuit that integrates the MBONs outputs, the fly's memory performance will be higher under learning by depression than under potentiation if the noise in the MBONs responses is of multiplicative nature (that is, if the noise in the MBONs responses across different trials is higher at higher MBONs firing rates).

Despite the recent success of artificial neural networks in solving complex tasks and achieving superhuman performance in a variety of tasks, the ability of biological brains to generalise the learnt knowledge and to quickly adapt to novel situations remains unmatched. This gap between specialised models and behavioural flexibility demonstrated by biological systems encourages researchers to weigh biological plausibility in the process of formulating machine learning models. In this sense, works in the broad field of optimisation and learning lie on a spectrum whose extremes have two diverse research approaches: the first can be described as a virtuous race toward higher performance and more challenging applications, often met at the cost of interpretability of the model and biological plausibility; the second tries to replicate how biological systems operate at a functional scale of modelling detail and gives higher priority to the system understanding rather than performance measures. We believe that the present work lies in between these complementary approaches, also definable as machine learning and computational neuroscience respectively. In this thesis, we will take inspiration from biology to develop machine learning models, but without the certainty that the formulated systems fall within the limits dictated by biological plausibility. Of course, the novel bio-inspired models are required to have practical advantages, measured in terms of performance, interpretability or computational cost, in comparison to pre-existing models. While an example of this line of thought can be found in the first two papers reported in the thesis, the third publication reported arises from a different, but complementary reasoning process. In the latter, we first formulated theoretically the model from abstract and desirable principles and then demonstrated its ability to explain neuroscientific, experimental findings. In summary, the connection, or flow of information, between the neuroscientific and machine learning fields is bi-directional across the thesis.

In complex natural environments, sensory systems are constantly exposed to a large stream of inputs. In order to encode the everchanging sensory surroundings efficiently, sensory systems rapidly adapt their responses to common stimuli that are parts of predictable patterns while retain their responsiveness to the novel, unexpected stimuli that usually contain behaviourally important information. Such a differential processing capacity is commonly referred to as novelty detection, whose neural signatures have been identified in human electroencephalogram (EEG) recording as mismatch negativity (MMN) and single-neuron electrophysiology as stimulus-specific adaptation (SSA). Despite the prevalence of the rare-selective responses in several sensory modalities, the intracellular and network mechanisms underlying are not yet fully understood. In this thesis, we employ the modeling tools in the computational neuroscience community, combined with computer simulation, to investigate the neuronal network dynamics producing novelty-detecting signals in the auditory and somatosensory cortex of the rodent model. Following an earlier auditory SSA mechanism in recurrent network with synaptic depression [Yarden and Nelken, 2017], we provide further support for the SSA circuitry hypothesis by using for the first time a similar recurrent network structure yet mediated with neuronal adaptation implemented in both simulated mean-field and physical neuromorphic spiking models. Our results suggest that SSA and its pertinent properties could arise from the differential excitation of tuned neuronal populations via propagation of population activity and short-term adaptation mechanism (neuronal adaptation or synaptic depression). In addition, we adapted and expanded this cortical circuit of the auditory SSA to a multi-scale thalamocortical network model that explains and reproduces physiologically observed neuronal response patterns that demonstrate signatures of both SSA and MMN in the somatosensory cortex [Musall et al., 2017]. Specifically, our results indicate that the novelty signal arises from the complex recurrent interplay between thalamic neurons and cortical neurons in layers 4 and 6. This work therefore provides a concrete mechanism that can serve as a starting point for further investigating the neural circuit mechanisms underlying novelty detection.

Attention mechanisms for context modelling are becoming ubiquitous in neural architectures in machine learning. The attention mechanism is a technique that filters out information that is irrelevant to a given task and focuses on learning task-dependent fixation points or regions. Furthermore, attention mechanisms suggest a question about a given task, i.e. 'what' to learn and 'where/how' to learn for task-specific context modelling. The context is the conditional variables instrumental in deciding the categorical distribution for the given data. Also, why is learning task-specific context necessary? In order to answer these questions, context modelling with attention in the vision and emotion domains is explored in this thesis using attention mechanisms with different hierarchical structures. The three main goals of this thesis are building superior classifiers using attention-based deep neural networks~(DNNs), investigating the role of context modelling in the given tasks, and developing a framework for interpreting hierarchies and attention in deep attention networks. In the vision domain, gesture and posture recognition tasks in diverse environments, are chosen. In emotion, visual and speech emotion recognition tasks are chosen. These tasks are selected for their sequential properties for modelling a spatiotemporal context. One of the key challenges from a machine learning standpoint is to extract patterns which bear maximum correlation with the information encoded in its signal while being as insensitive as possible to other types of information carried by the signal. A possible way to overcome this problem is to learn task-dependent representations. In order to achieve that, novel spatiotemporal context modelling networks and the mixture of multi-view attention~(MOMA) networks are proposed using bidirectional long-short-term memory network (BLSTM), convolutional neural network~(CNN), Capsule and attention networks. A framework has been proposed to interpret the internal attention states with respect to the given task. The results of the classifiers in the assigned tasks are compared with the 'state-of-the-art' DNNs, and the proposed classifiers achieve superior results. The context in speech emotion recognition is explored deeply with the attention interpretation framework, and it shows that the proposed model can assign word importance based on acoustic context. Furthermore, it has been observed that the internal states of the attention bear correlation with human perception of acoustic cues for speech emotion recognition. Overall, the results demonstrate superior classifiers and context learning models with interpretable frameworks. The findings are very important for speech emotion recognition systems. In this thesis, not only better models are produced, but also the interpretability of those models are explored, and their internal states are analysed. The phones and words are aligned with the attention vectors, and it is seen that the vowel sounds are more important for defining emotion acoustic cues than the consonants, and the model can assign word importance based on acoustic context. Also, how these approaches for emotion recognition using word importance for predicting emotions are demonstrated by the attention weight visualisation over the words. In a broader perspective, the findings from the thesis about gesture, posture and emotion recognition may be helpful in tasks like human-robot interaction~(HRI) and conversational artificial agents (such as Siri, Alexa). The communication is grounded with the symbolic and sub-symbolic cues of intent either from visual, audio or haptics. The understanding of intent is much dependent on the reasoning about the situational context. Emotion, i.e. speech and visual emotion, provides context to a situation, and it is a deciding factor in the response generation. Emotional intelligence and information from vision, audio and other modalities are essential for making human-human and human-robot communication more natural and feedback-driven.

Behavioural neuroscience uses a variety of animals to model human diseases, test novel drugs and study how certain factors affect or alter natural behaviour. In its arsenal, it contains a number of different experimental procedures involving navigation and locomotion tasks inside constrained environments and a number of analysis techniques to draw conclusions about various aspects of neuroscience such as the development of learning and memory. With the advancements in technology and the rise of artificial intelligence in many areas of our society, machine learning algorithms and applications are commonly used to draw, with limited user interaction and in a speedy manner, as much intelligence as possible from collections of data. Machine learning has greatly boosted behavioural neuroscience research but in many cases it provides experiment-specific analysis methods requiring domain knowledge in order to be used. This work addresses the first limitation of experiment-specific analysis methods by bringing an integration of common metrics used in different experimental procedures involving path analysis. For the second limitation it proposes a machine learning agnostic framework for data analysis in a common experimental procedure called Morris Water Maze which can also be used to other experiments involving behavioural categorisation tasks. In addition, it proposes a novel machine learning method for detailed analysis of locomotion that can be applied to any navigation task for both automatic categorisation and pattern recognition tasks. Other objectives of this study are to present detailed benchmarks of machine learning techniques that can be used for data analytics in behavioural neuroscience and to expand the usability of the methods it presents by making them easy to use by the research community. For this reason, all the source codes of the presented algorithms and pipelines is publicly available and, when applicable, graphical user interfaces or software tools have been engineered to help executing them.

Individuals use their cognitive abilities to make decisions, with the ultimate goal of improving their status. Decisions outcomes are used to learn the association between the decisions which lead to good results and those resulting in punishing outcomes. These associations might not be easily inferable because of environmental complexity or noisy feedback. Tasks in which outcomes probabilities are known are termed "decisions under risk". Researchers have consistently showed that people are risk averse when choosing among options featuring gains, while they are risk seeking when making decisions about options featuring losses. When the probabilities of the options are not clearly stated the task is known as "decisions under ambiguity". In this type of task individuals face an exploration-exploitation trade off: to maximise their profit they need to choose the best option but at the same time they need to discover which option leads to the best outcome by trial-and-error. The process of knowledge acquisition by interaction with the environment is called adaptive learning. Evidence from literature points in the direction of unskilled investors behaviour being consistent with naive reinforcement learning, simply adjusting their preference for which option to choose based on its recent outcomes. Experimental data from a binary choice task and a quasi-field scenario is used to test a combination of Reinforcement Learning and Prospect Theory. Both the investigations include reinforcement learning models featuring specific parameters which can be tuned to describe individual learning decision-making strategies. The first part is focused on integrating the two computational models, the second on testing it on a more realistic scenario. The results indicate that the combination of Reinforcement Learning and Prospect Theory could be a descriptive account of decision- making in binary decision tasks. A two-state space configuration, together with a non- saturating reward function appears to be the best setup to capture behaviour in said task. Moreover, analysing the parameters of the models it becomes evident that payoff variability has an impact on speed of learning and randomness of choice. The same modelling approach fails to capture behaviour in a more complex task, indicating that more complex models might be needed to provide a computational account of decisions from experience in non-trivial tasks.

Due to technological advances the amount of data that can be collected in modern science is increasing every day and neuroscience is no exception. Integrating large amounts of data at different spatial and temporal scales is essential for uncovering the underlying mechanisms of the brain but poses also new challenges since drawing conclusions from vast amounts of data is increasingly difficult. New automated and fast analysis methods that can make sense of large and complex data sets are therefore in need and will become increasingly important in the years and decades ahead. This work proposes new tools for the analysis of two important types of data commonly found in neuroscience. The first is behavioural data from rodent navigation tasks in the form of animal movement paths. Two novel classification methods based on machine learning algorithms are proposed here. The methods are able to automatically or semi-automatically reduce the complex movement paths of the animals to a series of stereotypical types of behaviour, leading toboth more detailed and consistent results. The second type of data considered here is electrophysiological data, in the form of extracellular multielectrode array (MEA) recordings which can record the electrical activity of thousands of neurons in parallel over long periods of time. Here a new highly-parallel data processing tool which can reduce the MEA data to a series of spike trains is presented. The tool can serve as basis for more sophisticated analyses like the reconstruction of the individual cell spike trains, for which machine learning methods are again essential. The results presented here show that machine learning algorithms and parallel processing architectures are both fundamental tools for coping with large and complex data sets, like the ones found in modern neuroscience.

Decision-making is an integral part of everyday life for animals of all species. Some decisions are rapid and based on sensory input alone, others rely on factors such as context and internal motivation. The possibilities for the experimental investigation of choice behaviour in mammals, especially in humans, are seemingly endless. However, neuroscience has struggled to define the neural circuitry behind decision-making processes due to the complex structure of the mammalian brain. For this work we turn to the honeybee for inspiration. With a brain composed of approximately one million neurons and sized at a tiny 1mm3, it may be assumed that such an insect produces mere `programmed' behaviours, yet, the honeybee exhibits a rich, elaborate behavioural repertoire and a large capacity for learning in a variety of different paradigms. Indeed, the honeybee has been identified as a powerful model for decision-making. Sequential sampling models, originating in psychology, have been used to explain rapid decision-making behaviours. Such models assume that noisy sensory evidence is integrated over time until a threshold is reached, whereby a decision is made. These models have proven popular because they are able to fit biological data and are furthermore supported by neural evidence. Additionally, they explain the speed-accuracy trade-off, a behavioural phenomenon also demonstrated in bees. For this work we examine honeybee choice behaviour in different levels of satiation, and show that hungry bees are faster and less accurate than partially satiated bees in a simple choice task. We suggest that differences in choice behaviour may be attributed to a simple mechanism which alters the level of the decision threshold according to how satiated the bee is. We further speculate that the honeybee olfactory system may be a drift-diffusion channel, and develop a simple computational model, based on honeybee neurobiology, with simulations that match behavioural results.

Virtual environments and serious games are a popular media used to fulfil a variety of purposes, including teaching and learning. The former are computer-generated environments that present three-dimensional spatial representations. The latter are games which fundamentally combine virtual environments and gamification, and are used for objectives other than pleasure and pure entertainment. Most recent works have investigated the effect of virtual reality technology on learners, sense of presence and so on. However, less investigated is the relationship between the achievement of learning objectives and the knowledge delivery methods utilised (knowledge representations and instruction modalities); or the effect of technology-enhanced learning on learner's mood after the intervention (and consequently, the learning) and whether there is any gender difference; or finally, the transfer of knowledge from virtual to the real world and its long-term retention; which are all elements investigated in this research. Two studies of procedural learning were conducted to investigate the elements highlighted above. The first study investigated on the requirements of a three-dimensional virtual environment as compared to Google Street View, instruction modalities such as textual, phone and companion, and short-term memory to learn a new route. The findings show that the virtual environment is better than Google Street View according to users' experience and having a companion to the task is a better instruction modality for route learning. The second study focused on ritual learning that is the case study of the research. The study investigated on the efficiency of a serious game as compared to PowerPoint note, collaboration with and without a coach, memory recall between short and long-term period and gender differences. The findings indicate that the serious game is better than PowerPoint note according to users' self-reported score and having the coach improves users' learning efficiency and moods. Also, knowledge of landmarks representations remains longer in users' memory if learnt from the serious game and factual knowledge remains longer in users' memory if learnt with the coach. Considering gender differences, women feel that the task is more enjoyable if learning takes place with the companion, and they recall more landmarks than men, whereas men take less time to complete the task. Apart from that, the ritual and navigation knowledge acquired in a virtual environment or a serious game can be competently used in reality, and it encourages users to remember more landmarks. The further findings from both studies also reveal that navigation in the virtual environment and serious game improve users' overall mood and happiness and women with improved happiness after the virtual training increase their learning performance. Also, younger players improve learning performance after learning in a virtual environment and serious game. To sum up, virtual environments and serious games can be used as a delivery method for procedural learning, in particular for ritual learning as they induce enjoyment, create an interesting experience and stimulate learning performance. Both representations also encourage landmarks memorization. Also, collaborative learning, in particular with a coach, is always the best method to convey, share and understand the knowledge. Finally, women enjoy learning with a companion.

Experimental research over the last two decades has shown that the anatomical connectivity among neurons is largely non-random across brain areas. This complex organisation shapes the flow of information, giving rise to specific pathways and motifs, which are ultimately responsible for processes like emotions, cognitive functions and behaviour, just to mention few. Due to the spectacular progress of technology, the study of the brain wiring diagram, known as connectomics, has received considerable attention in recent years, resulting in the proliferation of large data sets. From one side, this adds a significant contribution towards a better understanding of the complex processes that take place in the brain. On the other side, however, analysing such large connectivities is a hard task that has not yet found a satisfactory solution. Particular evidence has been found for bidirectional motifs,occurring when two neurons project onto each other via connections of equal strength, and unidirectional motifs, when one of the two connections is dominant. These specific motifs were found to correlate with short-term synaptic plasticity properties, which are related to resources availability for signal transmission. The aim of this thesis is to add a contribution to the ongoing efforts spent on answering the two main questions related to motif evidence: How can we satisfactory detect and measure motifs in large networks and why do they have the characteristics that we observe? Following existing literature, we hypothesise that bidirectional and unidirectional motifs appear as a consequence of learning processes, which move the distribution of the synaptic connections away from randomness through activity dependent synaptic plasticity. Based on this, we introduce a symmetry measure for global connectivity and a statistics-based heuristic algorithm for directed and weighted graphs that is able to detect overlapping bidirectional communities within large networks. On the other side, to address the why question we introduce an error-driven learning framework for short-term plasticity that acts jointly with Spike-Timing Dependent Plasticity, a well-known learning mechanism for long-term plasticity: By allowing synapses to change their properties,neurons are able to adapt their own activity depending on an error signal. This results in more rich dynamics and also, provided that the learning mechanism is target-specific, leads to specialised groups of synapses projecting onto functionally different targets, qualitatively replicating the experimental results of Wang and collaborators in 2006.

Although the basal ganglia have been widely studied and implicated in signal processing and action selection, little information is known about the active role that the striatal microcircuit plays in action selection in the basal ganglia-cortical-thalamic loops. To address this knowledge gap we use a large scale three dimensional spiking model of the striatum, combined with a rate coded model of the basal ganglia-cortical-thalamic loop, to asses the computational role the striatum plays in action selection. We identify robust transient phenomena generated by the striatal microcircuit, which temporarily enhances the difference between two competing cortical inputs. We show that this transient is sufficient to modulate decision making in the basal ganglia-thalamo-cortical circuit. We also find that the transient selection originates from a novel adaptation effect in single striatal projection neurons, which is amenable to experimental testing. Finally, we compared transient selection with models implementing classical steady-state selection. We challenged both forms of model to account for recent reports of paradoxically enhanced response selection in Huntington's disease patients. We found that steady-state selection was uniformly impaired under all simulated Huntington's conditions, but transient selection was enhanced given a sufficient Huntington's-like increase in NMDA receptor sensitivity. I propose a mechanistic underpinning to a novel neural compensatory mechanism, responsible for improved cognition in severe neuro-degeneration. Thus, our models provide an intriguing hypothesis for the mechanisms underlying the paradoxical cognitive improvements in manifest Huntington's patients, which is consistent with recent behavioural data.