The precise simulation of the human brain requires coupling different models in order to cover the different physiological and functional aspects of this extremely complex organ. Each of this brain models is implemented following specific mathematical and programming approaches, potentially leading to diverging computational behaviour and requirements. Such situation is the typical use case that can benefit from the Modular Supercomputing Architecture (MSA), which organizes heterogeneous computing resources at system level. This architecture and its corresponding software environment enable to run each part of an application or a workflow on the best suited hardware.This paper presents the MSA concept covering current hardware and software implementations, and describes how the neuroscientific workflow resulting of coupling the codes NEST and Arbor is being prepared to exploit the MSA.
Recent research on face analysis has demonstrated the richness of information embedded in feature vectors extracted from a deep convolutional neural network. Even though deep learning achieved a very high performance on several challenging visual tasks, such as determining the identity, age, gender and race, it still lacks a well grounded theory which allows to properly understand the processes taking place inside the network layers. Therefore, most of the underlying processes are unknown and not easy to control. On the other hand, the human visual system follows a well understood process in analyzing a scene or an object, such as a face. The direction of the eye gaze is repeatedly directed, through purposively planned saccadic movements, towards salient regions to capture several details. In this paper we propose to capitalize on the knowledge of the saccadic human visual processes to design a system to predict facial attributes embedding a biologically-inspired network architecture, the HMAX. The architecture is tailored to predict attributes with different textural information and conveying different semantic meaning, such as attributes related and unrelated to the subject’s identity. Salient points on the face are extracted from the outputs of the S2 layer of the HMAX architecture and fed to a local texture characterization module based on LBP (Local Binary Pattern). The resulting feature vector is used to perform a binary classification on a set of pre-defined visual attributes. The devised system allows to distill a very informative, yet robust, representation of the imaged faces, allowing to obtain high performance but with a much simpler architecture as compared to a deep convolutional neural network. Several experiments performed on publicly available, challenging, large datasets demonstrate the validity of the proposed approach.
The human lateral geniculate body (LGB) with its six sickle shaped layers (lam) represents the principal thalamic relay nucleus for the visual system. Cytoarchitectonic analysis serves as the groundtruth for multimodal approaches and studies exploring its function. This technique, however, requires experienced knowledge about human neuroanatomy and is costly in terms of time. Here we mapped the six layers of the LGB manually in serial, histological sections of the BigBrain, a high-resolution model of the human brain, whereby their extent was manually labeled in every 30th section in both hemispheres. These maps were then used to train a deep learning algorithm in order to predict the borders on sections in-between these sections. These delineations needed to be performed in 1 µm scans of the tissue sections, for which no exact cross-section alignment is available. Due to the size and number of analyzed sections, this requires to employ high-performance computing. Based on the serial section delineations, high-resolution 3D reconstruction was performed at 20 µm isotropic resolution of the BigBrain model. The 3D reconstruction shows the shape of the human LGB and its sublayers for the first time at cellular precision. It represents a use case to study other complex structures, to visualize their shape and relationship to neighboring structures. Finally, our results could provide reference data of the LGB for modeling and simulation to investigate the dynamics of signal transduction in the visual system.
The exchange of ideas between computer science and statistical physics has advanced the understanding of machine learning and inference significantly. This interdisciplinary approach is currently regaining momentum due to the revived interest in neural networks and deep learning. Methods borrowed from statistical mechanics complement other approaches to the theory of computational and statistical learning. In this brief review, we outline and illustrate some of the basic concepts. We exemplify the role of the statistical physics approach in terms of a particularly important contribution: the computation of typical learning curves in student teacher scenarios of supervised learning. Two, by now classical examples from the literature illustrate the approach: the learning of a linearly separable rule by a perceptron with continuous and with discrete weights, respectively. We address these prototypical problems in terms of the simplifying limit of stochastic training at high formal temperature and obtain the corresponding learning curves.
The Human Brain Project (HBP) ( https://humanbrainproject.eu/ ) is a large-scale flagship project funded by the European Commission with the goal of establishing a research infrastructure for brain science. This research infrastructure is currently being realised and will be called EBRAINS ( https://ebrains.eu/ ). The wide ranging EBRAINS services for the brain research communities require diverse access, processing and storage capabilities. As a result, it will strongly rely on e-infrastructure services. The HBP led to the creation of Fenix ( https://fenix-ri.eu/ ), a collaboration of five European supercomputing centres, who are providing a set of federated e-infrastructure services to EBRAINS. The Fenix architecture has been designed to uniquely address the need for a wide spectrum of services, from high performance computing (HPC) to on-demand cloud technologies to identity and access federation, for facilitating ease of access and usage of distributed e-infrastructure resources. In this article we describe the underlying concepts for an audience of computational science end-users and developers of domain-specific applications, workflows and platforms services. To exemplify the use of Fenix, we will discuss selected use cases demonstrating how brain researchers can use the offered infrastructure services and describe how access to these resources can be obtained.
Nowadays, clinicians have multiple tools that they can use to stimulate the brain, by means of electric or magnetic fields that can interfere with the bio-electrical behaviour of neurons. However, it is still unclear which are the neural mechanisms that are involved and how the external stimulation changes the neural responses at network-level. In this paper, we have exploited the simulations carried out using a spiking neural network model, which reconstructed the cerebellar system, to shed light on the underlying mechanisms of cerebellar Transcranial Magnetic Stimulation affecting specific task behaviour. Namely, two computational studies have been merged and compared. The two studies employed a very similar experimental protocol: a first session of Pavlovian associative conditioning, the administration of the TMS (effective or sham), a washout period, and a second session of Pavlovian associative conditioning. In one study, the washout period between the two sessions was long (1 week), while the other study foresaw a very short washout (15 min). Computational models suggested a mechanistic explanation for the TMS effect on the cerebellum. In this work, we have found that the duration of the washout strongly changes the modification of plasticity mechanisms in the cerebellar network, then reflected in the learning behaviour.
In the last decade, Sentiment Analysis and Affective Computing have found applications in different domains. In particular, the interest of extracting emotions in healthcare is demonstrated by the various applications which encompass patient monitoring and adverse events prediction. Thanks to the availability of large datasets, most of which are extracted from social media platforms, several techniques for extracting emotion and opinion from different modalities have been proposed, using both unimodal and multimodal approaches. After introducing the basic concepts related to emotion theories, mainly borrowed from social sciences, the present work reviews three basic modalities used in emotion recognition, i.e. textual, audio and video, presenting for each of these i) some basic methodologies, ii) some among the widely used datasets for the training of supervised algorithms and iii) briefly discussing some deep Learning architectures. Furthermore, the paper outlines the challenges and existing resources to perform a multimodal emotion recognition which may improve performances by combining at least two unimodal approaches. architecture to perform multimodal emotion recognition.
The ‘BigBrain’ is a high-resolution data set of the human brain that enables three-dimensional (3D) analyses with a 20 µm spatial resolution at nearly cellular level. We use this data set to explore pre-α (cell) islands of layer 2 in the entorhinal cortex (EC), which are early affected in Alzheimer’s disease and have therefore been the focus of research for many years. They appear mostly in a round and elongated shape as shown in microscopic studies. Some studies suggested that islands may be interconnected based on analyses of their shape and size in two-dimensional (2D) space. Here, we characterized morphological features (shape, size, and distribution) of pre-α islands in the ‘BigBrain’, based on 3D-reconstructions of gapless series of cell-body-stained sections. The EC was annotated manually, and a machine-learning tool was trained to identify and segment islands with subsequent visualization using high-performance computing (HPC). Islands were visualized as 3D surfaces and their geometry was analyzed. Their morphology was complex: they appeared to be composed of interconnected islands of different types found in 2D histological sections of EC, with various shapes in 3D. Differences in the rostral-to-caudal part of EC were identified by specific distribution and size of islands, with implications for connectivity and function of the EC. 3D compactness analysis found more round and complex islands than elongated ones. The present study represents a use case for studying large microscopic data sets. It provides reference data for studies, e.g. investigating neurodegenerative diseases, where specific alterations in layer 2 were previously reported.
The study of the visual system of the brain has attracted the attention and interest of many neuro-scientists, that derived computational models of some types of neuron that compose it. These findings inspired researchers in image processing and computer vision to deploy such models to solve problems of visual data processing. In this paper, we review approaches for image processing and computer vision, the design of which is based on neuro-scientific findings about the functions of some neurons in the visual cortex. Furthermore, we analyze the connection between the hierarchical organization of the visual system of the brain and the structure of Convolutional Networks (ConvNets). We pay particular attention to the mechanisms of inhibition of the responses of some neurons, which provide the visual system with improved stability to changing input stimuli, and discuss their implementation in image processing operators and in ConvNets.
In recent years, Independent Component Analysis (ICA) has successfully been applied to remove noise and artifacts in images obtained from Three-dimensional Polarized Light Imaging (3D-PLI) at the mesoscale (i.e., 64 μm). Here, we present an automatic denoising procedure for gray matter regions that allows to apply the ICA also to microscopic images, with reasonable computational effort. Apart from an automatic segmentation of gray matter regions, we applied the denoising procedure to several 3D-PLI images from a rat and a vervet monkey brain section.
We are entering an age of ‘big’ computational neuroscience, in which neural network models are increasing in size and in numbers of underlying data sets. Consolidating the zoo of models into large-scale models simultaneously consistent with a wide range of data is only possible through the effort of large teams, which can be spread across multiple research institutions. To ensure that computational neuroscientists can build on each other’s work, it is important to make models publicly available as well-documented code. This chapter describes such an open-source model, which relates the connectivity structure of all vision-related cortical areas of the macaque monkey with their resting-state dynamics. We give a brief overview of how to use the executable model specification, which employs NEST as simulation engine, and show its runtime scaling. The solutions found serve as an example for organizing the workflow of future models from the raw experimental data to the visualization of the results, expose the challenges, and give guidance for the construction of an ICT infrastructure for neuroscience.
New computing technologies, such as affective computing and sentiment analysis, are raising a strong interest in different fields, such as marketing, politics and, recently, life sciences. Examples of possible applications in the last field, regard the detection and monitoring of depressive states or mood disorders and anxiety conditions. This paper aims to provide an introductory overview of affective computing and sentiment analysis, through the discussion of the main processing techniques and applications. The paper concludes with a discussion relative to a new approach based on the integration of sentiment analysis and affective computing to obtain a more accurate and reliable detection of emotions and feelings for applications in the life sciences.
Nowadays, much is known about the functions of the components of the human auditory system. Computational models of these components are widely accepted and recently inspired the work of researchers in pattern recognition and signal processing. In this work we present a novel filter, which we call COPE (Combination of Peaks of Energy), that is inspired by the way the sound waves are converted into neuronal firing activity on the auditory nerve. A COPE filter creates a model of the pattern of the neural activity generated by a sound of interest and is able to detect the same pattern and modified versions of it. We apply the proposed method on the task of event detection for surveillance of roads. For the experiments, we use a publicly available data set, namely the MIVIA road events data set. The results that we achieve (recognition rate equal to 94% and false positive rate lower than 4% ) and the comparison with existing methods demonstrate the effectiveness of the proposed bio-inspired filters for audio analysis.
Three-dimensional Polarized Light Imaging (3D-PLI) is a promising technique to reconstruct the nerve fiber architecture of human post-mortem brains from birefringence measurements of histological brain sections with micrometer resolution. To better understand how the reconstructed fiber orientations are related to the underlying fiber structure, numerical simulations are employed. Here, we present two complementary simulation approaches that reproduce the entire 3D-PLI analysis: First, we give a short review on a simulation approach that uses the Jones matrix calculus to model the birefringent myelin sheaths. Afterwards, we introduce a more sophisticated simulation tool: a 3D Maxwell solver based on a Finite-Difference Time-Domain algorithm that simulates the propagation of the electromagnetic light wave through the brain tissue. We demonstrate that the Maxwell solver is a valuable tool to better understand the interaction of polarized light with brain tissue and to enhance the accuracy of the fiber orientations extracted by 3D-PLI.
Contemporary simulation technology for neuronal networks enables the simulation of brain-scale networks using neuron models with a single or a few compartments. However, distributed simulations at full cell density are still lacking the electrical coupling between cells via so called gap junctions. This is due to the absence of efficient algorithms to simulate gap junctions on large parallel computers. The difficulty is that gap junctions require an instantaneous interaction between the coupled neurons, whereas the efficiency of simulation codes for spiking neurons relies on delayed communication. In a recent paper [15] we describe a technology to overcome this obstacle. Here, we give an overview of the challenges to include gap junctions into a distributed simulation scheme for neuronal networks and present an implementation of the new technology available in the NEural Simulation Tool (NEST 2.10.0). Subsequently we introduce the usage of gap junctions in model scripts as well as benchmarks assessing the performance and overhead of the technology on the supercomputers JUQUEEN and K computer.
Collaboration is an essential feature of human social interaction. Briefly, when two or more people agree on a common goal and a joint intention to reach that goal, they have to coordinate their actions to engage in joint actions, planning their courses of actions according to the actions of the other partners. The same holds for teams where the partners are people and robots, resulting on a collection of technical questions difficult to answer. Human-robot collaboration requires the robot to coordinate its behavior to the behaviors of the humans at different levels, e.g., the semantic level, the level of the content and behavior selection in the interaction, and low-level aspects such as the temporal dynamics of the interaction. This forces the robot to internalize information about the motions, actions and intentions of the rest of partners, and about the state of the environment. Furthermore, collaborative robots should select their actions taking into account additional human-aware factors such as safety, reliability and comfort. Current cognitive systems are usually limited in this respect as they lack the rich dynamic representations and the flexible human-aware planning capabilities needed to succeed in tomorrow human-robot collaboration tasks. Within this paper, we provide a tool for addressing this problem by using the notion of deep hybrid representations and the facilities that this common state representation offers for the tight coupling of planners on different layers of abstraction. Deep hybrid representations encode the robot and environment state, but also a robot-centric perspective of the partners taking part in the joint activity.
Primate cortices are organized into different layers which constitute a compartmental structure on a functional level. We show how composite structural elements form building blocks to define canonical elements for columnar computation in cortex. As a further abstraction, we define a dynamical three-stage model of a cortical column for processing that allows to investigate the dynamic response properties of cortical algorithms, e.g., feedforward signal integration as feature detection filters, lateral feature grouping, and the integration of modulatory (feedback) signals. Using such multi-stage cortical model, we investigate the detection and integration of spatio-temporal motion measured by event-based (frame-less) cameras. We demonstrate how the canonical neural circuit can improve such representations using normalization and feedback and develop key computational elements to map such a model onto neuromorphic hardware (IBM’s TrueNorth chip). This makes a step towards implementing real-time and energy-efficient neuromorphic optical flow detectors based on realistic principles of computation in cortical columns.
Although the domain of hard real-time systems has thoroughly been elaborated in academia, architectural issues did not receive the attention they deserve, and in most cases just off-the-shelf computer systems are used as execution platforms — with no guarantee that they are able to meet the temporal requirements specified. Therefore, a novel asymmetrical multiprocessor architecture for embedded control systems is presented. Resembling the structure of the human brain, it inherently supports temporal determinism.
Human brain atlases [1] are indispensable tools to achieve a better understanding of the multilevel organization of the brain through integrating and analyzing data from different brains, sources, and modalities while considering the functionally relevant topography of the brain [4]. The spatial resolution of most of these electronic atlases is in the range of millimeters, which does not allow the integration of the information at the level of cortical layers, columns, microcircuits or cells. Therefore, we introduced in 2013 the first BigBrain data set with a resolution of 20 m isotropic. This data set allows to specify morphometric parameters of human brain organization, which serve as a “gold standard” for neuroimaging data obtained at a lower spatial resolution. It provides, in addition, an essential basis for realistic brain models concerning structural analysis and simulation [2]. For the generation of other, even higher-resolution data sets of the human brain, we developed an improved and more efficient data processing workflow employing high performance computing to 3D reconstruct histological data sets. To facilitate the analysis of intersubject variability on a microscopic level, the new processing framework was applied for reconstructing a second BigBrain data set with 7676 sections. Efficient data processing of a large amount of data sets with a complex nested reconstruction workflow using large number of compute nodes required optimized distributed processing workflows as well as parallel programming. A detailed documentation of the processing steps and the complex inter-dependencies of the data sets at each level of the multi-step reconstruction workflow was essential to enable transformations to images of the same histological sections obtained with even higher spatial resolution. We have addressed these challenges, and achieved efficient high throughput processing of thousands of images of histological sections in combination with sufficient flexibility, based on an effective, successive coarse-to-fine hierarchical processing.
