Fast Quaternion Product Units for Learning Disentangled Representations in SO(3)

Real-world 3D structured data like point clouds and skeletons often can be represented as data in a 3D rotation group (denoted as $\mathbb{SO}(3)$). However, most existing neural networks are tailored for the data in the Euclidean space, which makes the 3D rotation data not closed under their algebraic operations and leads to sub-optimal performance in 3D-related learning tasks. To resolve the issues caused by the above mismatching between data and model, we propose a novel non-real neuron model called \textit{quaternion product unit} (QPU) to represent data on 3D rotation groups. The proposed QPU leverages quaternion algebra and the law of the 3D rotation group, representing 3D rotation data as quaternions and merging them via a weighted chain of Hamilton products. We demonstrate that the QPU mathematically maintains the $\mathbb{SO}(3)$ structure of the 3D rotation data during the inference process and disentangles the 3D representations into ``rotation-invariant'' features and ``rotation-equivariant'' features, respectively. Moreover, we design a fast QPU to accelerate the computation of QPU. The fast QPU applies a tree-structured data indexing process, and accordingly, leverages the power of parallel computing, which reduces the computational complexity of QPU in a single thread from $\mathcal{O}(N)$ to $\mathcal {O}(\log N)$. Taking the fast QPU as a basic module, we develop a series of quaternion neural networks (QNNs), including quaternion multi-layer perceptron (QMLP), quaternion message passing (QMP), and so on. In addition, we make the QNNs compatible with conventional real-valued neural networks and applicable for both skeletons and point clouds. Experiments on synthetic and real-world 3D tasks show that the QNNs based on our fast QPUs are superior to state-of-the-art real-valued models, especially in the scenarios requiring the robustness to random rotations.<br>

Fast Quaternion Product Units for Learning Disentangled Representations in SO(3)

Real-world 3D structured data like point clouds and skeletons often can be represented as data in a 3D rotation group (denoted as $\mathbb{SO}(3)$). However, most existing neural networks are tailored for the data in the Euclidean space, which makes the 3D rotation data not closed under their algebraic operations and leads to sub-optimal performance in 3D-related learning tasks. To resolve the issues caused by the above mismatching between data and model, we propose a novel non-real neuron model called \textit{quaternion product unit} (QPU) to represent data on 3D rotation groups. The proposed QPU leverages quaternion algebra and the law of the 3D rotation group, representing 3D rotation data as quaternions and merging them via a weighted chain of Hamilton products. We demonstrate that the QPU mathematically maintains the $\mathbb{SO}(3)$ structure of the 3D rotation data during the inference process and disentangles the 3D representations into ``rotation-invariant'' features and ``rotation-equivariant'' features, respectively. Moreover, we design a fast QPU to accelerate the computation of QPU. The fast QPU applies a tree-structured data indexing process, and accordingly, leverages the power of parallel computing, which reduces the computational complexity of QPU in a single thread from $\mathcal{O}(N)$ to $\mathcal {O}(\log N)$. Taking the fast QPU as a basic module, we develop a series of quaternion neural networks (QNNs), including quaternion multi-layer perceptron (QMLP), quaternion message passing (QMP), and so on. In addition, we make the QNNs compatible with conventional real-valued neural networks and applicable for both skeletons and point clouds. Experiments on synthetic and real-world 3D tasks show that the QNNs based on our fast QPUs are superior to state-of-the-art real-valued models, especially in the scenarios requiring the robustness to random rotations.<br>

Analysis of the limiting behavior of a biological neurons system with delay

In this work an analytical and numerical analysis of the limiting behaviors of a system consisting of a pair of biological neurons was carried out. In this case connection between neurons will occur with a delay. As a neuron model, the FitzHugh-Nagumo model was chosen as a model that can reproduce many dynamic behaviors of a real neuron and, at the same time, is not very complex computationally.

Channel block of the astrocyte network connections accounting for the dynamical transition of epileptic seizures

Epilepsy has been found to be modulated by the astrocyte systems in experiments, and tremendous modeling studies have unveiled the roles of astrocyte cellular functions such as the calcium and potassium channels in the epileptic seizures. However, little attention has been paid to the structure changes of astrocytes in the epileptic seizures in the scale of networks. This paper first constructs a neuron-astrocyte network model to explain the experimental observation that astrocytes mainly induce epilepsy by blocking the channels of the astrocyte gap junction in the network scale. Such model is used to discuss potential seizure induction process in the network by changing the connection intensity of the astrocyte gap junction. The simulation results show that a decrease of the gap junction intensity changes the firing pattern of the population of neurons from slow periodical firing to high-frequency epileptic seizures, featuring epileptic patterns of depolarization blocks. This further verifies that epileptic seizures are experimentally induced via the channel block of the astrocyte gap junctions. Because of the heterogeneous structure of the real neuron-astrocyte network, the effect of changing astrocyte network structures on the seizure activities is then studied in two typical network structures: the regular neighboring connection and the random connection. The results show that an increase of the number of regular connections of the regular neighboring astrocyte network could inhibit the induction and spread of the epileptic seizures. The epileptic inhibition can be achieved similarly by increasing the connection probability of the random astrocyte network. These findings further provide evidence for the experimental phenomena of the protective response of gliosis to epilepsy with increasing gap junctions. Above all, the simulation results suggest a potential pathway of epilepsy treatment by targeting the astrocyte gap junctions.

CONTROLE DE CAOS NO MODELO NEURONAL DE HINDMARSH-ROSE COM PARÂMETROS INCERTOS

In bioengineering there is a great motivation in studying the Hindmarsh-Rose (HR) neuron model due to the fact that it represents well the biological neuron, making it possible to simulate several behaviors of a real neuron, including periodic, aperiodic and chaotic behaviors, for example. Based on this model, this article proposes applying a linear optimal control design to the uncertain and chaotic behavior established by changes in the parameters of the system. To do so, the mathematical system of the RH model and its chaotic behavior are presented; afterwards, the fixed parametersare replaced by uncertain ones, and the chaotic dynamics of the system is investigated. At last, the linear optimal control is proposed as a method for controlling the chaotic behavior of the model, and numerical simulations are presented to show the efficiency of this proposal.

HippoUnit: A software tool for the automated testing and systematic comparison of detailed models of hippocampal neurons based on electrophysiological data

Anatomically and biophysically detailed data-driven neuronal models have become widely used tools for understanding and predicting the behavior and function of neurons. Due to the increasing availability of experimental data from anatomical and electrophysiological measurements as well as the growing number of computational and software tools that enable accurate neuronal modeling, there are now a large number of different models of many cell types available in the literature. These models were usually built to capture a few important or interesting properties of the given neuron type, and it is often unknown how they would behave outside their original context. In addition, there is currently no simple way of quantitatively comparing different models regarding how closely they match specific experimental observations. This limits the evaluation, re-use and further development of the existing models. Further, the development of new models could also be significantly facilitated by the ability to rapidly test the behavior of model candidates against the relevant collection of experimental data. We address these problems for the representative case of the CA1 pyramidal cell of the rat hippocampus by developing an open-source Python test suite, which makes it possible to automatically and systematically test multiple properties of models by making quantitative comparisons between the models and electrophysiological data. The tests cover various aspects of somatic behavior, and signal propagation and integration in apical dendrites. To demonstrate the utility of our approach, we applied our tests to compare the behavior of several different hippocampal CA1 pyramidal cell models from the ModelDB database against electrophysiological data available in the literature, and concluded that each of these models provides a good match to experimental results in some domains but not in others. We also show how we employed the test suite to aid the development of models within the European Human Brain Project (HBP), and describe the integration of the tests into the validation framework developed in the HBP, with the aim of facilitating more reproducible and transparent model building in the neuroscience community.Author summaryAnatomically and biophysically detailed neuronal models are useful tools in neuroscience because they allow the prediction of the behavior and the function of the studied cell type under circumstances that are hard to investigate experimentally. However, most detailed biophysical models have been built to capture a few selected properties of the real neuron, and it is often unknown how they would behave under different circumstances, or whether they can be used to successfully answer different scientific questions. To help the modeling community develop better neural models, and make the process of model building more reproducible and transparent, we developed a test suite that enables the comparison of the behavior of models of neurons in the rat hippocampus and their evaluation against experimental data. Applying our tests to several models available in the literature, we show that each model is able to capture some of the important properties of the real neuron but fails to match experimental data in other domains. We also use the test suite in the model development workflow of the European Human Brain Project to aid the construction of better models of hippocampal neurons and networks.

ESTUDO DO COMPORTAMENTO DINÂMICO DO MODELO NEURONAL DE HINDMARSH-ROSE

Based on the Hindmarsh-Rose (RH) neuronal model for nerve impulse transmission, this paper aims to study the properties and dynamic behavior of the non-linear chaotic system that describes neuronal bursting in a single neuron. On the part of bioengineering, there is great motivation in the study of the HR model because it is well representative of the biological neuron, being able to simulate several behaviors of a real neuron, among them periodic, aperiodic and chaotic behavior. The literature suggests that the chaotic behaviorrepresents in the human being the epileptic or convulsive state. Through computer simulations, considering the system parameters, it was analyzed that the stability is highly sensitive to the initial conditions and producing oscillations, more so, when the oscillation increases the random behavior tends to increase making the system unpredictable.

Neurons with Multiplicative Interactions of Nonlinear Synapses

Neurons are the fundamental units of the brain and nervous system. Developing a good modeling of human neurons is very important not only to neurobiology but also to computer science and many other fields. The McCulloch and Pitts neuron model is the most widely used neuron model, but has long been criticized as being oversimplified in view of properties of real neuron and the computations they perform. On the other hand, it has become widely accepted that dendrites play a key role in the overall computation performed by a neuron. However, the modeling of the dendritic computations and the assignment of the right synapses to the right dendrite remain open problems in the field. Here, we propose a novel dendritic neural model (DNM) that mimics the essence of known nonlinear interaction among inputs to the dendrites. In the model, each input is connected to branches through a distance-dependent nonlinear synapse, and each branch performs a simple multiplication on the inputs. The soma then sums the weighted products from all branches and produces the neuron’s output signal. We show that the rich nonlinear dendritic response and the powerful nonlinear neural computational capability, as well as many known neurobiological phenomena of neurons and dendrites, may be understood and explained by the DNM. Furthermore, we show that the model is capable of learning and developing an internal structure, such as the location of synapses in the dendritic branch and the type of synapses, that is appropriate for a particular task — for example, the linearly nonseparable problem, a real-world benchmark problem — Glass classification and the directional selectivity problem.

Decoding sound level in the marmoset primary auditory cortex

Neurons that respond favorably to a particular sound level have been observed throughout the central auditory system, becoming steadily more common at higher processing areas. One theory about the role of these level-tuned or nonmonotonic neurons is the level-invariant encoding of sounds. To investigate this theory, we simulated various subpopulations of neurons by drawing from real primary auditory cortex (A1) neuron responses and surveyed their performance in forming different sound level representations. Pure nonmonotonic subpopulations did not provide the best level-invariant decoding; instead, mixtures of monotonic and nonmonotonic neurons provided the most accurate decoding. For level-fidelity decoding, the inclusion of nonmonotonic neurons slightly improved or did not change decoding accuracy until they constituted a high proportion. These results indicate that nonmonotonic neurons fill an encoding role complementary to, rather than alternate to, monotonic neurons. NEW & NOTEWORTHY Neurons with nonmonotonic rate-level functions are unique to the central auditory system. These level-tuned neurons have been proposed to account for invariant sound perception across sound levels. Through systematic simulations based on real neuron responses, this study shows that neuron populations perform sound encoding optimally when containing both monotonic and nonmonotonic neurons. The results indicate that instead of working independently, nonmonotonic neurons complement the function of monotonic neurons in different sound-encoding contexts.

Neural Flows in Hopfield Network Approach

In most of the applications involving neural networks, the main problem consists in finding an optimal procedure to reduce the real neuron to simpler models which still express the biological complexity but allow highlighting the main characteristics of the system. We effectively investigate a simple reduction procedure which leads from complex models of Hodgkin-Huxley type to very convenient binary models of Hopfield type. The reduction will allow to describe the neuron interconnections in a quite large network and to obtain information concerning its symmetry and stability. Both cases, on homogeneous voltage across the membrane and inhomogeneous voltage along the axon will be tackled out. Few numerical simulations of the neural flow based on the cable-equation will be also presented.