We have presented brain-inspired reinforcement learning using a model of the Basal Ganglia (BG) implemented on the neuromorphic computer SpiNNaker. The novelty of this work is the inclusion of the the Substantia Nigra pars compacta population of BG that releases the neurochemical Dopamine (DA). Neuromodulation by DA allows dynamic synaptic weight learning by rewarding desired actions. This is unlike previous models that focused on action-selection using static weights. Furthermore, we have scaled up the model such that there are two channels, each of which are trained to select specific actions in response to input cues provided by periodic spike trains. Also, both fast and slow excitatory input pathways to the BG model from the cortex are used, unlike in previous work where only the fast synapses were modelled. To test the trained network, we simulated a robotic navigation system. The model is required to demonstrate a wall-following behaviour in a room environment of different dimensions. Thus, a ‘move forward’ action is selected for no detected obstacle; a ‘turn left’ action is selected otherwise. The environment is sensed at periodic intervals and live spike injection is used to provide cue to the model to take the required action. Our results show that the two-channel BG network trained with instrumental conditioning avoided clashing into corners successfully and retained its wall-following behaviour. Overall, our work provides a neuromorphic framework for implementing robust decision-making and action-selection systems on low-power real-time neuromorphic devices.

We have made a comparative study of three optimisation algorithms viz. Random Search (RS), Grid Search (GS) and Bayesian Optimization (BO) to find optimal hyperparameter combinations in an existing brain-inspired thalamocortical model that can simulate brain signals such as local field potentials (lfp) and electroencephalogram (eeg). The layout and parameters for the model are sourced from anatomical and physiological data. However, there is a lot of missing data in such sources due to obvious constraints in wet-lab experimental studies. In our previous work, the missing data are set by trial and error. As the scale of the model gets larger though, the combinatorics of the hyperparameters explode and manual parameter tuning gets non-trivial. The goal of this study is to identify the optimisation algorithm (among the three abovementioned) that gives the best performance at minimal computational costs; performance is evaluated by setting an objective, which is to search for hyperparameter combinations that can simulate theta (4 – 8 Hz), alpha (8 – 13 Hz) and beta (13 – 30 Hz) rhythms, which are typically observed in eeg and lfp. Each optimisation algorithm is tested on a small model (thalamus only) with eight hyperparameters and a large model (thalamocortical) with maximum of fifteen hyperparameters. The performance metric for each algorithm is measured by the number of times the objective is achieved during a fixed number of trials. Our results demonstrate that BO performs the best in reaching the objective with a 30.5% better performance compared to GS and 13% better than RS. In comparison, GS performance is lower with an exponential increase in time with increasing grid size. Overall, our study demonstrates the suitability of using the BO for optimising hyperparameter search in our thalamocortical network model of the visual pathway.

Following the initial implementation of a full-scale spiking neural network (SNN) of the cortical microcircuit on NEST, the work was replicated to simulate on SpiNNaker, the Juelich CPU cluster, and the Sussex GPU cluster, in order to compare the performances on the different platforms. All of these researches use the Leaky Integrate and Fire (LIF) model as the basic unit of spiking neurons. In comparison, Izhikevich's spiking neuron models (IZK) can mimic a larger variety of known cortical neuronal dynamics. In spite of this versatility, the IZK neuron is easy to implement and computes fast. In this work, we implement the above-mentioned cortical microcircuit at a reduced-scale and using IZK neurons on SpiNNaker. This is aligned with our ongoing research on a reduced-scale thalamocortical circuit of vision with changing IZK neuron dynamics on SpiNNaker. We validate our SNN with the LIF-based full-scale cortical microcircuit by providing Poisson noise inputs, and measuring objectively the outputs in terms of spike rate, irregularity and synchrony. Our reduced-scale SNN shows similar dynamics to the full-scale SNN and operates within the Asynchronous Irregular regime defined by set bounds on the three quantitative attributes. Next, we test our SNN with inputs from a Dynamic Vision Sensor- (DVS-)based electronic retina (e-retina) that converted a simple periodic environmental input to spike trains. With current parameter settings, the model output identifies the low-frequency, but not the high frequency periodic inputs.

We present a collated set of algorithms to obtain objective measures of synchronization in brain time-series data. The algorithms are implemented in MATLAB; we refer to our collated set of ‘tools' as SyncBox. Our motivation for SyncBox is to understand the underlying dynamics in an existing population neural network, commonly referred to as neural mass models, that mimic Local Field Potentials of the visual thalamic tissue. Specifically, we aim to measure the phase synchronization objectively in the network response to periodic stimuli; this is to mimic the condition of Steady-state-visually-evoked-potentials (SSVEP), which are scalp Electroencephalograph (EEG) corresponding to periodic stimuli. We showcase the use of SyncBox on our existing neural mass model of the visual thalamus. Following our successful testing of SyncBox, it is currently being used for further research on understanding the underlying dynamics in enhanced neural networks of the visual pathway. Link to SyncBox: https://github.com/PranavMahajan25/SyncBox

This paper studies sleep stage classification using NeuCube, a Spiking Neural Network (SNN) architecture, simulated on SpiNNaker, a neuromorphic computer. The sleep electroencephalogram (EEG) time series is converted to spikes and provided as an input to NeuCube. Relevant feature vectors are extracted at different stages of training. We used six standard machine learning classifiers on different combinations of these feature vectors and calculated 5-fold cross-validation accuracy. We observed that the gradient boosted decision trees classifier performed the best by achieving 81.25% accuracy on a combination of two feature vectors. An evaluation of the results using confusion matrices and classification reports showed that the Awake, N2, SWS and REM sleep stages can be classified with ≥ 80% F1-score using the gradient boosted decision trees algorithm. Overall, our proof-of-concept work towards autonomous sleep-stage classification using NeuCube shows promise and will form the base for continued research in this direction.

We have presented a Spiking Convolutional Neural Network (SCNN) that incorporates retinal foveal-pit inspired Difference of Gaussian filters and rank-order encoding. The model is trained using a variant of the backpropagation algorithm adapted to work with spiking neurons, as implemented in the Nengo library. We have evaluated the performance of our model on two publicly available datasets - one for digit recognition task, and the other for vehicle recognition task. The network has achieved up to 90% accuracy, where loss is calculated using the cross-entropy function. This is an improvement over around 57% accuracy obtained with the alternate approach of performing the classification without any kind of neural filtering. Overall, our proof-of-concept study indicates that introducing biologically plausible filtering in existing SCNN architecture will work well with noisy input images such as those in our vehicle recognition task. Based on our results, we plan to enhance our SCNN by integrating lateral inhibition-based redundancy reduction prior to rank-ordering, which will further improve the classification accuracy by the network.

Designing a robust classification mechanism with a higher accuracy for Electroencephalogram (EEG) signals is a challenging task. In this paper, a metaheuristic based multi-objective fuzzy modelling mechanism based on the One-Against-One (OAO) strategy has been developed for classification of multiclass steady state visual evoked potential (SSVEP) data. In this work, three different flickering frequencies in 10Hz, 14Hz and 21Hz were used to elicit the SSVEPs. The recorded EEG signals were segmented into 2.5-second long epochs and features were extracted using the Discrete Wavelet Transform (DWT) method. The proposed classification mechanism provides higher classification accuracy compared to baseline classification algorithms based on adaptive neuro fuzzy inference system (ANFIS) and artificial neural networks (ANNs), this is achieved by simultaneously improving two objectives: the prediction accuracy and interpretability of the fuzzy rule-based systems. The results highlight that by searching for both optimal parameters and structure of the classifier, the generalisation capability and interpretability are improved.

Among different machine learning algorithms AdaBoost is a classification technique, which improves the classification accuracy by increasing the weights of the misclassified data. To overcome the problem of misclassification in Real AdaBoost algorithm, of the already classified samples, concept of margin is employed in the Parameterized AdaBoost algorithm. The new parameter, introduced in Parameterized AdaBoost, corresponding to the margin is chosen randomly between 0 to 1. However, the margin value is different for different classification problem. Hence, in this paper, the parameter corresponding to the margin is adapted by learning the parameter value with the help of Differential Evolutionary algorithm corresponding to the optimal classification accuracy. Experiment for the support of the proposed Adaptive Parameterized AdaBoost Algorithm has been conducted with different standard database given by UCI Machine Learning Repository. In addition, an application of Adaptive Parameterized AdaBoost is performed in EEG Motor Imagery Classification. Finally the EEG Motor Imagery Classified data (Left/Right) is tested in a robot.

The memory network is a result of current dipoles created in the brain. Localizing the source of these current flows is known as source localization, and it could potentially reveal which parts of the brain are actually responsible for a particular brain activity. It would also increase the spatial resolution of an EEG recording by identifying the true source of multiple correlated readings. In our experiments, we employed memory networks to classify perception of emotional instances conveyed in facial expressions as well as to localize sources. These networks were created by selectively evaluating EEG channel signals pairwise for Granger causality. Channel selection was based on clustering of EEG features by Self Organizing Feature Map (SOFM). Principal Component Analysis (PCA) was employed for dimension reduction and noise elimination of EEG features. Finally a new metric based on Fischer's discriminant was used to compare different source localization techniques, where real source locations are unknown. The perception of the stimuli was classified as belonging to one the following classes i) Happy ii) Sad iii) Fear iv) Relaxed. The created memory networks could classify perception of emotional content in 90.64% of cases. Comparison by the proposed Fischer Discriminant based metric revealed that the proposed network identification technique performs better at source localization as compared to independent component based source localization.

This paper introduces a novel approach to examine the scope of touch perception as a possible modality of treatment of patients suffering from certain mental disorder using a Radial Basis function induced Back Propagation Neural Network. Experiments are designed to understand the perceptual difference of schizophrenic patients from normal and healthy subjects with respect to four different touch classes, including soft touch, rubbing, massaging and embracing and their three typical subjective responses such as pleasant, acceptable, and unpleasant. Experiments undertaken indicate that that the frontal part of the scalp map of healthy subjects carry more blood during touch perception than those obtained for the schizophrenic patients. Further, for normal subjects and schizophrenic patients, the average percentage accuracy in classification of all the three classes including pleasant, acceptable or unpleasant is comparable with their respective oral responses. In addition, for schizophrenic patients, the percentage accuracy for acceptable class is very poor of the order of below 10%, which for normal subjects is quite high (46%). Performance analysis reveals that the proposed classifier outperforms its competitors with respect to classification accuracy in all the above three classes. A well known statistical test confirms that the proposed classifier outperforms all its competitors along with principal component analysis as feature selector by a large margin.

This paper proposes novel algorithms for data-point and feature selection of motor imagery electroencephalographic signals for classifying motor plannings involved in car- driving including braking, acceleration, left steering control and right steering control. Variants of neural network classifiers such as linear support vector machines, and kernel-based support vector machines including radial basis function kernel, polynomial kernel and hyperbolic kernel have been applied to classify the various cognitive tasks. Experimental finding reveals that the proposed data-point and feature selection technique altogether provides better classification accuracies (more than 88%) for all cognitive tasks in comparison with using factor analysis for data-point reduction and feature selection. It is also observed that power spectral density and discrete wavelet transform features are selected among the list of electroencephalographic features for holding the top two rank values for cognitive task classification during car-driving. From the experimental result, it is confirmed that support vector machines with radial basis function along with power spectral density outperforms the remaining feature-classifier pairs in terms of average classification accuracy.

This paper describes a study of the effects of variations in the thalamocortical synaptic activity on alpha rhythms (8 - 13 Hz) using a computational model. The study aims to investigate alpha rhythm slowing associated with Alzheimer's Disease. It is observed that for a certain range of values of the input, an increase in inhibitory activity results in an increase of the lower alpha band (8 - 10 Hz) power and a corresponding decrease in the upper alpha band (11 - 13 Hz) power, thus indicating a slowing of the alpha rhythms. On the other hand, an increase in the excitatory synaptic activity results in an overall shift of the peak power in the output signal from the lower alpha band to the upper alpha band. However, for values of input outside this range, the output signal shows a bifurcation in behaviour and enters a limit cycle mode. In this state, the output power lies dominantly in the lower alpha band. Variation in the inhibitory or excitatory synaptic parameters has little or no effect on the frequency band of the output power.

In this paper, we present biologically inspired means to enhance perceptually important information retrieval from rank-order encoded images. Validating a retinal model proposed by VanRullen and Thorpe, we observe that on average only up to 70% of the available information can be retrieved from rank-order encoded images. We propose a biologically inspired treatment to reduce losses due to a high correlation of adjacent basis vectors and introduce a filter-overlap correction algorithm (FoCal) based on the lateral inhibition technique used by sensory neurons to deal with data redundancy. We observe a more than 10% increase in perceptually important information recovery. Subsequently, we present a model of the primate retinal ganglion cell layout corresponding to the foveal-pit. We observe that information recovery using the foveal-pit model is possible only if FoCal is used in tandem. Furthermore, information recovery is similar for both the foveal-pit model and VanRullen and Thorpe's retinal model when used with FoCal. This is in spite of the fact that the foveal-pit model has four ganglion cell layers as in biology while VanRullen and Thorpe's retinal model has a 16-layer structure.

We propose a model of the primate retinal ganglion cell layout corresponding to the foveal-pit to test rank-order codes as a means of sensory information transmission in primate vision. We use the model for encoding images in rank-order. We observe that the model is functional only when the lateral inhibition technique is used to remove redundancy from the sampled data. Further, more than 80% of the input information can be decoded by the time only up to 10% of the ganglion cells of our model have fired their first spikes.