Agentic AI aims to create systems that set their own goals, adapt proactively to change, and refine behavior through continuous experience. Recent advances suggest that, when facing multiple and unforeseen tasks, agents could benefit from sharing machine-learned knowledge and reusing policies that have already been fully or partially learned by other agents. However, how to query, select, and retrieve policies from a pool of agents, and how to integrate such policies remains a largely unexplored area. This study explores how an agent decides what knowledge to select, from whom, and when and how to integrate it in its own policy in order to accelerate its own learning. The proposed algorithm, Modular Sharing and Composition in Collective Learning (MOSAIC), improves learning in agentic collectives by combining (1) knowledge selection using performance signals and cosine similarity on Wasserstein task embeddings, (2) modular and transferable neural representations via masks, and (3) policy integration, composition and fine-tuning. MOSAIC outperforms isolated learners and global sharing approaches in both learning speed and overall performance, and in some cases solves tasks that isolated agents cannot. The results also demonstrate that selective, goal-driven reuse leads to less susceptibility to task interference. We also observe the emergence of self-organization, where agents solving simpler tasks accelerate the learning of harder ones through shared knowledge.

Predictive Coding (PC) has emerged as a prominent theory underlying information processing in the brain. The general concept for learning in PC is that each layer learns to predict the activities of neurons in the previous layer, which enables local computation of error as well as in-parallel learning across layers. Deep Bi-directional Predictive Coding (DBPC) is proposed here as a new learning algorithm that enables neural networks to simultaneously perform classification and reconstruction tasks using the same learned weights. Building on existing PC approaches, DBPC supports both feedforward and feedback propagation of information. Each layer in the network trained using DBPC learns to predict the activities of neurons in the previous and next layers, enabling the network to simultaneously perform classification and reconstruction tasks using feedforward and feedback propagation, respectively. DBPC also relies on locally available information for learning, thus enabling in-parallel learning across all layers in the network. DBPC enables the training of both fully connected networks and convolutional neural networks. The classification accuracies of DBPC on the MNIST, Fashion-MNIST, and CIFAR-10 datasets (99.58%, 92.42%, and 74.29%, respectively) exceed those of well-established PC-based benchmark approaches (including FIPC 3 and iPC) and are competitive with state-of-the-art Error-Backpropagation-based methods (including ResNet and DenseNet) on MNIST, Fashion-MNIST, and EuroSAT datasets. Importantly, DBPC achieves these results using significantly smaller networks for MNIST, Fashion-MNIST, and CIFAR-10 datasets (0.425, 1.004, and 1.109 million parameters), and every representation estimated in DBPC can be used for the reconstruction of inputs. The significant benefit of DBPC is its ability to achieve this performance using locally available information and in-parallel learning mechanisms, which results in an efficient training protocol. Overall, we demonstrate that DBPC is a much more efficient approach for training networks that can perform both classification and reconstruction simultaneously.

Behavioral variability across individuals leads to substantial performance differences during cognitive tasks, although its neuronal origin and mechanisms remain elusive. Here we use recurrent neural networks trained on a multisensory decision-making task to investigate inter-subject behavioral variability. By uniquely characterizing each network with a random synaptic-weights initialization, we observed a large variability in the level of accuracy, bias and decision speed across these networks, mimicking experimental observations in mice. Performance was generally improved when networks integrated multiple sensory modalities. Additionally, individual neurons developed modality-, choice- or mixed-selectivity, these preferences were different for excitatory and inhibitory neurons, and the concrete composition of each network reflected its preferred behavioral strategy: fast networks contained more choice- and mixed-selective units, while accurate networks had relatively less choice-selective units. External modulatory signals shifted the preferred behavioral strategies of networks, suggesting an explanation for the recently observed within-session strategy alternations in mice.

Spiking neural networks (SNNs) mimic their biological counterparts more closely than their predecessors and are considered the third generation of artificial neural networks. It has been proven that networks of spiking neurons have a higher computational capacity and lower power requirements than sigmoidal neural networks. This article introduces a new type of SNN that draws inspiration and incorporates concepts from neuronal assemblies in the human brain. The proposed network, termed as class-dependent neuronal activation-based SNN (CDNA-SNN), assigns each neuron learnable values known as CDNAs which indicate the neuron's average relative spiking activity in response to samples from different classes. A new learning algorithm that categorizes the neurons into different class assemblies based on their CDNAs is also presented. These neuronal assemblies are trained via a novel training method based on spike-timing-dependent plasticity (STDP) to have high activity for their associated class and low firing rate for other classes. Also, using CDNAs, a new type of STDP that controls the amount of plasticity based on the assemblies of pre- and postsynaptic neurons is proposed. The performance of CDNA-SNN is evaluated on five datasets from the University of California, Irvine (UCI) machine learning repository, as well as Modified National Institute of Standards and Technology (MNIST) and Fashion MNIST, using nested cross-validation (N-CV) for hyperparameter optimization. Our results show that CDNA-SNN significantly outperforms synaptic weight association training (SWAT) ( ) and SpikeProp ( ) on 3/5 and self-regulating evolving spiking neural (SRESN) ( ) on 2/5 UCI datasets while using the significantly lower number of trainable parameters. Furthermore, compared to other supervised, fully connected SNNs, the proposed SNN reaches the best performance for Fashion MNIST and comparable performance for MNIST and neuromorphic-MNIST (N-MNIST), also utilizing much less (1%-35%) parameters.

Knowing expected milk yield can help dairy farmers in better decision-making and management.The objective of this study was to build and compare predictive models to forecast daily milk yield over a long duration.A machine-learning pipeline was provided and five baseline models as well as a novel stacking model were developed for the prediction of milk yield on the CowNflow dataset using 414 Holstein cattle records collected from 1983 to 2019.Four different feature selection methods were performed to evaluate the essential biological characteristics and feeding-related features which affect milk yield.The results showed that the overall performance of predictive models improved after proper feature selection, with an R 2 value increased to 0.811, and a root mean squared error (RMSE) decreased to 3.627.The stacking model achieved the best performance with an R 2 value of 0.85, a mean absolute error (MAE) of 2.537 and an RMSE of 3.236.This research provides benchmark information for the prediction of milk yield on the CowNflow dataset and identifies useful factors such as dry matter (DM) intake and lactation month in long-term milk yield prediction.

In this paper, a Gaussian mixture synapse model for a time-varying Weight Spiking neural classifier (GWS) is proposed. A two-layer Spiking Neural Network (SNN) is developed, in which the synapses are modeled using multiple gaussian functions spread across the simulation interval. An error function is evaluated using the spike times of the output neuron. The learning algorithm updates the parameters of the gaussians based on the gradient of this error function. Insights that endorse the faster and efficient decision-making capability of the GWS classifier are demonstrated using the spike times of output neurons. The performance of the GWS is evaluated on datasets in the UCI machine learning repository and compared with other existing spiking neural classifiers. Also, the robustness of GWS is further evaluated by using it to classify the complex real-world dataset generated using electroencephalogram (EEG) signals.

Multi-label learning and continual multi-label learning are crucial challenges in machine learning, particularly in handling complex data with multiple overlapping labels over time. Recent research works try to tackle the effect of data imbalance as it makes multi-label learning more challening. This work introduces an adaptive margin spiking neural net-work (AM-SNN) architecture coupled with a novel imbalance-sensitive loss function designed to enhance robustness against class imbalance in these settings. AM-SNN employs two output layers: one for predictions and another for margin values, with a unique loss function leveraging cosine similarity between predictions and ground truths to improve confidence in the model's predictions. Experiments show that AM-SNNs trained with the proposed loss function outperform state-of-the-art loss functions on metrics such as the imbalance-weighted F1 score and the F1 score for the most imbalanced class on several multi-label learning datasets. In continual multi-label learning, AM-SNNs surpass Bipolar SNNs and the CIFDM benchmark on large datasets - Birds, Human, and Eukaryote.

This paper develops a new approach to estimate predicted class probabilities in deep Spiking Neural Networks (SNN) that encourages faster classification. The proposed approach utilizes the temporal separation between the first spikes generated by the output neurons to estimate the predicted class probabilities which are then used with cross entropy loss for training the network. This maximizes the separation between the first spikes generated by the neuron associated with the correct class and neurons associated with other classes. Higher classification performance is obtained by maximising the tem-poral separation, which also drives the correct class neuron to spike earlier in the simulation. As a consequence, the predicted class may be determined from the first spike in the output layer, leading to quicker classification. The sensitivity factor for each neuron in the network is estimated via error-backpropagation during training. Using Spike Timing Dependent Plasticity (STDP) regulated by the estimated sensitivity factors, the network weights are updated. It results that the learning method is termed as Temporal Separation Modulated Spike Timing Dependent Plasticity (TSM-STDP). On the benchmark MNIST dataset, the performance of TSM-STDP has been assessed, and the evaluation results are compared with those of other learning methods for SNNs. Additionally, a histogram of the output layer's first spikes demonstrated that the right class neurons spiked earlier in the simulation than other class neurons, enabling faster classification. On real-world Attention Deficit Hyperactivity Disorder (ADHD) detection dataset, the effectiveness of TSM-STDP has also been assessed and compared with other available approaches. The per-formance comparison results clearly show that TSM-STDP can achieve classification performance comparable to other existing learning algorithms on benchmark and real-world datasets while requiring less time for classification.

A novel 3D deep convolutional neural network (3D-CNN) model called MEGNet3D has been proposed in the paper. MEGNet3D is designed to differentiate between amyotrophic lateral sclerosis (ALS) and healthy individuals from their resting state (eyes open and eyes closed condition) sensor-level magnetoencephalography (MEG) data. The raw MEG data is initially transformed into their time-frequency representation, which are then used as inputs to MEGNet3D. Both magnetometer and gradiometer recordings have been investigated separately. The proposed model exhibits an accuracy of over 75% for most classification conditions. Thus, MEGNet3D is capable of handling high subject variability and shows that spectral-temporal representation of resting-state MEG data yields relevant neural markers related to the existence of ALS. Furthermore, it has also been observed resting state with eyes closed yields better classification accuracy as compared to the resting state with eyes open condition.

Spiking Neural Networks (SNNs), An alternative to sigmoidal neural networks, include time into their operations using discrete signals called spikes. Employing spikes enables SNNs to mimic any feedforward sigmoidal neural network with lower power consumption. Recently a new type of SNN has been introduced for classification problems, known as Degree of Belonging SNN (DoB-SNN). DoB-SNN is a two-layer spiking neural network that shows significant potential as an alternative SNN architecture and learning algorithm. This paper introduces a new variant of Spike-Timing Dependent Plasticity (STDP), which is based on the assembly of neurons and expands the DoB-SNN's training algorithm for multilayer architectures. The new learning rule, known as assembly-based STDP, employs trained DoBs in each layer to train the next layer and build strong connections between neurons from the same assembly while creating inhibitory connections between neurons from different assemblies in two consecutive layers. The performance of the multilayer DoB-SNN is evaluated on five datasets from the UCI machine learning repository. Detailed comparisons on these datasets with other supervised learning algorithms show that the multilayer DoB-SNN can achieve better performance on 4/5 datasets and comparable performance on 5th when compared to multilayer algorithms that employ considerably more trainable parameters.

Spiking neural networks (SNNs) as the third generation of artificial neural networks are closer to their biological counterparts than their predecessors. SNNs have a higher computational capacity and lower power requirements than networks of sigmoidal neurons. In this paper, a new spiking neural network for pattern classification referred to as Degree of Belonging SNN (DoB-SNN) is introduced. DoB-SNN is inspired by a neuronal assembly where each neuron has a degree of belonging to every class of data being process. DoB-SNN clusters the neurons during the training process using DoBs to allocate a group of neurons to each class. A new training algorithm is presented to adjust DoBs along with the network's synaptic weights, based on Spike-Timing Dependent Plasticity (STDP) and neurons' activity for training samples. The performance of DoB-SNN is evaluated on five datasets from the UCI machine learning repository. Nested Cross-Validation is employed to determine the network's hyperparameters for each dataset and thoroughly assess generalisation capability. A detailed comparison on these datasets with three other supervised learning algorithms, including SpikeProp, SWAT, and SRESN is provided. The results show that no algorithm significantly outperforms DoB-SNN, Whereas DoB-SNN has significantly better performance than others for Liver disorders dataset (>6.10%,$p < 0.01$). Accuracies obtained by DoB-SNN are significantly greater than SWAT for both Iris and Breast Cancer (>1.69%,$p < 0.001$) and significantly better than SpikeProp for Iris (1.62%,$p=0.04$). In all comparisons, DoB-SNN used the smallest network, among others. DoB-SNN therefore offers significant potential as alternative SNN architecture and learning algorithm.

This paper proposes a Distributed Coding Spiking Neural Network (DC-SNN) with a self-regulated learning algorithm to deal with pattern classification problems. DC-SNN employs two hidden layers. First hidden layer has receptive field neurons that convert the real-valued input features to spike patterns and the second hidden layer employs LIF neurons with inhibitory interconnections. The second hidden layer has been termed as the distributed coding layer in the rest of the paper. The inhibitory interconnections in distributed coding layer will ensure that each neuron in this layer learns a distinct spike pattern from input feature space. The synaptic weights between layers and the weights of lateral inhibitory connections are learned using a self-regulated learning algorithm. Self-regulation identifies neurons for updating in the output layer and distributed coding layer and also adapts the learning rate based on the temporal separation between spikes in the output layer. It also skips learning from samples which are correctly classified with higher temporal separation and hence prevents over-training. The detailed performance comparisons of DC-SNN with other algorithms for SNNs in the literature using six benchmark data set from the UCI machine learning repository has been presented. Further, the performance of DC-SNN is evaluated on a real-world brain computer interface problem for classification of electroencephalogram (EEG) signals recorded during motor-imagery tasks. The results clearly indicate that the proposed DC-SNN architecture provides slightly better generalization ability and is suitable for deep spiking networks.

Generative Adversarial Networks (GAN) have led to important advancements in generation of time-series data in areas like speech processing. This ability of GANs can be very useful for Brain-Computer Interfaces (BCIs) where collecting large number of samples can be expensive and time-consuming. To address this issue, this paper presents a new approach for generating artificial electroencephalography (EEG) data for motor imagery. GANs here use a generator and discriminator networks that consist of Bidirectional Long Short Term Memory neurons. Trained models are evaluated using the dataset 2b from the BCI competition IV. The dataset consists of trials with left and right hand motor imagery. Separate GANs are trained to generate artificial EEG samples corresponding to the two types of trials present in the data set. For the purpose of evaluation, the time-frequency characteristics of the real and artificial EEG signals are compared using Short-Term Fourier Transform and Welch's power spectral density. The results indicate that GANs can capture important characteristics of motor imagery EEG data such as power variations in the beta-band. The power variation in the artificial generated and original signal was in the similar frequency bin when looked at Welch's power spectral density.

This paper presents a new learning algorithm developed for a three layered spiking neural network for pattern classification problems. The learning algorithm maximizes the interclass margin and is referred to as the two stage margin maximization spiking neural network (TMM-SNN). In the structure learning stage, the learning algorithm completely evolves the hidden layer neurons in the first epoch. Further, TMM-SNN updates the weights of the hidden neurons for multiple epochs using the newly developed normalized membrane potential learning rule such that the interclass margins (based on the response of hidden neurons) are maximized. The normalized membrane potential learning rule considers both the local information in the spike train generated by a presynaptic neuron and the existing knowledge (synaptic weights) stored in the network to update the synaptic weights. After the first stage, the number of hidden neurons and their parameters are not updated. In the output weights learning stage, TMM-SNN updates the weights of the output layer neurons for multiple epochs to maximize the interclass margins (based on the response of output neurons). Performance of TMM-SNN is evaluated using ten benchmark data sets from the UCI machine learning repository. Statistical performance comparison of TMM-SNN with other existing learning algorithms for SNNs is conducted using the nonparametric Friedman test followed by a pairwise comparison using the Fisher's least significant difference method. The results clearly indicate that TMM-SNN achieves better generalization performance in comparison to other algorithms.

In this paper, an automatic seizure detection technique using multichannel EEG is proposed based on Metacognitive Complex-valued Interval Type-2 Fuzzy Inference System (McCIT2FIS). A wavelet chaos theory based feature extraction is employed to extract the features from EEG signal as it can handle the non stationarity in data and Sparse Multinomial Logistic Regression via Bayesian L1 Regularisation (SBMLR) based feature selection is employed to select the most discriminative features. McCIT2FIS is employed to classify the samples as either interictal or ictal EEG segment as it has been shown to be capable of handling noisy data by virtue of Interval Type-2 fuzzy sets, and is good at classification because of its ability to handle complex-valued data. Further, we have also shown that the feature selected using SBMLR can be successfully mapped back to the channels allowing us to identify the epileptogenic regions of the brain. The performance of the McCIT2FIS was also compared with the support vector machines and the results indicate that McCIT2FIS is better capable of detecting seizure based on EEG signals.

This paper presents a two stage learning algorithm for a Growing-Pruning Spiking Neural Network (GPSNN) for pattern classification problems. The GPSNN uses three layered network architecture with input layer employing a modified population coding and, leaky integrate-and-fire spiking neurons in the hidden and output layers. The class label for a sample is determined according to the output neuron with minimum spike latency. The learning algorithm for the GPSNN employs a two stage learning mechanism. In the first stage, the hidden layer is grown and adapted to map the inputs to a hyperdimensional space. In the second stage, the hidden layer neurons with low dominance are pruned and the response of the most dominant neurons is mapped to the output space. The proposed approach has been evaluated on benchmark data sets from the UCI machine learning repository and the results were compared with batch as well as online spiking neural networks. The results clearly highlight that the GPSNN can achieve better performances using a compact network structure.

In this paper, we develop a new sequential learning algorithm for a spiking neural network classifier. The algorithm handles the input features that are not in the form of a spike train but in a real-valued (analog) form. The sequential learning algorithm evolves the number of spiking neuron automatically based on the information present in the current sample and results in a compact architecture. Hence, it is referred to as a Minimal Spiking Neural Network (MSNN). The learning algorithm can either add a new neuron to the network or update the parameters of the existing neurons based on the information contained in the arriving samples. The update rule uses excitatory/inhibitatory rule to capture the knowledge contained in the current sample. Performance evaluation of the proposed MSNN is presented using two benchmark problems from the UCI machine learning repository, namely, the Iris flower classification and Wisconsin breast cancer problem and the results are compared with other existing spiking neural algorithms like SpikeProp, MuSpiNN and Multi-spike learning algorithms. The results clearly indicate the better performance of MSNN with a compact architecture.

This paper presents an evolving spiking neural network namely, `Basis Coupled Evolving Spiking Neural Network (BCESNN)' and its learning algorithm to solve real-valued pattern recognition problems. BCESNN is a two-layered neuron model with afferent neurons in the input layer and efferent neurons in the output layer. The afferent neurons in the input layer convert the real-valued input feature to a train of spikes using a bank of Gaussian Receptive Field (GRF) for each individual feature. The number of GRF per feature is fixed a priori. Each efferent neuron in the output layer is associated to a class. Efferent neurons are integrate-and-fire type neuron. BCESNN has an evolving architecture that uses basis coupled rank order learning (BCROL) algorithm to estimate the number of output neurons and the network parameters. Each sample is presented only once to the network. When a new sample is presented to the network either we add a neuron or we update an existing neuron. Weight estimation for added neuron is done using BCROL and weight update is done using Euclidean distance based distance measure. In the performance section we conducted three different experiments. Firstly we compared the performance of BCROL against Rank Order Learning(ROL). Next, we evaluated the performance of BCESNN on benchmark classification problems from the UCI machine learning repository. Finally, we evaluated the performance of a sparsely connected BCESNN against fully connected BCESNN where connectivity refers to the number of GRF connected to the afferent neurons.