Owing to the generation of vast amount of unlabelled dynamic data and the need to analyze them, deep unsupervised learning based clustering algorithms are gaining importance in the field of data ...science. Since the task of automated feature extraction is proficiently combined with the machine learning models in deep unsupervised learning algorithms, they are identified to be superior as compared to conventional dynamic similarity measure based clustering methods. In this context, the authors present a recurrent neural network (RNN) based clustering algorithm optimization, where the vital information representing the dynamic data (or time-series data) is extracted first and subsequently clustered using a soft clustering algorithm. This methodology not only ensures dynamic component extraction in terms of static features but also clusters them efficiently using an evolutionary clustering algorithm called Neuro-Fuzzy C-Means (NFCM) clustering, which reduces the large-scale optimization problem of FCM to small-scale along-with identification of optimal number of clusters. The proposed algorithm has been implemented on three different test data sets collected from machine learning repository and it was found that the results are 98-100% accurate.
Calcium spiking can be used for drug screening studies in pharmaceutical industries. However, performing experiments for multiple drugs and doses are highly expensive. The oscillatory behavior of ...calcium spiking data demonstrates extreme nonlinearity and phase singularity. This makes it more challenging to construct physics-based models for the experimental observations. In this scenario, data based modelling, such as Artificial Neural Networks (ANN), and thereafter the model based prediction of calcium profiles may offer a cost-effective and time saving solution. Therefore, a novel ANN building algorithm is presented in the current work, where data based simultaneous estimation of ANN architecture and nonlinear activation function stands out as the main highlight. The resultant ANN was then used to learn the oscillatory behavior in calcium ion concentration data, obtained from hippocampal neurons of rats by fluorescent labelling and confocal imaging. The paper shows that the novel technique can be used in general for emulating biochemical oscillations (with or without drug injection) and can be implemented to predict the cell-drug responses for intermediated doses. The proposed algorithm can also be used for obtaining high resolution data from low resolution experimental measurements.
Mixed-precision quantization has emerged as a solution in recent times for accurate inference of Deep Neural Networks on edge. However the prior-art is far from being deployable on embedded devices ...due to various practical limitations. In this work, a pipeline is designed that a) performs layer-wise assignment of bit-precisions, b) builds the quantized graph, c) deploys the graph on the embedded device (Galaxy S23), and d) measures the accuracy and on-device performance. This pipeline is optimized using multi-objective Bayesian Optimization for simultaneously maximizing the accuracy and minimizing the on-device inference time, resulting in a Pareto list. The best configuration among them resulted in 3.16, 2.8, and 2.57 times model compression, 31%, 26% and 18% latency improvement and -0.01, 0.27, and 0.08 accuracy drop for ResNet18, MobileNetV2, and InceptionV3, respectively, on ImageNet, establishing a new benchmark in mixed-precision quantization.
Neural synchronicity plays a vital role in monitoring the functions that are cognitive. Any disturbance identified in the neural synchrony might lead to a diseased state. In the case of in vitro cell ...recordings, the neurons demonstrate significant heterogeneity in the firing pattern. Thus, the task of automated identification of synchronous and asynchronous neurons from a large population of neuronal cells remains challenging. To address this issue, an efficient unsupervised machine learning approach has been proposed for a system of primary cultures of hippocampal neurons. Here, a confocal microscope is used for imaging of intracellular calcium using Fluo-4 as the fluorescent indicator. The obtained static images are transformed into time-varying data of cytosolic calcium. Subsequently, an intelligent artificial neural network (ANN) assisted fuzzy clustering algorithm is proposed for grouping the synchronous neurons from a heterogeneous set of calcium data that are spiking in nature. This novel algorithm enables a drastic variable reduction followed by the implementation of a global optimization algorithm to solve the problem in Fuzzy C-means (FCM) clustering. Additionally, the proposed technique computes the optimal cluster number and the hyper-parameters involved in ANNs. To validate the result obtained from ANN assisted FCM, a correlation coefficient, and a spiking pattern plot is analyzed for both the synchronous and asynchronous neuronal cells. Besides this, the proposed algorithm is compared with the traditional FCM, where the solution quality is found to be improved along-with an 88% reduction in decision variable count. The complete novel framework combines the aspects of calcium imaging, ANN-assisted FCM, validation, and comparison, which as a whole, can be used for quick and effective quantification of synchronicity.
Wake models play an important role in wind farm layout optimization and control studies and it is, therefore, important to model wake effects in accurate and efficient ways. The power production from ...a wind farm is estimated using analytical models such as Jensen model in the wind industry, as they are simple and their computational cost is significantly less compared to the high-fidelity models involving Large Eddy Simulations. As these analytical models are used in an iterative setup like control and optimization of wind layouts, simulation cost involved in these models is extremely important. However, most of the analytical models assume linear expansion of wake while modeling wake effects, which makes it inaccurate. In this paper, the authors present a data driven approach under the framework of machine learning to impart the effect of nonlinear expansion of wake in the analytical models and thereby empowering them to be more accurate. Such a wake model was developed by integrating Artificial Neural Networks (ANNs) and Jensen model, where the expansion of wake is assumed as nonlinear and is modeled using ANNs inside the Jensen model setup, thus establishing the nonlinearity between inter-turbine distance and reduction in wind speed. To prove the efficacy of the proposed model, the results are compared with the predictions of analytical Jensen model, and it has been shown that the proposed model performs better than Jensen, demonstrating the importance of nonlinear expansion of wake and its effect in power calculation of a wind farm.
Climate change and global energy crisis are two primary drivers behind the search of several forms of renewable energy sources e.g., wind energy. In studies of wind farm layout design and control, ...wind frequency map (WFM) plays a crucial role. WFM is a joint probability mapping between wind direction and speed considering their time series behavior. The limited past data suppresses the ability to represent long term variabilities in wind which leads to inaccurate estimation of WFMs resulting in unrealistic calculation of wind power for wind farms. Hence, in this paper, Generative Adversarial Networks (GANs) are explored for generation of wind frequency maps using limited wind time series data. GAN s are data driven probabilistic techniques, which captures distribution hidden in data. The distribution of wind scenarios, once captured by GANs, can generate new scenarios. The success of GAN's ability in capturing the probability distribution hidden in a wind dataset was demonstrated by calculating power production from an optimal layout using the available scenarios and scenarios generated by GANs. Trained GANs can generate many scenarios in future horizon which can be used for robust design of wind farms. Thus, this study can be helpful in efficient design and control of wind farms under wind state uncertainty.
Wind energy is now the world's second-fastest-growing electricity source. The power output of the wind farm depends on wind characteristics like wind speed and direction and wind farm layout. ...Specifically, these wind characteristics are modeled using a probability density function built using local wind measurements over the farm, called the Wind Frequency Maps (WFMs). The conventional approach for modeling this dynamic data is to perform manual feature extraction followed by static data clustering since the data is unlabeled. Nonetheless, since the features to be extracted are based on heuristics and may lead to information loss, this technique is inefficient. Thus, in this study, the wind characteristics data is treated in the form of images that are essentially the surface plots corresponding to the joint probability mass functions built over 12 direction sectors and 16-speed sectors. Moreover, the WFMs are modeled using a novel unsupervised Deep Learning framework where the required features are extracted using convolutional auto-encoders, followed by applying a soft clustering algorithm that can identify optimal cluster number. Here, 1400 such WFMs, were generated, 11 latent vectors were extracted, and finally, the images were grouped into 4 clusters with varying wind characteristics. Two of these clusters are found to be relatively denser. Further, this study will help perform wind farm layout optimization under uncertainty and control studies.
The practical optimal control problems often contain multiple conflicting objectives leading to a set of decision vectors called Pareto Optimal solutions. Population based evolutionary optimizers ...such as Genetic Algorithms are known to have many advantages over the classical methods when it comes to solving multi-objective optimization problems. However, their applicability in solving large scale optimization problems such as the multi-objective optimal control is severely limited due to the computational time issues. In this paper, we, therefore, present a novel neural network based strategy which reduces the size of optimal control problem (number of decision variables and constraints) by several folds. The reformulation of optimal control problem into a simple weight training exercise allows the implementation of evolutionary solvers to achieve high quality Pareto solutions. We demonstrate our technique for (i) the design of a plug flow reactor with conflicting energy and conversion costs and (ii) the control of a fed batch bioreactor with a conflict between yield and productivity.
Unsupervised learning based clustering methods are gaining importance in the field of data analytics, owing to the features they possess, such as high accuracy, simple implementation and fast ...computation, when compared with conventional supervised learning methods. Among several types of clustering techniques, those implying optimization routines are found to be more efficient. However, explosion in number of decision variables is making these algorithms computationally intensive. The authors present an efficient two-stage optimization based fuzzy clustering formulation, which works through variable reduction approach. The membership values associated with each data point, forming the majority of decision variables, are estimated under an artificial neural networks framework. The reduction in decision variables allows the implementation of evolutionary optimization solvers to solve the single objective constrained optimization problem of fuzzy clustering increasing the chance of finding global optima. Additionally, this formulation estimates the optimal network topology and optimal number of clusters, which is not estimated rather assumed by other formulations. The proposed algorithm has been implemented on three different test data sets and the efficacy of the novel approach has been demonstrated by comparing the obtained clustering results with that of conventional fuzzy clustering approach.