Recent advances in Artificial Intelligence (AI) on the Internet of Things (IoT) devices have realized Edge AI in several applications by enabling low latency and energy efficiency. However, deploying state-of-the-art Convolutional Neural Networks (CNNs) such as VGG-16 and ResNets on resource-constrained edge devices is practically infeasible due to their extensive parameter counts and floating-point operations (FLOPs). Thus, the concept of network pruning as a type of model compression is gaining attention for accelerating CNNs on low-power devices. State-of-the-art pruning approaches, either structured or unstructured , postulate a three-stage: training-pruning-retraining pipeline, which results in an inevitable retraining overhead. In this work, we posit an orthogonal conjecture on structured pruning at initialization to find sparse subnetworks realizing ≈ 60% less training time than conventional pruning. Moreover, conventional pruning focuses on identifying the saliency at filter granularity while ignoring the importance of characteristics at layer granularity. In contrast, the proposed complexity-driven approach leverages the intrinsic complexities of CNN layers to guide the filter-pruning process without requiring dense pre-training of models. Particularly, we characterize the importance of CNN layers with respect to parameters, FLOPs, and memory-based complexities to work in tandem with filter-pruning in a structured manner. Experiments show the competitive performance of our approach in terms of accuracy and acceleration for all three modes of pruning, namely parameter-aware (PA), FLOPs-aware (FA), and memory-aware (MA). For example, reducing ≈ 70% parameters, ≈ 50% FLOPs, and ≈ 50% memory from MobileNetV2 did not result in any accuracy loss, unlike state-of-the-art approaches. Lastly, we present a trade-off between different resources and accuracy, which can be helpful for developers in making the right decisions in resource-constrained IoT environments.

There has been an increase in the implementation of Artificial Intelligence (AI) in the dairy industry for Milk Quality Analysis (MQA). However, traditional Machine Learning (ML) algorithms may not be effective due to non-linearity in milk spectral data and the requirement of pre-processing. Important features from the spectral data may be lost during the pre-processing stage, which is a severe problem. Deep Learning (DL) can help by eliminating the need for pre-processing, thereby avoiding the loss of information. Although traditional DL methods have been used in dairy farming applications, fewer studies indicate the use of DL for MQA. Therefore, there is a need to develop novel DL models for MQA to improve the classification accuracy for milk quality monitoring. This study proposes a Hybrid Blended Deep Learning (HyBDL) approach for better classification accuracy and lower prediction errors. The proposed model outperformed classical DL and Blended DL models in terms of overall accuracy, loss, and class-wise accuracy used in this study. The model achieved 98.03% accuracy and lower Mean Squared Error (MSE) scores for each iteration, and its power consumption, energy consumption, and training time were evaluated. To support our work, we calculated the reproducibility score for all the models, representing how consistent the results are when repeated multiple times. Time complexity analysis of the models is performed to compare the resource consumption and training times for the base learners and HyBDL model. To further validate the performance of our model, we have trained it on different resource-intensive edge devices, such as the NVIDIA Jetson Nano and a low-end device. Edge devices can be used in dairy processing plants to provide real-time milk quality predictions making it essential to this field of research. Our proposed HyBDL model outperformed all the other models by having a low deviation score of 0.0037 for ten iterations and 0.0077 for 100 iterations showing high reproducibility.

Air pollution has become an environmental threat globally. Therefore, air pollution estimation and prediction are crucial for effecting timely control measures. This paper proposes an ensemble deep particulate forecaster (EDPF) to predict hourly PM_2.5 concentrations for a day ahead. Air quality data from Thiruvananthapuram, Kerala (an Indian state), is used as a case study to demonstrate the model. The effect of other pollutant concentrations and various meteorological parameters on PM_2.5 prediction is first studied using Pearson correlation analysis and recursive feature elimination method. Thus, the most relevant features for the forecast of PM_2.5 concentrations are identified. Then preliminary PM_2.5 predictions are made using two base predictors: an attention-based convolutional neural network - bidirectional long short-term memory (CNN-BiLSTM) model and a bidirectional long short-term memory model (BiLSTM). Finally, a feed-forward deep neural network (DNN) model combines the preliminary predictions yielding the final PM_2.5 predictions. The proposed model is validated for long-term forecasting of PM_2.5 at Thiruvananthapuram, and the model attains root mean squared error of 12.96 µg/m³ and mean absolute error of 9.28 µg/m³ for forecasting 24 h ahead PM_2.5 concentrations. The experimental results demonstrate that the model performs better compared to various deep learning models in the literature. The effectiveness of the model in all four seasons is also evaluated, which reveals winter is the most challenging season to forecast.

Hand motion tracking traditionally requires highly complex and expensive systems in terms of energy and computational demands. A low-power, low-cost system could lead to a revolution in this field as it would not require complex hardware while representing an infrastructure-less ultra-miniature (~ 100μm - [1]) solution. The present paper exploits the Multiple Point Tracking algorithm developed at the Tyndall National Institute as the basic algorithm to perform a series of gesture recognition tasks. The hardware relies upon the combination of a stereoscopic vision of two novel Lensless Smart Sensors (LSS) combined with IR filters and five hand-held LEDs to track. Tracking common gestures generates a six-gestures dataset, which is then employed to train three Machine Learning models: k-Nearest Neighbors, Support Vector Machine and Random Forest. An offline analysis highlights how different LEDs' positions on the hand affect the classification accuracy. The comparison shows how the Random Forest outperforms the other two models with a classification accuracy of 90-91 %.