Real-time modulation recognition is crucial for modern communication systems in various cognitive radio tasks. While prior works have employed deep learning techniques to address this challenge, few are feasible for real-time applications. Spiking Neural Networks (SNNs) present a promising alternative to conventional deep learning approaches, enabling low-power hardware implementations. However, existing SNN-based modulation recognition methods often lag behind traditional techniques or necessitate high sample rate implementations. This work introduces an SNN architecture that utilizes a low-resolution quantizer in the receiver and operates at a lower rate than the quantizer, resulting in significant area and power savings when integrated into a system. We experimentally determine the optimal quantizer resolution and the ratio of quantizer-to-SNN rate. The optimized network achieves an average classification accuracy of 68.45% on the RadioML2018.01A dataset, utilizing a 4-bit quantizer and running at a rate 16 times lower than the quantizer. This performance is comparable to conventional neural networks and surpasses that of previous spiking-based methods, especially at low signal-to-noise ratio (SNR) conditions.

Analog-to-digital converters (ADCs) are essential components in modern signal processing systems, but their performance is often constrained by non-idealities such as mismatches, gain errors, offsets, etc. Two most popular ADC toplogies, viz., Successive Approximation Register (SAR) ADC whose performance is mostly affected by capacitor mismatch and Pipelined ADCs whose performance is mostly affected by capacitor mismatch and gain errors are discussed in this paper. Machine learning (ML)-based calibration techniques have recently emerged as effective tools to mitigate these challenges and enhance ADC performance. This paper provides a comprehensive overview of ML-driven approaches for calibrating SAR and Pipelined ADCs, emphasizing key methodologies, advantages, and limitations. Additionally, traditional Least Mean Squares (LMS)-based calibration methods are discussed and shown to be a limiting case of ML-based calibration.

In an era, where the exponential growth of image data driven by the Internet of Things (IoT) is outpacing traditional storage solutions, this work explores and advances the potential of implicit neural representation (INR) as a transformative approach to image compression. INR leverages the function approximation capabilities of neural networks to represent various types of data. While previous research has employed INR to achieve compression by training small networks to reconstruct large images, no work has explored past the fundamental barrier of using one network per image. This work proposes a novel advancement by breaking this barrier and representing multiple images with a single network. By modifying the loss function during training, the proposed approach allows a small number of weights to represent a large number of images, even those significantly different from each other. A thorough analytical study of the convergence of this new training method is also carried out, establishing upper bounds that not only confirm the method’s validity but also offer insights into optimal hyperparameter design. The proposed method is evaluated on the Kodak, ImageNet, and CIFAR-10 datasets. Experimental results demonstrate that all 24 images in the Kodak dataset can be represented by linear combinations of two sets of weights, achieving a peak signal-to-noise ratio (PSNR) of 26.5 dB with as low as 0.2 bits per pixel (BPP). The proposed method matches the rate-distortion performance of state-of-the-art image codecs, such as BPG, on the CIFAR-10 dataset. Additionally, the proposed method maintains the fundamental properties of INR, such as arbitrary resolution reconstruction of images.

Spiking Neural Networks (SNNs) are closely related to the dynamics of the human brain and use spatiotemporal encoding of information to generate spikes. Implementing various neuronal models in hardware is a popular field of research aiming to mimic biological behavior. The leaky integrate-and-fire model of the neuron is generally chosen for hardware implementation owing to its simplicity and accuracy in modeling the neuron. This paper proposes an infinite impulse response (IIR) filter-based neuron model and describes a backpropagation-based training algorithm for an SNN built using the proposed neurons. The trained network is implemented on an Ultra96-V2 FPGA to validate the design and demonstrate the power and resource efficiency. The implemented design achieves an accuracy of 98.91% on the MNIST dataset and classifies images at 13,021 frames-per-second (FPS) with a 200 MHz clock while consuming$\approx 7.5\times$higher resource efficiency than previous publications.