A mixed-signal oscillatory neural network for scalable analog computations in phase domain

Neuromorphic Computing and Engineering, Journal Year: 2023, Volume and Issue: 3(3), P. 034004 - 034004

Published: July 24, 2023

Abstract Digital electronics based on von Neumann’s architecture is reaching its limits to solve large-scale problems essentially due the memory fetching. Instead, recent efforts bring near computation have enabled highly parallel computations at low energy costs. Oscillatory neural network (ONN) one example of in-memory analog computing paradigm consisting coupled oscillating neurons. When implemented in hardware, ONNs naturally perform gradient descent an landscape which makes them particularly suited for solving optimization problems. Although ONN computational capability and link with Ising model are known decades, implementing a remains difficult. Beyond oscillators’ variations, there still design challenges such as having compact, programmable synapses modular large problem instances. In this paper, we propose mixed-signal named Saturated Kuramoto (SKONN) that leverages both digital domains efficient hardware implementation. SKONN computes phase domain while propagating information digitally facilitate scaling up size. SKONN’s separation between propagation enhances robustness enables feed-forward showcased first time. Moreover, leads unique binarizing dynamics suitable NP-hard combinatorial finding weighted Max-cut graph. We find accuracy good Goemans–Williamson 0.878-approximation algorithm Max-cut; whereas time only grows logarithmically. report Weighted experiments using 9-neuron proof-of-concept printed circuit board (PCB). Finally, present low-power 16-neuron integrated illustrate ability XOR function.

Language: Английский

---

Training energy-based single-layer Hopfield and oscillatory networks with unsupervised and supervised algorithms for image classification

Neural Computing and Applications, Journal Year: 2023, Volume and Issue: 35(25), P. 18505 - 18518

Published: June 10, 2023

Abstract This paper investigates how to solve image classification with Hopfield neural networks (HNNs) and oscillatory (ONNs). is a first attempt apply ONNs for classification. State-of-the-art are multi-layer models trained supervised gradient back-propagation, which provide high-fidelity results but require high energy consumption computational resources be implemented. On the contrary, HNN ONN single-layer, requiring less resources, however, they necessitate some adaptation as not directly applicable novel brain-inspired computing paradigm that performs low-power computation attractive edge artificial intelligence applications, such In this paper, we perform by exploiting their auto-associative memory (AAM) properties. We evaluate precision of state-of-the-art unsupervised learning algorithms. Additionally, adapt equilibrium propagation (EP) algorithm single-layer AAM architectures, proposing AAM-EP. test validate on images handwritten digits using simplified MNIST set. find learning, reaches 65.2%, 59.1% precision. Moreover, show AAM-EP can increase up 67.04% 62.6% ONN. While intrinsically meant tasks, best our knowledge, these best-reported precisions performing digits.

Language: Английский

---

Computing with oscillators from theoretical underpinnings to applications and demonstrators

Deleted Journal, Journal Year: 2024, Volume and Issue: 1(1)

Published: Dec. 4, 2024

Networks of coupled oscillators have far-reaching implications across various fields, providing insights into a plethora dynamics. This review offers an in-depth overview computing with covering computational capability, synchronization occurrence and mathematical formalism. We discuss numerous circuit design implementations, technology choices applications from pattern retrieval, combinatorial optimization problems to machine learning algorithms. also outline perspectives broaden the understanding oscillator

Language: Английский

---

SIFT-ONN: SIFT Feature Detection Algorithm Employing ONNs for Edge Detection

Published: April 11, 2023

Mobile robot navigation tasks can be applied in various domains, such as space, underwater, and transportation industries, among others. In navigation, robots analyze their environment from sensors navigate safely up to target points by avoiding obstacles. Numerous methods exist perform each task. this work, we focus on localization based feature extraction algorithms using images sensory data. ORB, SURF are state-of-the-art for feature-based thanks fast computation time, even if ORB lacks precision. SIFT is high precision detection but it slow not compatible with real-time robotic applications. Thus, our explore how speed algorithm employing an unconventional computing paradigm oscillatory neural networks (ONNs). We present a hybrid SIFT-ONN that replaces the of Difference Gaussian ONNs performing image edge detection. report performances, which similar algorithm.

Language: Английский

---

Oscillatory neural network learning for pattern recognition: an on-chip learning perspective and implementation

Frontiers in Neuroscience, Journal Year: 2023, Volume and Issue: 17

Published: June 15, 2023

In the human brain, learning is continuous, while currently in AI, algorithms are pre-trained, making model non-evolutive and predetermined. However, even AI models, environment input data change over time. Thus, there a need to study continual algorithms. particular, investigate how implement such on-chip. this work, we focus on Oscillatory Neural Networks (ONNs), neuromorphic computing paradigm performing auto-associative memory tasks, like Hopfield (HNNs). We adaptability of HNN unsupervised rules on-chip with ONN. addition, propose first solution using digital ONN design. show that architecture enables efficient Hebbian Storkey hundreds microseconds for networks up 35 fully-connected oscillators.

Language: Английский

---

A color extraction algorithm by segmentation

Scientific Reports, Journal Year: 2023, Volume and Issue: 13(1)

Published: Dec. 2, 2023

Abstract The segmentation and extraction on color features can provide useful information for many different application domains. However, most of the existing image processing algorithms feature are gray image-based consider only one-dimensional parameters. In order to carry out a fast accurate extraction, this paper proposes algorithm by that is called This compared under distribution situations, effect also shown combination algorithms. Experimental results show such has some advantages segmentation. fuzzy preprocessing, gives location method region interest. Moreover, with other algorithms, presented in not higher accuracy, shorter time stronger anti-interference ability, but better more divergent edge. evaluation proposed demonstrates dominance over extraction. These researches new way thinking segmentation, which an important theoretical references practical significance.

Language: Английский

---

Design of oscillatory neural networks by machine learning

Frontiers in Neuroscience, Journal Year: 2024, Volume and Issue: 18

Published: March 4, 2024

We demonstrate the utility of machine learning algorithms for design oscillatory neural networks (ONNs). After constructing a circuit model oscillators in machine-learning-enabled simulator and performing Backpropagation through time (BPTT) determining coupling resistances between ring oscillators, we associative memories multi-layered ONN classifiers. The machine-learning-designed ONNs show superior performance compared to other methods (such as Hebbian learning), they also enable significant simplifications topology. that single-layer ones. argue can be valuable tool unlock true computing potential hardware.

Language: Английский

---

Sub-Pixel Edge Detection Algorithm Based on Improved Zernike Moments

2022 4th International Conference on Communications, Information System and Computer Engineering (CISCE), Journal Year: 2024, Volume and Issue: unknown, P. 1087 - 1091

Published: May 10, 2024

Language: Английский

---

A prior knowledge-embedded deformable convolutional network for vision-based fabric defect detection

Textile Research Journal, Journal Year: 2024, Volume and Issue: unknown

Published: Oct. 7, 2024

The vision-based defect detection of textile surface is an essential problem in evaluating the appearance quality. In previous studies on detection, following two points were neglected: (1) proportion defects fabric small, resulting a low signal-to-noise ratio images and lack sampling features; (2) irregular shape can overlap position, leading to localization difficulty. this paper, we propose prior knowledge-embedded deformable convolutional network (PKE-DCNet) based deep learning address these issues. First, feature extraction method with information designed shape-matching clustering module region-biased for detecting complex shape. Then, boundary boxes adaptive generation proposed anchor-free search mechanism edge contour computation detect key region. Extensive experiments mixed datasets demonstrated that PKE-DCNet reached overall mAP 95.36% six types within speed 322 FPS, which was better than state-of-the-art methods.

Language: Английский

---

Two-Layered Oscillatory Neural Networks with Analog Feedforward Majority Gate for Image Edge Detection Application

2022 IEEE International Symposium on Circuits and Systems (ISCAS), Journal Year: 2023, Volume and Issue: unknown, P. 1 - 5

Published: May 21, 2023

The increasing volume of smart edge devices, like cameras, and the growing amount data to treat incited development light Artificial Intelligence (AI) solutions with neuromorphic computing. Oscillatory Neural Network (ONN) is a promising computing approach which uses networks coupled oscillators, their inherent parallel synchronization compute. Also, ONN phase allows limit voltage amplitude reduce power consumption. Low-power, fast, computation properties make attractive for AI. In state-of-the-art, built fully-connected architecture, coupling defined from unsupervised learning perform auto-associative memory tasks, Hopfield Networks. However, allow solve beyond associative applications, there need explore further architectures. this work, we propose novel architecture cascaded analog ONNs interconnected an feedforward majority gate layer. particular, show can image detection task using two layers. This is, our best knowledge, first analog-based solution cascade ONNs.

Language: Английский

---