• artificial intelligence
• fpga
• hls4ml
• keras
• machine learning
• monitoring
• optical networks
• pynq
• python
• qkeras
• tensorflow
• vivado
• vivado_hls

In recent years, the rapid advancement of technology and the proliferation of connected devices have transformed the way data is generated, processed, and utilized. This evolution has led to the emergence of two key paradigms in modern computational environments: cloud computing and edge computing. Both have been introduced to address specific needs in the digital era, driven by the exponential growth of data and the demand for faster, more efficient processing.

These two paradigms of data processing cater to different scenarios and requirements.

Cloud computing represents a shift away from localized hardware and software to a centralized infrastructure. It enables users to access computational services, including machine learning (ML) algorithms, storage, and processing power, over the internet on a pay-as-you-go basis. It relies on centralized data centers with immense computational resources, capable of handling large-scale ML models and data-intensive applications. Cloud computing has revolutionized industries by providing scalability, flexibility, and cost efficiency; this has been particularly beneficial for businesses managing large-scale applications like artificial intelligence (AI), machine learning (ML), and big data analytics.

Cloud computing is ideal for centralized environments where latency and resources consumption are not critical factors, as data must be transmitted to and from remote servers for processing. This centralized approach allows for scalability and high-performance model training, making it suitable for applications requiring extensive computational resources [1].

While cloud computing excels at processing large datasets and running complex algorithms, it has limitations, particularly in scenarios requiring real-time responses or low-latency communication. This is where edge computing comes into play, introduced as a complementary paradigm to cloud computing.

Edge computing brings computation closer to the data source. Instead of transmitting data to remote cloud servers, edge computing performs processing locally on edge devices such as IoT sensors, smartphones, or embedded systems. This approach reduces latency, minimizes bandwidth usage, and enables real-time processing, which is crucial for time-sensitive applications such as autonomous vehicles, healthcare monitoring, and industrial automation. However, edge devices are constrained by limited computational power, memory, and energy, making it challenging to implement complex ML algorithms designed for the cloud.

The fundamental difference between cloud and edge computing lies in the trade-off between centralized power and decentralized efficiency. While cloud computing thrives on its ability to scale and handle sophisticated models, it often falls short in meeting the low-resource usage and energy-efficient requirements of edge environments. Moreover, the reliance on continuous communication with remote data centers introduces latency and energy inefficiencies that undermine the real-time requirements of edge applications. As a result, ML models designed for the cloud cannot be directly deployed to edge devices due to their large size, high memory demands, and energy consumption. This necessitates the development of optimized, lightweight models tailored for edge computing, often achieved through techniques such as quantization, compression, and hardware-aware design tools like Hls4ml, which ensure that these models can operate effectively within the constraints of edge devices [2] [3].

The efficient implementation of ML algorithms in dedicated hardware devices at the “edge”, offers numerous advantages. Edge processing and data compression can greatly reduce data rates, and the energy required for data movement and the resources utilization of the dedicated hardware. Furthermore, real-time data processing and interpretation can greatly accelerate decision-making, hypothesis testing and even enable just-in-time interventions. Staggering data rates and massive datasets are generated across a broad range of modern scientific applications in high energy physics, sciences, and astrophysics. Low-latency ML is required for real-time decision making with a range of requirements from tens of nanoseconds to sub-millisecond. In many ways, techniques for resource-constrained ML implementations are similar whether targeting low resources consumption, ultra-low latency and high throughput [4].

In this work, we explore how ML algorithms, traditionally implemented on resource-intensive platforms, can be effectively adapted for embedded hardware deployable at the edge. To address these challenges, this work explores the use of ML models to assess and classify the operational conditions of optical fibers utilized in transport optical networks. Specifically, the thesis investigates whether the working regime of optical fibers lies within the linear or nonlinear range, aiming to optimize the system’s performance by identifying the optimal input power settings, indeed, especially using coherent optical transmission systems working in a non-linear regime can significantly degrade the system performance. The foundation of this approach relies on a dataset provided by Nokia Bells Lab comprising normalized values representing the probability density functions (PDFs) of signal-to-noise ratios (SNRs). By leveraging advanced techniques such as compression and quantization, we aim to minimize the computational and memory requirements of the utilized ML models without compromising their predictive performance. This enables their implementation on energy-efficient hardware, making real-time, low-latency processing feasible for edge applications. The outcomes of this research demonstrate the potential of these techniques to bring powerful machine learning capabilities closer to data sources, addressing the unique challenges of edge computing environments.