Data centers (DCs) are now heavily involved in large-scale artificial intelligence (AI) learning processes. Fiber optic systems are the primary communication medium for enabling DC networks due to their high speed and low energy consumption per bit. Traditionally, fiber optic systems are designed to ensure nearly error-free communication, which increases their power consumption and latency. In this work, we revisit the communication requirements in DCs for AI training by analytically examining how communication channel noise affects the learning process. Our analysis shows that channel errors introduce bounded noise in the model weights, which can be maintained within an acceptable range. Therefore, relaxing the error-free requirement has a negligible impact on AI learning performance while significantly improving power efficiency and latency. Additionally, we propose an optimized stack layer for the optical communication link with minimal modifications to the IEEE 802.3bs standard, tailored for AI training. To validate our approach, we simulate the training of the ChatGPT-2 model on a DC with 32 workers and noisy communication links, showing that the training process tolerate a bit error rate of up to$1 \mathrm{e}-4$.

In high-rate or long-haul optical fiber transmissions, correcting chromatic dispersion (CD) is critical but challenging and energy-intensive due to the large filter tap size required for CD compensation (CDC). Overlap-save (OLS) is a common frequency-domain CDC technique that uses fast Fourier transform (FFT). However, hardware constraints-such as power, memory, latency, and chip area-limit the FFT size. This limitation makes OLS too complex or even infeasible in dispersive channels where the number of taps approaches or exceeds the FFT size. We introduce the time-segmented overlap-free (TS-OLF) technique, a novel frequency-domain block filtering method to enable low-complexity CDC under FFT-size limitations. TS-OLF divides the signal into non-overlapping blocks and segments the filter accordingly to enable filtering operations with any FFT size. It aggregates the results of different filter segments directly in the frequency domain, therefore, using a minimum number of FFT operations. We show that TS-OLF achieves consistently lower complexity than OLS when the filter size is more than half the FFT size, and unlike OLS, can handle filter sizes that exceed the FFT size. TS-OLF also outperforms other filter segmenting methods, providing significant complexity improvements.

According to the Gaussian noise (GN) model, nonlinear interference noise (NLIN) in fiber depends on the signal power spectral density (PSD). Consequently, optimizing the PSD of the pulse that modulates data, as the main factor influencing the PSD of the launched signal into the fiber, can effectively minimize fiber NLIN. In this study, we first employ the calculus of variations to identify the optimal band-limited pulse PSD that minimizes fiber NLIN. Next, we add other communication requirements, such as zero inter-symbol interference (ISI) and fast decay over time, as constraints to our design problem. For this case, we develop a general pulse model and formulate the design problem as an optimization problem. By solving this optimization problem, we find the optimal pulse PSD that not only minimizes NLIN power in fiber but also meets practical requirements. We study the time-domain impact of the designed modulating pulse PSD on the launched signal properties to gain insights into the nonlinearity benefits we achieve. We further analytically demonstrate that our designed pulse has favorable properties for the Godard timing recovery method. Through extensive simulations using the split-step Fourier method on a fiber with typical parameters and considering practical transmitter/receiver limitations, we illustrate the superior system reach and achievable data rate of our optimized pulses compared to existing pulse shapes.

Digital backpropagation (DBP) is the most effective fiber optic compensation technique. While DBP can increase the achievable data rate, its high computational complexity prevents its use in real-time applications. Parallelization is one of the most promising techniques to allow real-time DBP. Existing parallel DBP (PDBP) are based on coupled nonlinear Schrodinger equation (NLSE), which only considers cross-phase modulation and discards the rest of the inter-channel impairments. Therefore, these methods lose their performance in today's wavelength-division multiplexing systems. We fill this gap by proposing a PDBP capable of working with regular NLSE and compensating for all fiber impairments, including four-wave mixing. We analytically derive the compensation equations for the parallel scheme and handle them efficiently to reduce the complexity. Our simulations show a significant improvement in the sense of optimal launched power, achievable data rate, and system reach over existing DBPs with comparable complexity.

We propose a hybrid probabilistic-geometric constellation shaping method for optical fiber communication systems that is non-linearity tolerant and compatible with probabilistic fold shaping (PFS) architecture. To do so, the impact of non-linear interference noise (NLIN) in the shaping process is considered. This is a challenging process because NLIN is itself constellation dependent. This study copes with this issue by developing a deep learning-based shaping method that takes the enhanced Gaussian noise (EGN) model of fiber into account to model NLIN. Our hybrid shaping scheme maximizes the generalized mutual information (GMI) rate of the optical system while the impacts of shaping on the NLIN power are considered. Our results show that the proposed hybrid scheme results in significant reach improvements and outperforms recent hybrid shaping methods. Interestingly, our hybrid shaping scheme suggests distributions that are much different from the traditional Maxwell-Boltzmann distributions.

Non-line-of-sight (NLOS) error is one of the dominant sources of error in localization applications. Existing algorithms rely on solving a set of highly nonlinear equations to compensate for this error, which is intractable in practice. In this paper, we propose an efficient NLOS identification algorithm based on supervised machine learning. This approach enables us to improve localization accuracy by taking advantage of the NLOS measurements if the location of the reflector is known. Hence, our approach can be employed in combination with 5G intelligent reflecting surface systems to provide location-based wireless services. We also analytically derive the Cramer-Rao lower bound for the localization problem at hand. Finally, we investigate the performance of our proposed NLOS identification algorithm under different simulation setups.

Sparse signals are encountered in many modern technologies. Compressed sensing methods create the opportunity of sampling sparse signals significantly lower than the Nyquist rate. However, sparse signal recovery still remains a challenge. In this work, we show that detecting the support of the sparse signal can be viewed as a classification problem, and hence efficiently be solved using neural networks (NN). We find the best NN configuration in support detection for random data sets, where an accuracy of 98% is achieved in highly sparse signals. After detecting the support of the signal, we show that signal recovery can be done efficiently. In this work, we develop two recovery methods. The simulation studies confirm that our proposed methods outperform counterpart algorithms in both accuracy and convergence speed.