Abtin received both his Bachelor and Master degrees from the University of Tehran in Computer Science and Multimedia Systems, respectively. In 2017 he was awarded “Top inventor of the University of Tehran” for creating a speech recognition device for the Persian language called “Ravannevis”. In addition he got the first-rank student in the Master’s level.

Abtin’s Master’s thesis was in the area of configuring topology of classifiers in real-time large-scale stream mining systems, an area where he hold a publicationin the prestigeous Springer Nature Journal.

Prior to joining Telecom Paristech,.Abtin worked as a research assistant at the Goethe University (Frankfurt, Germany).
In his free time Abtin is a passionate fan of learning, innovating, and swimming.

Publications:

Email:  abtin.shahkarami@telecom-paris.fr

LinkedIn

ORCID

Abtin’s Blog – the home of his blog is here

Abtin’s Website

State-of-play: Update on Vinod’s’s research in FONTE

• Capacity limits of NFDM optical fibre networks
  The demand for information traffic has risen significantly in recent years, led to the so -called “Capacity Crunch” in optical fiber. Nonlinear Frequency Division Multiplexing (N FDM) has been proposed to solve this problem. This method enhances the Wavelength Division Multiplexing (WDM) by replacing the Fourier Transform with the Nonlinear Fourier Transform (NFT). The NFT is based on the key observation that there is a linear oper ator associated with the propagation equation whose spectrum is invariant during propagation. Problematically, the state-of-the-art NFT implementation takes 𝑂(𝑁log2𝑁), where N is the number of signal samples. This makes it difficult, in multi-user nonlinear optical fiber communication, to carry out both nonlinear modulation and multiplexing. A variety of approaches has been proposed to enhance NFDM to address this problem. We, however, consider a machine learning approach to tackle the nonlinearity problem in optical fiber communication. Two different scenarios of learning the optical fiber communication channel using deep neural networks were performed in the past year. This approach provided a considerable run -time gain while maintaining acceptable accuracy. However, despite the fact that we received a guarantee from the results that the architecture is free of overfitting, we discerned that this approach might not be able to generalize well to the system variations. In view of this, we ruminated about a deep learning-based approach for optical fiber communication that is more environment-independent. Towards this, we came up with a series of theories that cast light on a novel and promising approach to strengthening the sequence of symbols at the transmitter against error throughout the channel. This approach is based on deep representation learning and feature transfer. In this document, we will elaborate on these discussions. (from: FONTE Deliverable D4.3)

• Multi-user communication and information theory
  Nonlinear frequency-division multiplexing (NFDM) has been introduced to address nonlinearity in optical fiber communication. This approach is a signal multiplexing scheme based on the nonlinear Fourier transform (NFT). Problematically, the computational complexity of the inverse nonlinear Fourier transform (NFT), typically conducted by integral equations, is high. This makes it hard, in multi -user nonlinear optical fiber communication, to perform both nonlinear modulation and multiplexing. A varie ty of approaches has been proposed to enhance NFDM to address this problem. We, however, consider a machine learning approach to tackle the nonlinearity problem in optical fiber communication. An end to end communication system was simulated. The equalizer at the receiver, which was based on Backpropagation (BP) algorithm, was targeted to be optimized, as BP has high complexity. A neural network (NN) was succeeded to be devised that approximates the high-complexity BP, with the accuracy of 99.9% in noiseless condition, while having 87.3% lower complexity. This NN equalizer also works well in noisy environments. In this article, after providing a brief background on machine learning and neural networks, we elaborate on this NN equalizer and analyse the results. We also discuss how machine learning can enhance the achievable information rate of fiber-optic communication systems by reducing the receiver complexity. (from: FONTE Deliverable D4.2)

• Principles of linear and nonlinear frequency-division multiplexing
  Wavelength-division multiplexing (WDM) and nonlinear frequency-division multiplexing (NFDM) are the two multiplexing schemes for optical fiber communication. In WDM, which is the same as linear frequency-division multiplexing (FDM) in radio communication systems, user’s signals are linearly multiplexed in the frequency domain. However, in nonlinear channels, such as optical fibers, linear multiplexing causes interactions. To address this, NFDM has been proposed. In NFDM, which is based on the nonlinear Fourier transform (NFT), users’ signals are multiplexed in the nonlinear Fourier domain and propagate independently in a lossless noiseless optical fiber modeled by the nonlinear Schrödinger (NLS) equation. In light of recent notable progress in these schemes, ever-increasing attention has been attracted to this area. In this report, mathematical principles underlying modulation and multiplexing in linear and nonlinear communication systems are reviewed.(from: FONTE Deliverable D4.1)