ComNum - Séminaires

Le 15/01/2020 [ComNum] Multiple access with finite payloads and many users

Auteur(s) & Affilliation(s) du séminaire :

Yuri Polyanskiy

Présentation du séminaire :

in 3C14 at 10am


One of the next frontiers in communication technology is to re-think methods of sharing
spectrum among wireless transmitters. In this talk, I will discuss recent results on
characterizing the fundamental tradeoff between the density of users (ratio of active users and the number of degrees of freedom available per frame) and 
the minimal required energy-per-bit (Eb/N0). In the regime of interest for the Internet-of-Things, 
the payload of each user is about 100 data bits. This is too small for the asymptotic Shannon theory to provide any guidance and thus we needed to develop new tools for analysis.
A surprising discovery is existence of coded-access schemes that are able to almost perfectly reject
the multi-user interference, so that increasing the density of users
(without increasing space-time-frequency resources) does not lead to
any deterioration of service. Practically important is that known
MAC architectures are not capable of attaining this effect. A similar analysis and effects
arise in the problem of (unsourced) random-access over the AWGN channel, which we will
discuss as well.

Contact(s) :

Wigger Michele

Le 14/01/2020 [ComNum PhD seminar] DNN assisted Sphere Decoder

Auteur(s) & Affilliation(s) du séminaire :

Askri Aymen

Présentation du séminaire :

in 3C41 at 2pm


For a wide class of MIMO space-time transmission schemes, Maximum Likelihood decoding requires us to solve an integer least-squares problem. The sphere decoder (SD) algorithm is one of the optimal decoders that require high complexity. Our work addresses this issue by using artificial intelligence to reduce the computational complexity. To that end, we train a feedforward neural network (NN) to predict the number of lattice points inside the sphere. Since the number of lattice points is intelligently learnt by a NN, the SD algorithm starts the tree search with a radius that allows a small number of points falling inside the sphere. This results in a significant complexity reduction.



Contact(s) :

Ciblat Philippe

Image Retour haut de page