Pruning and Quantizing Neural Belief Propagation Decoders

Machine Learning for Multimedia Communications

“…The challenges above triggered significant research in redesigning channel encoding/decoding processes so that the decoding performance is maintained high for short and medium codeblock lengths, and both the decoding complexity and the decoding delay are greatly reduced. To this aim, machine learning has been proposed for channel encoding [ 66 , 67 , 68 ], channel decoding [ 69 , 70 , 71 , 72 , 73 , 74 , 75 , 76 , 77 , 78 , 79 ], and building end-to-end communication systems where channel encoding and decoding are jointly considered [ 80 ]. Different machine learning methods have been examined, for example, neural networks are used in [ 69 , 70 , 71 , 72 , 73 , 74 , 75 , 76 ], while designs based on reinforcement learning methods are presented in [ 77 , 78 , 79 ].…”

“…To this aim, machine learning has been proposed for channel encoding [ 66 , 67 , 68 ], channel decoding [ 69 , 70 , 71 , 72 , 73 , 74 , 75 , 76 , 77 , 78 , 79 ], and building end-to-end communication systems where channel encoding and decoding are jointly considered [ 80 ]. Different machine learning methods have been examined, for example, neural networks are used in [ 69 , 70 , 71 , 72 , 73 , 74 , 75 , 76 ], while designs based on reinforcement learning methods are presented in [ 77 , 78 , 79 ].…”

“…In addition, the parameters are learned iteration by iteration, avoiding problems with vanishing gradients. Pruning the neural belief propagation decoder is proposed in [ 75 ] where weights show how much each check node (in the Tanner graph representation) affects the decoding performance. The message exchange schedule of the belief propagation decoder is modeled as a Markov Decision Process (MDP), and reinforcement learning is to find the optimal scheduling [ 77 , 78 ].…”

Low-complexity Near-optimum Symbol Detection Based on Neural Enhancement of Factor Graphs

“…Thereby, heavily parametrized DNNs can be thinned out and memory storage as well as computational expenses can be saved [22]. The application of pruning techniques to NBP-based DNNs in the context of Tanner graphs was recently proposed in [23]. Following this approach, we investigate the effect of pruning for the NBP-based symbol detection.…”

Ανάλυση Επίδοσης Σφάλματος Σχημάτων Τούρμπο Κωδίκων Με Αναδιατάκτες Άρτιων-Περιττών Θέσεων, Βάσει Του Φάσματος Αποστάσεών Τους

“…Γι' αυτά τα μήκη, οι καθιερωμένες μέθοδοι σχεδίασης κωδίκων προσέγγισης της χωρητικότητας, όπως τα διαγράμματα ExIT, δεν έχουν την ίδια σημασία όσο για τα μεγάλα, οπότε θα μπορούσε να ακολουθηθεί μια σχεδίαση βάσει αλγορίθμων της τεχνητής νοημοσύνης [183,184]. Τέλος, για τα εν λόγω μήκη στην αρθρογραφία εξετάζονται και νευρωνικοί αποκωδικοποιητές αλγεβρικών κωδίκων [84,85,185], μιας και οι γνωστοί κώδικες προσέγγισης της χωρητικότητας φαίνεται να υστερούν.…”