On December 19th 2018 was interesting workshop related to Artificial Intelligence. interesting topic was about deep and narrow neural networks. Dr. Oseledets using Tensor Train proof that Deep Neural Network outperform shallow ones. He show for class of Multiplicative RNNs (Y. Wu, S. Zhang, Y. Zhang, Y. Bengio, R. Salakhutdinov., “On Multiplicative Integration with Recurrent Neural Networks”,2016) that Multiplicative RNNs are equivalent Tensor Train and shallow net univalent canonical decomposition. Using two side reduction he proof theorem that “a random d-dimensional TT-tensor with probability 1.0 has exponentially large CP-rank”, or from Neural Network point of view an RNN (of the form discussed earlier) with random weights can be exactly mimicked with a shallow net only of exponentially larger width (layers). Moreover they announce analytical method to compress neural networks using low-rank approximation. Extended version: https://goo.gl/vNBxcv
Today! Selected areas and carriers for 5G deployment in Russia.
http://www.cnews.ru/news/top/2018-12-24_gde_v_rossii_v_pervuyu_ochered_zarabotaet_5gspisok Earlier there were several 5G pre-commercial trials (see the map).
Vodafone is starting a 5G trial in UK. I'm looking forward to performance results when 5G hardware is also available on the consumer side. Especially, the new 3.4 GHz spectrum band should offer up to Gigabit throughput. What is about Deutsche Telecom ? They had plans to do it in 2018 in downtown Berlin at 3.7 GHz. Regarding UE: it is expected the first 5G devices will be mobile hotspots, which will roll out toward the end of 2018 or beginning 2019. Seems two main streams in 5G application development: (1) automotive life (from self-driven cars to smart cities); (2) smart healthcare (from wearable sensing to remote surgery). ... And maybe industrial automation (robot control systems), but usually it is not required wireless, and all control can be provided by fiber on the factory.
Couple of very attractive/interesting papers about equation solving in noisy conditions (Gaussian channels).
(A + n1) x = b + n2, here n1 is parameter estimation error (noise matrix); n2 is observation error (noise vector) -- both Gaussian distributed. But what’s ever doing using THEORY OF INFORMATION (Shannon) and MACHINE LEARNING – finally we obtain LINEAR ALGEBRA and THEORY OF MEASUREMENTs ! The second surprise has come from this paper: [ref. 2]. Seems one of the most important problems in mathematical society is GIBBS SAMPLING, which is private/particular case of Metropolis–Hastings algorithm for random restarts by one of dimensions. Interesting your opinion, if you have time to type several sentences about. Channel estimation is a bottleneck for massive-MIMO system both TDD and FDD, because high spatial resolution is required enough channel estimation precision that is challenge problem in high noise environment, where noise includes thermal noise, intra-cell / inter-cell interference, pilot contamination (because of not enough orthogonal pilots for measurements), hardware impairement effects (PIM, etc.). And if massive-MIMO concept, which is based on coherent processing allows to obtain significant gain for data layer, it cannot be applied steight forward way for channel estimation, because basic CSI is estimated antenna port-by-antenna port and there is no gain from massive-MIMO at channel estimation stage.
Rome-Moscow school of Matrix Methods and Applied Linear Algebra (August – September, 2018) Hometask after lecture #1 and #2:
Recommended books for deep study the subject
Huawei, together with China Unicom Group, completed field verification of the industry's first FDD-based Massive MIMO technology, using the existing two-antenna receiving terminal on the 20MHz spectrum to achieve a peak network rate of 697.3Mbps. Massive MIMO solutions have three typical characteristics: AAU hardware form, 3D user-level beamforming, multi-user MIMO.
Lately, the research world has increasingly relied on Cooper’s famous law, that most capacity increases in cellular networks will be due to denser and denser cell deployments, to get us towards the promised land of 5G data rates.
Last year we have some progress in mesh cooperative network, which removes edges between cells. This is research work, but the same time it is a room behind the door of densification. The first experiments show that such approach allows to create very similar conditions for all users, but it cannot further increase spectral efficiency as well as data rates. It is just a way for radio resource redistribution. What if we ever reached a point, where adding more infrastructure did not allow to increase capacity? Will cellular networks eventually become interference-overloaded? Will densification be the death of 5G? We will strengthen our communication and cooperation with scientists from all over the world, sponsor the research of those who move in the same direction as us, get actively involved in international industry and standards organizations, and attend all kinds of academic seminars. We should have a cup of coffee with more bright minds to absorb their brilliant ideas and sense the direction of future development.
The most efficient consortium between academia and industry in 5G area is hosted by NYU. A lot of interesting videos for beginners and professionals can be watched on IEEE.TV: https://ieeetv.ieee.org/event-showcase/the-brooklyn-5g-summit. |
AuthorPrincipal Engineer in wireless communication for PHY and MAC layers. Archives
December 2018
Categories |