Many fields have evolved in the last decade through the development of deep learning. 5G and artificial intelligence (AI) have become the most popular area for the development of wireless communications, and competitions are an unprecedented way to promote the integration of 5G and AI by involving academic and industrial R&D to solve typical wireless problems with AI and machine learning (ML) tools. The Wireless Communication AI Competition (WAIC) has constantly contributed to developing AI combined with the wireless communication area for over three years. It has drawn thousands of qualified experts who devote themselves to this advancement. As one of the competition’s strongest participants, Samsung R&D Institute China - Beijing (SRC-B) has achieved several honors in the past. This year, researchers at SRC-B won second and third place in the “AI-based joint Channel Estimation and Channel State Information Feedback” track.

Wireless Communication AI Competition (WAIC)

WAIC is hosted by the 5G+AI Work Group from the International Mobile Telecommunications-2020 (IMT-2020) Promotion Group, which is committed to conducting research on the requirements, theories, and technologies of deep integration of 5G and AI and promoting the international standardization and industrialization process of 5G and AI integration. More than 300 participants from operators, vendors, and universities all over the world took part in the competition.

The third WAIC was held online in September and October 2022 and consisted of two tracks: “AI-based Joint Channel Estimation and Channel State Information Feedback” and “AI-based High-Precision Positioning.” Both tracks are typical use cases in the 3rd Generation Partnership Project (3GPP) release 18 study items on AI and ML for the New Radio (NR) air interface.

SRC-B has always played an essential role in the IMT-2020 and WAIC. For this year, SRC-B mainly focused on the “AI-based Joint Channel Estimation and Channel State Information Feedback” track and placed second and third.

AI-based Joint Channel Estimation and Channel State Information Feedback

Accurate channel estimation and effective channel state information (CSI) feedback are the basic conditions to ensure the performance of wireless communication systems. This topic focuses on joint channel estimation and CSI feedback. This topic’s challenge also includes certain time intervals between the pilot signal and the target data transmission moment, as well as the channel’s time variance.
Two teams from SRC-B designed two distinguished solutions and won second and third place in this topic. One focused on wireless communication techniques, and another focused on AI.

Mixer and Vector solutions

The WAIC aims to inspire the participants to take advantage of the AI network’s strong approximating and sequence learning ability to design channel estimation and transmit channel information in 64-bit joint solutions.

The standard research team adopts a multilayer perceptron (MLP)–mixer for channel estimation and a transformer for CSI feedback as their baseline. The fundamental techniques were drawn from two aspects. One aspect is data preprocessing based on insights into communication data, which includes, but is not limited to, OCC (Orthogonal Cover Code) decoding the pilot signal, data dimension rearranged according to the antenna array layout, data augmentation by antenna array dimension flipping, adding phase shifting and linear scaling on signal amplitude. It is verified that these preprocessing methods positively contribute to the final result because of the enhancement of the AI model’s generalization ability.

Another key challenge comes from the non-differentiability of the vector quantization layer, which caused a considerable barrier to network training. Multistage training is introduced, where training is prepared by skipping the quantization layer. The quantization layer training and the non-quantization section will pull each other to convergence in an iterative way.

Mixer and Reparameterization

The AI research team proposed a new solution that can overcome the non-differentiability of the quantization layer rather than the traditional linear or vector quantization operation, where the 64 feedback bits are treated to present the entire information sequence. Under this structure, discrete distribution can be changed into continuous distribution by involving the Gumbel distribution. As such, the quantization layer can be trained by gradient backpropagation as a part of the AI network. Therefore, the whole solution is called “mixer and reparameterization,” where AI experts think out of the box within the wireless communication domain.