R1-2600135
discussion
6GR CSI: Considerations for Evaluation of AI/ML-based Solution
From InterDigital
Summary
This document from InterDigital presents 3 Observations and 3 Proposals regarding the evaluation methodology for AI/ML-based Channel State Information (CSI) processing in 6G, arguing that AI/ML solutions introduce hidden complexities (LCM, hardware, energy, interoperability) that vary by use case and must be rigorously benchmarked against state-of-the-art non-AI/ML solutions, not just legacy NR designs.
Position
InterDigital presents a technical case against unqualified adoption of AI/ML for CSI processing by emphasizing hidden costs in Model Life Cycle Management (LCM) overhead, computational and hardware complexity, energy efficiency versus spectral gain, and multi-vendor interoperability testing. They propose studying these four specific dimensions to decide whether an AI/ML-based solution should be supported for CSI processing in 6GR. They require that evaluation benchmarks use state-of-the-art non-AI/ML solutions like sparsity-based frequency/spatial domain designs rather than legacy NR CSI-RS inherited from LTE. They also support identifying specific use cases (CSI-RS overhead reduction, CSI prediction, CSI compression) before discussing evaluation metrics, arguing that two-sided models like CSI compression impose significantly more signaling, computational complexity, power, and testing burdens than single-sided models.
Key proposals
- Proposal 1 (Sec 3): Study Model Life Cycle Management Overhead, Computational/Hardware Complexity, Energy efficiency, and Interoperability/Testing as aspects for deciding whether an AI/ML-based solution should be supported for CSI processing in 6GR.
- Proposal 2 (Sec 3): Support identification of relevant use cases prior to discussing evaluation assumptions and metrics, due to varying degrees of AI/ML complexity per use case.
- Proposal 3 (Sec 4): Support use of state-of-the-art non-AI/ML based solutions, such as sparsity-based frequency/spatial domain solutions, as the benchmark for evaluation of AI/ML use cases, rather than NR CSI-RS design which is inherited from LTE.
- Observation 1 (Sec 3): There are various aspects related to AI/ML-based solutions that weigh on the decision between an AI/ML-based and a non-AI/ML-based solution.
- Observation 2 (Sec 3): The degree of complexity of AI/ML-based CSI processing varies according to the use case.
- Observation 3 (Sec 4): There are non-AI/ML based solutions for scalable CSI-RS design, e.g., sparsity-based frequency/spatial domain solutions, that offer a good performance at a lower complexity.