Reliability feedback usually arrives after the expensive step
A conventional receiver learns whether a transport block passed its cyclic redundancy check only after equalization and decoding. For ultra-reliable low-latency communications, that sequence delays link adaptation and retransmission decisions. This preprint proposes predicting the post-decoding CRC outcome directly from the 5G New Radio physical-uplink resource grid and least-squares channel estimates.
The predictor combines a lightweight convolutional neural network with a fixed front end that extracts multi-scale time–frequency energy features. Two front ends are studied. One uses wavelet scattering with fixed Gabor filters; the other uses FFT-based scattering with geometrically spaced bandpass masks. Because the filters are fixed, the learned component does not need to discover the entire signal representation from data.
Radio context and uncertainty govern early action
The modulation and coding scheme index is supplied as an additional conditioning input. That is an important piece of radio context: the same received-energy pattern can imply a different decoding risk under a different code rate or modulation order. The abstract reports that both hybrid scattering variants outperform a pure CNN baseline on a multi-MCS 5G NR uplink dataset, with MCS conditioning improving reliability across the evaluated channel conditions.
An evidential deep-learning extension estimates epistemic uncertainty in a single forward pass and supports a conservative decision rule. This matters because an early predictor should be able to abstain or defer to the conventional chain when it is unsure. The abstract describes the front ends as lightweight and compatible with GPU-accelerated real-time baseband inference, but it does not provide end-to-end latency, block-error-rate operating points, calibration error, or performance under unseen channel distributions.
The contribution is therefore best understood as moving a decision boundary earlier in the receiver pipeline. The decisive deployment test will not be average classification accuracy alone. It will be whether uncertainty-aware early feedback preserves the stringent reliability target while saving enough time to change what the scheduler or retransmission loop can do.
Research notes
Neural CRC Prediction for 5G NR URLLC
- Author: Prashanth Murthy
- Public record: arXiv
- What is established: The method combines fixed wavelet or FFT scattering features, a lightweight CNN, MCS conditioning, and an evidential uncertainty extension to predict CRC outcomes before full decoding.
- Read with care: The abstract gives qualitative comparisons but not the latency, reliability operating point, calibration error, or out-of-distribution results needed for a real URLLC deployment claim.