ELEC5508 Chap.9 AI Signal Processing for Wireless
AI Signal Processing for Wireless
This chapter begins with the learning objective rather than the network name. It distinguishes supervised regression, supervised classification, unsupervised learning and reinforcement learning, then maps regression to channel or parameter estimates and classification to bit decisions.
Conventional pilot-based processing is compared with convolutional OFDM receivers and model-driven networks, with held-out BER and distribution shift treated as essential evidence. A neural receiver must be described as a communication system, not only as a network architecture. Inputs, pilots, target labels, activation, loss, output likelihoods and deployment channels all define the task.
Convolution is useful because an OFDM grid carries local time-frequency correlation, while model-driven learning can preserve known iterative structure and learn only selected parameters. Neither choice removes the need for controlled baselines, fair overhead accounting and fallback behaviour outside the training distribution.
What this chapter covers
- 01
Supervised, unsupervised and reinforcement learning
- 02
Regression, classification and loss choice
- 03
Forward pass, backpropagation and updates
- 04
Convolution across OFDM time and frequency
- 05
Conventional and learned receiver pipelines
- 06
Model-driven learning and mismatch testing
Worked example · free
Audit one gradient update
- +1The prediction is 2 and the initial loss is 9.
- +1The weight gradient is -12.
- +1The bias gradient is -6.
- +1Updated parameters are w = 1.6 and b = 0.3.
- +1The new prediction is 3.5 and new loss is 2.25.
Key terms
- Supervised learning
- Learning an input-output mapping from labelled training pairs.
- Regression
- Prediction of a continuous target such as a channel coefficient.
- Binary cross-entropy
- A loss for Bernoulli labels and probability outputs in binary classification.
- Convolutional receiver
- A receiver using shared local filters across a time-frequency observation grid.
- Model-driven learning
- A trainable architecture that retains the structure of a classical iterative algorithm.
AI Signal Processing for Wireless FAQ
Why are CNNs suitable for OFDM receivers?
An OFDM grid has local correlation across neighbouring subcarriers and symbols because of coherence bandwidth and coherence time. Two-dimensional convolution exploits that structure with shared filters instead of flattening every observation into an unrelated feature.
What makes a fair neural receiver comparison?
Hold modulation, coding, bandwidth, pilot overhead, channel test set and output metric constant. Train without test leakage, then compare held-out BER or block error rate, inference cost and robustness to channel mismatch.
What is model-driven learning?
It preserves the stages or iterations of a known signal-processing algorithm but learns selected parameters such as step sizes or weights. This can need less data and be easier to interpret, though wrong model assumptions can still limit performance.
How do I audit a learned wireless receiver?
Begin by naming its input, output, training target and loss, then draw the conventional receiver blocks it replaces or augments. Hold modulation, coding, pilot overhead, channel samples and the communication metric constant when comparing methods. Separate training performance from inference performance and keep related channel realisations out of both training and test sets.
Test one mismatch at a time, such as a new fading distribution, Doppler range, SNR, antenna correlation or oscillator impairment. Report bit or block errors alongside inference cost and training cost. Finally, define what happens outside the represented domain: confidence monitoring, recalibration or a fallback signal-processing path makes the limitation operationally visible.
Exam move
Write the conventional receiver chain first, then label the learned replacement's input, output, target and loss. Recompute a small gradient update by hand and design mismatch tests that change amplitude distribution and Doppler separately. Begin with supervised, unsupervised and reinforcement learning and give each one a wireless decision rather than a generic software example.
For supervised work, separate regression from classification. Pair a continuous channel estimate with mean squared error and a bit probability with sigmoid output and binary cross-entropy, while remembering that other well-justified objectives are possible. Work one-neuron forward, loss, gradient and update calculations until the derivative sign can be checked by whether the next prediction moves toward the target.
Then draw the conventional OFDM receiver: pilots, channel estimation, interpolation, detection, demapping and bit likelihoods. Place a convolutional receiver beside it and show exactly which blocks are absorbed into the learned map. Explain why local convolution matches coherence across neighbouring subcarriers and symbols.
For model-driven learning, identify the retained classical iteration and the small trainable parameter set. Build a fair evaluation table with identical modulation, coding, bandwidth, pilot overhead, channel test samples and BER metric. Split data by independent channel scenarios to prevent leakage.
Add mismatch tests one factor at a time: Rayleigh to Rician, slow to fast fading, changed SNR, oscillator impairment or antenna correlation. Compare accuracy with inference cost and training cost. Finish with a safety route such as confidence monitoring, fallback processing or recalibration when observations leave the represented domain.
Working through AI Signal Processing for Wireless in ELEC5508? Sia is AskSia’s AI Electrical Engineering tutor — ask any ELEC5508 AI Signal Processing for Wireless question and get a clear, step-by-step explanation grounded in how ELEC5508 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.