Advanced Classical RF-ML

The Week 7 move, repeated for the architecture that took over the field: build it by hand once, and it stops being magic.

Watch and build along

Exercises

• Implement single-head scaled dot-product attention in pure NumPy; verify shapes and outputs against torch.nn.functional.scaled_dot_product_attention.
• Follow the lecture to build the character-level transformer end to end, typing every line yourself.
• Ablate: remove positional encodings and document exactly how and why performance degrades.

Checkpoint

You can state what Q, K, and V are, their shapes through a multi-head block, and why attention cost scales with sequence length squared.

Study

• Read the Vision Transformer paper for the patch-embedding idea, then a transformer-based modulation classification paper such as this one to see the same trick on I/Q data.
• Survey the current state of the art on RML2016.10a by searching arXiv and Hugging Face Papers (the old Papers With Code leaderboards shut down in 2025).

Exercises

• Build a small transformer for RML2016.10a: embed I/Q segments as patches, add a class token, train at a parameter count matched to your Phase 3 ResNet.
• Compare accuracy vs SNR curves against your CNN and ResNet - three-way plot, honest conclusion about where attention helps and where it does not.
• Visualize attention maps on a few examples; note what the model attends to at low SNR.

Checkpoint

You can articulate when a transformer beats a CNN on signals (long-range structure, large data) and when it loses (small data, strong local priors).

Labeled RF data is scarce; unlabeled RF is infinite. This is arguably the most practically valuable idea in ML for RF right now.

Study

• SimCLR (contrastive learning) and Masked Autoencoders - read both; you will implement the second.

Exercises

• Pretrain a masked autoencoder on unlabeled I/Q frames (your Phase 2 synthetic generator, cranked up to hundreds of thousands of frames, plus RML data stripped of labels).
• Linear-probe evaluation: freeze the encoder, train a linear head with only 1 percent and 10 percent of RML labels; compare against training your Phase 3 model from scratch on the same label budgets.
• Plot accuracy vs label fraction for both approaches - the signature plot of self-supervised learning.

Checkpoint

At the 1 percent label budget, your pretrained encoder should clearly beat from-scratch training, and you can explain why the gap closes as labels grow.

Everything so far was curated data. Real spectrum is messier, and the synthetic-to-real gap is where deployed RF-ML lives or dies. Hardware week: you need the RTL-SDR Blog v4 from here on.

Study

• PySDR's capture and hardware chapters (you met them in Week 4; now use them for real).
• Skim local band plans so you know what you are hearing: FM broadcast, ADS-B at 1090 MHz, POCSAG paging, ISM-band traffic.

Exercises

• Write a capture pipeline: tune, record I/Q to disk with metadata (frequency, sample rate, time), slice into frames.
• Curate a small labeled dataset of 4-6 over-the-air signal classes you can reliably identify; document your labeling procedure.
• The domain-gap experiment: test your best RML-trained model on your captures, then fine-tune on a small labeled split. Quantify the gap before and after.
• Optional with a PlutoSDR or HackRF One: generate your own modulated signals and loop them back over a coax cable into the RTL-SDR - a controlled bridge between synthetic and over-the-air. Transmit only into cable or dummy load unless you hold an amateur license.

Checkpoint

You have a dataset nobody else has, and a measured number for how much synthetic training transfers to reality.

Study

• PyTorch ONNX export and post-training quantization docs.
• GNU Radio Embedded Python Blocks - how to put your model inside a flowgraph.

Exercises

• Export your best model to ONNX; quantize to int8; measure the accuracy and latency changes at each step.
• Build a GNU Radio flowgraph: RTL-SDR source, windowed framing, your model in an embedded Python block, live prediction display.
• Measure sustained throughput; it must keep up with the sample rate without dropping frames on your laptop.

Checkpoint

A window on your screen names the modulation of live off-air signals in real time. Show someone; this is the demo that lands jobs.

Harden the live classifier into something honest about its own limits - the difference between a demo and a tool.

Tasks

• Add out-of-distribution detection (an energy-based or Mahalanobis score on features) so the system flags "unknown signal" instead of guessing confidently on things it never trained on.
• Validate: feed it a signal class you deliberately held out of training and confirm it abstains rather than misclassifies.
• Write the project README: architecture, dataset, domain-gap numbers, live throughput, OOD behavior, honest failure cases.

Definition of done

A stranger with an RTL-SDR could clone the repo, run one command, and watch live classifications with abstention on unknowns.