DSP 101
M12 · L04
Module 12: Capstone & Real-World DSP

Machine Learning Meets DSP

Neural networks now live inside every layer of the signal-processing stack — but classical DSP still runs the show where interpretability and reliability matter most.

01 / 11
DSP 101
M12 · L04
Two Paradigms

Model vs. Data

Classical DSP
First-principles math — optimal when signal statistics are known
Machine Learning
Learned from data — adapts when statistics are unknown or non-stationary
02 / 11
DSP 101
M12 · L04
Neural Denoising

Dilated Convolutions

Stack 1-D convolutions with dilation factors 1, 2, 4, …, 512. Receptive field grows exponentially with depth while parameters grow linearly.

Dilated Convolution
(x *_d h)[n]=\sum_{k=0}^{K-1}h[k]\,x[n-d\cdot k]
03 / 11
DSP 101
M12 · L04
Modulation Classification

IQ Samples In, Labels Out

  • Feed raw complex IQ samples into a CNN
  • Learns richer features than any hand-crafted cumulant extractor
  • RadioML: >95% accuracy across 24 modulation classes at SNR > 10 dB
  • Penalty: black-box decisions, needs labeled training data
  • Classical AMC still wins at low SNR and in regulated spectrum audits
04 / 11
DSP 101
M12 · L04
Learned Equalization

RNN & Transformer Equalizers

Sequence models learn end-to-end channel inversion — no explicit channel estimate needed. Outperform MMSE in high-mobility channels. Fail catastrophically out-of-distribution.

+
High-mobility channels
⚠
New channel types
✓
Classical degrades gracefully
05 / 11
DSP 101
M12 · L04
Model-Based Deep Learning

Algorithm Unrolling

  • Convert an iterative algorithm (ISTA, turbo decoder) into a fixed-depth network
  • Replace fixed parameters with learned weights
  • Each layer has a clear signal-processing meaning
  • LISTA: 10 layers ≈ 1000 ISTA iterations — 100× speedup
  • Best of both worlds: interpretability + data-driven optimization
06 / 11
DSP 101
M12 · L04
When to Use What

Classical DSP Still Wins

Use Classical DSP When
Safety-critical · standards-compliant · real-time deterministic · no training data
Use ML When
Plentiful labeled data · interpretability not required · complex non-stationary statistics
07 / 11
DSP 101
M12 · L04
Hybrid Systems

DSP + ML Pipeline

  • Apply ML where statistics are hard to model analytically
  • Retain classical DSP for analytically tractable stages
  • Learned filter banks: trainable STFT windows fine-tuned end-to-end
  • MFCCs (mel scale + log + DCT): the classic hand-crafted speech feature these learned banks replace
  • Neural beamformers: MVDR replaced by complex-valued CNN
  • Deep unfolded OFDM receivers: iterative channel estimation refinement
08 / 11
DSP 101
Knowledge Check

Check whatstuck

Four questions on ML and DSP — dilated convolutions, neural equalizers, algorithm unrolling, and when classical DSP still wins.

Question 1 of 0
Score 0/0

09 / 11
DSP 101
M12 · L04
Key Takeaways

Six Things to Remember

  • ML excels at unknown/non-stationary statistics; classical DSP at known models
  • Dilated CNNs: exponential receptive field, linear parameter count
  • AMC with CNNs surpasses hand-crafted cumulants at high SNR
  • Neural equalizers beat MMSE in high-mobility — but fail out-of-distribution
  • Algorithm unrolling yields interpretable, data-efficient hybrid networks
  • Classical DSP is still mandatory for safety-critical certified systems
10 / 11
DSP 101
Complete
Module 12 Complete

DSP-101 Finished

From the Nyquist theorem to neural equalizers — you now have a complete picture of digital signal processing, from mathematical foundations to real-world systems.

Back to Overview
DSP-101 Course Page →
All lessons Read article
11 / 11