ADR-006: Stay in DSP, skip pattern-recognition / ML topics

Status

Accepted

Context

Exposure to a real time-series pattern recognition curriculum (a 2nd-year BSc Signal Processing Systems elective) made it visible which feature-extraction topics the dsp-workshop is missing compared to a typical PR course:

Time-frequency: STFT, wavelets, MFCC
Spectral features: centroid, bandwidth, rolloff, flux
Statistical descriptors of windowed time-series
Dimensionality reduction: PCA
Classifiers: k-NN, DTW
Deep learning: RNN / LSTM / GRU / CNN
Evaluation metrics: precision, recall, F1

This raises the question of how far the workshop should extend into pattern recognition territory.

Constraints: - ADR-003 frames the workshop as DSP, with basics (learning path) and topics (research-flavoured reference). The tagline is “DSP, from math to metal”, not “Pattern Recognition, from math to metal”. - The repo is maintained by one author whose centre of gravity is classical DSP, multirate, and embedded signal processing, not deep learning. - Topic pages are written at a quality bar (math, Python, embedded companion, open questions) that does not scale to fashionable-but-shallow ML coverage.

Decision

1. Extend in DSP-flavoured feature extraction only

Add the following topic pages, in this priority order. Each fits the existing “math + Python + (optional) embedded.qmd + open questions” template:

STFT (foundational time-frequency transform)
Spectral features (centroid, bandwidth, rolloff, flux; small consolidation topic)
Wavelets (PhD-era source material available in _raw/)
MFCC (with a wakeword + DTW running example as the embedded companion)
Statistical features for windowed time-series (mean, variance, skew, kurtosis, energy; short consolidation page)

See docs/topic-roadmap.md for sources, sections, and demo ideas per topic.

2. Do NOT add standalone PR / ML topics

The following are explicitly out of scope as standalone topic pages:

Classifiers: k-NN, SVM, decision trees, and similar
Dimensionality reduction: PCA, t-SNE, UMAP (despite their usefulness in feature analysis)
Deep learning: RNN, LSTM, GRU, CNN
Evaluation metrics: precision, recall, F1, ROC

The justification is identity, scope, and authority: this is a DSP workshop, the maintainer’s strongest contribution is on the signal side, and good treatments of these topics already exist elsewhere (scikit-learn docs, fast.ai, Pete Warden’s TinyML book).

3. Classical, training-free PR is allowed as supporting material

Methods that are training-free, template-based, and DSP-adjacent (DTW, matched filtering, simple thresholding rules, hand-crafted classifiers built from spectral or statistical features) MAY appear inside an embedded.qmd or topic page as part of an end-to-end example. They do not get their own dedicated topic pages. The wakeword + DTW example under MFCC is the reference case.

4. Pointers, not coverage, for ML topics

Where a topic naturally invites a deep-learning continuation (MFCC + CNN for keyword spotting, PPG + LSTM for arrhythmia, anomaly detection + autoencoder), the topic page ends with a “What next” or “Open questions” section that names the technique and points at an authoritative external reference. No tutorials, no code, no claims of expertise we do not have.

Consequences

Positive: the workshop gains credible PR-adjacency without losing its DSP identity; pages can be used as companion material for PR courses that build on DSP fundamentals.
Positive: scope stays bounded; five new topics is a finite amount of work and each one stands alone.
Positive: the maintainer writes in their strongest idiom, keeping the quality bar where it has been.
Negative: students wanting a complete PR pipeline must read elsewhere for the classifier and deep-learning halves; the workshop is openly half a story for that audience.
Negative: requires discipline to refuse drift toward “just one more ML topic” when a related opportunity (lecture, contribution, request) presents itself.

# ADR-006: Stay in DSP, skip pattern-recognition / ML topics ## Status Accepted ## Context Exposure to a real time-series pattern recognition curriculum (a 2nd-year BSc Signal Processing Systems elective) made it visible which feature-extraction topics the dsp-workshop is missing compared to a typical PR course: - Time-frequency: STFT, wavelets, MFCC - Spectral features: centroid, bandwidth, rolloff, flux - Statistical descriptors of windowed time-series - Dimensionality reduction: PCA - Classifiers: k-NN, DTW - Deep learning: RNN / LSTM / GRU / CNN - Evaluation metrics: precision, recall, F1 This raises the question of how far the workshop should extend into pattern recognition territory. Constraints: - ADR-003 frames the workshop as DSP, with basics (learning path) and topics (research-flavoured reference). The tagline is "DSP, from math to metal", not "Pattern Recognition, from math to metal". - The repo is maintained by one author whose centre of gravity is classical DSP, multirate, and embedded signal processing, not deep learning. - Topic pages are written at a quality bar (math, Python, embedded companion, open questions) that does not scale to fashionable-but-shallow ML coverage. ## Decision ### 1. Extend in DSP-flavoured feature extraction only Add the following topic pages, in this priority order. Each fits the existing "math + Python + (optional) embedded.qmd + open questions" template: 1. STFT (foundational time-frequency transform) 2. Spectral features (centroid, bandwidth, rolloff, flux; small consolidation topic) 3. Wavelets (PhD-era source material available in `_raw/`) 4. MFCC (with a wakeword + DTW running example as the embedded companion) 5. Statistical features for windowed time-series (mean, variance, skew, kurtosis, energy; short consolidation page) See `docs/topic-roadmap.md` for sources, sections, and demo ideas per topic. ### 2. Do NOT add standalone PR / ML topics The following are explicitly out of scope as standalone topic pages: - Classifiers: k-NN, SVM, decision trees, and similar - Dimensionality reduction: PCA, t-SNE, UMAP (despite their usefulness in feature analysis) - Deep learning: RNN, LSTM, GRU, CNN - Evaluation metrics: precision, recall, F1, ROC The justification is identity, scope, and authority: this is a DSP workshop, the maintainer's strongest contribution is on the signal side, and good treatments of these topics already exist elsewhere (scikit-learn docs, fast.ai, Pete Warden's TinyML book). ### 3. Classical, training-free PR is allowed as supporting material Methods that are training-free, template-based, and DSP-adjacent (DTW, matched filtering, simple thresholding rules, hand-crafted classifiers built from spectral or statistical features) MAY appear inside an `embedded.qmd` or topic page as part of an end-to-end example. They do not get their own dedicated topic pages. The wakeword + DTW example under MFCC is the reference case. ### 4. Pointers, not coverage, for ML topics Where a topic naturally invites a deep-learning continuation (MFCC + CNN for keyword spotting, PPG + LSTM for arrhythmia, anomaly detection + autoencoder), the topic page ends with a "What next" or "Open questions" section that names the technique and points at an authoritative external reference. No tutorials, no code, no claims of expertise we do not have. ## Consequences - **Positive**: the workshop gains credible PR-adjacency without losing its DSP identity; pages can be used as companion material for PR courses that build on DSP fundamentals. - **Positive**: scope stays bounded; five new topics is a finite amount of work and each one stands alone. - **Positive**: the maintainer writes in their strongest idiom, keeping the quality bar where it has been. - **Negative**: students wanting a complete PR pipeline must read elsewhere for the classifier and deep-learning halves; the workshop is openly half a story for that audience. - **Negative**: requires discipline to refuse drift toward "just one more ML topic" when a related opportunity (lecture, contribution, request) presents itself.