AI and ML Learning — the written record and what is next

Projects Tracker · AI / ML learning · track 1 of 5 ReMapping →

The status page for one long project: working through Andrew Ng's two specializations and turning each part into a written note in the engine folder Machine Learning Notes/. The notes are the deliverable, not the videos — so progress is measured in what exists as prose.

In one line. The ML Specialization is written up in full — 13 notes, 164 sections, ~106,500 words. The Deep Learning Specialization is the live half: its Course 1 is covered by the network reference, and Course 2 is watched up to RMSprop but not written. Next: dl-regularization.md, from source already captured.

Contents

Where this stands
The written record
The gap — Deep Learning Course 2
What the remaining courses are worth
The next step
How a part gets written
Where the sources are
How the two folders stay in step
Open question — interactive plots

Where this stands

Two programmes, eight courses. The first is finished as far as the writing goes; the second is the live one.

Programme	Course	Became	State
ML Specialization	1 · Supervised learning: regression and classification	notes 1–2 — linear regression, logistic regression	written
	2 · Advanced learning algorithms	notes 3–6 — the network reference, softmax, model development, decision trees	written
	3 · Unsupervised learning, recommenders, RL	tracks 2–4 — clustering, anomaly detection, collaborative filtering, RL, deep Q-learning	written
DL Specialization	1 · Neural networks and deep learning	the reference, Parts I–IV, plus appendices A and B	written
	2 · Improving deep neural networks	track 5 — three planned `dl-` notes	watched, not written
	3 · Structuring machine learning projects	—	not started most transferable to the RF goal
	4 · Convolutional neural networks	—	not started optional for the RF goal
	5 · Sequence models	—	not started optional for the RF goal

Table 1: Every course in both specializations, what it turned into, and its state. The neural-network reference does double duty: it is the deep end of ML Specialization Course 2 and the whole of Deep Learning Specialization Course 1.

Watched, not written is a backlog, not progress: the source is captured, the record is not made. Only written counts as done.

The written record

Counted live from the notes folder every time this page loads — there is no copy of these numbers to go stale, and a note added there appears here on the next reload.

13 notes, 164 sections, 106,505 words, 74 Chart.js figures, 56 NeuralFlow diagrams, 843 KB.

Track	#	Note	Sections	Words	State
1 · classical spine	1	`linear-regression.md`	15	16,035	complete
1 · classical spine	2	`logistic-regression.md`	8	7,751	complete
1 · classical spine	3	`neural-networks-reference.md`	41	35,095	complete
1 · classical spine	4	`multiclass-softmax.md`	12	5,716	complete
1 · classical spine	5	`model-development.md`	14	6,637	complete
1 · classical spine	6	`decision-trees.md`	13	5,605	complete
2 · unsupervised	1	`ul-clustering.md`	9	4,118	complete
2 · unsupervised	2	`ul-anomaly-detection.md`	10	3,455	complete
3 · recommenders	1	`rec-collaborative-filtering.md`	8	3,413	complete
4 · reinforcement learning	1	`rl-foundations.md`	9	3,798	complete
4 · reinforcement learning	2	`rl-deep-q-learning.md`	10	4,129	complete
appendices	A	`backpropagation.md`	9	4,147	superseded by Part IV
appendices	B	`fbnn-reorganized.md`	6	6,606	superseded by Part IV
5 · deep learning	1	`dl-regularization.md`	0	0	planned
5 · deep learning	2	`dl-optimization.md`	0	0	planned
5 · deep learning	3	`dl-tuning-batchnorm.md`	0	0	planned

The gap — Deep Learning Course 2

The only track with a backlog. The material has been studied; what is missing is the write-up. The course was followed video by video into numbered LaTeX notes. Those are your own writing, not course material, so they live outside the engine at ~/Documents/My LaTeX Video Notes/ — the archive holds what Coursera published, and for Course 2 that is nothing.

Audited 24 August 2026, and they hold up better than a file count suggested. Every note compiles to a 2–5 page PDF of 270–1,000 words. Nothing is a stub or a placeholder, and nothing is missing an image — the figures are drawn natively in TikZ and pgfplots rather than pulled in as picture files, so each PDF is self-contained and portable. Only four .png files exist across the whole set, and all four resolve.

Week	Becomes	Subject	Captured	Written
1	`dl-regularization.md`	train/dev/test, bias and variance, $L_2$ and dropout, normalization, initialization, gradient checking	14 of 14 lectures complete 45 pp · 7,400 words · 25 figures	nothing
2	`dl-optimization.md`	mini-batch, exponentially weighted averages, momentum, RMSprop, Adam, learning-rate decay, local optima	7 of 10 lectures to RMSprop 26 pp · 5,100 words · 9 figures	nothing
3	`dl-tuning-batchnorm.md`	hyperparameter search, batch normalization, softmax at scale, frameworks	none — 11 lectures to watch	nothing

Table 2: Course 2 audited against the official Coursera syllabus rather than by counting files. Lecture counts exclude the guest interviews Coursera lists as videos - one in week 1 and one in week 2 - which is why week 1 is 14 lectures out of 15 videos. Week 2 stops after RMSprop, so Adam is the exact point to resume from.

Course 2 Week 1 needs nothing watched. All fourteen of its lectures are covered — one of them twice, since Dropout Regularization was written up in two framings (inverted dropout, and a layer-$\ell$ example), which is why the folder holds fifteen files for fourteen lectures. At 7,400 words of source for a note that will land around 6,000–8,000 in house style, this is a one-to-one basis to write from, not a skeleton. dl-regularization.md is writable today.

What is genuinely missing, checked against the official syllabus on 24 Aug 2026 — three lectures to finish Week 2, then eleven for Week 3. That is 14 videos, about 75 minutes:

The two guest interviews (Yoshua Bengio in Week 1, Yuanqing Lin in Week 2) are listed by Coursera as videos but are not lectures, and are not counted above.

Write from ~/Documents/My LaTeX Video Notes/Video notes/, not from Working copies/. The working copies are three days newer and thinner: on Why Regularization Reduces Overfitting the numbered version has six TikZ figures and the working copy has none, and the same pattern holds for dropout and for input normalization. Newer did not mean better. Keep both; write from the numbered set.

Settled on 26 Jul 2026, do not re-decide. These are three dl- notes in the notes folder — not new Parts on the network reference, and not a separate folder. The reasoning is in Machine Learning Notes/CLAUDE.md. Overlaps are linked, never re-derived: bias and variance already have a section in model-development.md, $L_2$ in linear-regression.md.

What the remaining courses are worth

Three courses remain after Course 2, and they are not equally worth the months. Measured against what is actually being built here — the notes in this folder, and the 5G Systems track next door — they differ a lot.

First, the thing to be clear about: no course in the Deep Learning Specialization teaches radio. No 5G, no NR, no antenna switching, no insertion loss — not in Course 2, and not in 3, 4 or 5. Ng teaches deep learning. The 5G track has its own sources and shares nothing with this one except the maths.

Course	Teaches	Worth doing?
C2 · Improving Deep Neural Networks	Regularization, optimization (Adam), tuning, batch norm, frameworks	yes The baseline craft — can you actually make a model train. Finish it; the notes are already planned.
C3 · Structuring ML Projects	Orthogonalization, single-number evaluation metric, satisficing vs optimizing metrics, train/dev distribution mismatch, human-level performance, error analysis, transfer and multi-task learning	the best of the four It is about defining the metric and diagnosing the system — which is what a systems architect does. Satisficing versus optimizing metrics is a genuinely reusable idea — it is how you frame any "maximise X subject to a budget on Y" problem. Short course. Do it.
C4 · Convolutional Neural Networks	Convolutions, ResNets, object detection, face recognition, style transfer	optional Computer vision. Excellent in its own right; simply not what anything else in this engine is about.
C5 · Sequence Models	RNNs, GRU/LSTM, word embeddings, attention, transformers	optional Same story — taught through NLP and speech, and unconnected to the rest of the work here.

Table 3: The Deep Learning Specialization course by course. Course 3 is the surprise - it is the most transferable of the four, because it is about defining a metric and diagnosing a system under constraints, which is what the rest of this folder is about.

So: finish C2, then C3. Treat C4 and C5 as optional. They are excellent courses and worth doing for their own sake — they are simply several months, and nothing else in this engine depends on them.

And note what the specialization never covers: reinforcement learning. It is not in any of the five courses. You already have it, from the Machine Learning Specialization's Course 3 — rl-foundations.md and rl-deep-q-learning.md are written and complete.

The next step

Write dl-regularization.md from the captured Deep Learning Course 2, Week 1 notes — the source is complete, nothing needs watching first.
Wire it in: nav strip top and foot, a card and lookup rows in the notes folder's index.md, and a new track section for deep learning.
Resume the videos at Course 2 Week 2 — Adam optimization (the first uncaptured one).
Then write dl-optimization.md once Course 2, Week 2 is captured through learning-rate decay.
Keep the Coursera subscription active while the specialization is open.

How a part gets written

The full rules — house style, notation, the verify loop, the gotchas — are in Machine Learning Notes/CLAUDE.md and the engine's own CLAUDE.md. Read both before appending anything. The three that are non-negotiable: one part per pass, the existing notation, and nothing from outside the two specializations.

Figure 1: The loop for one part. The check is on notation and numbers together, and it comes before the note is wired into the hub — a section that fails it is rewritten, not patched afterwards.

Where the sources are

md-engine/Courses (RAW)/ is the course archive — raw material for several projects, not just this one, kept in the engine folder so it sits beside the notes written from it. Like the notes folder, it reports itself: the rows below are walked from disk on every page load, so a deck dropped in there shows up here on the next reload.

Your own notes live outside the engine. ~/Documents/My LaTeX Video Notes/ — 98 files, 9.3 MB — holds the numbered LaTeX write-ups (see §3), the thinner working copies, a command LaTeX document, a recovered sigmoid_simulation.ipynb, and the old two-line Master Tracker that this page replaces. They were briefly filed inside the archive on 24 Aug 2026 and moved back out the same day: the archive is for material the course published, and these are yours.

What is still missing is Coursera's own Course 2 slide decks — the archive holds Course 1's Week1–4.pdf and nothing for Course 2, so those LaTeX notes are the only Course 2 source there is. Enough to write from; not the printed deck.

Two different Week1–4.pdf sets exist, one in the archive above and one in ~/Documents/Notion/AI_Learning/Neural_Network/. Same names, different files, different page counts (Course 1's Week 2 deck is 51 pages in one set against 48 in the other). Both are handwritten scans of Course 1, not Course 2. Which one the reference was written from is not recorded — worth settling before Parts I–III are ever revised against a source.

How the two folders stay in step

Nothing on this page is copied from the notes folder — it is read from it. The engine widget widgets/ml_status.py runs server-side on every page load, so the table in §2 is counted fresh each time. There is no sync step and no command to remember.

Direction	What crosses	Where it shows
notes → here	every `.md` scanned for sections, words, charts and diagrams, joined with that folder's `status.json` for track and state	§2, and the summary line
here → notes	the live tick boxes under The next step	the "Where the work is" card on the notes hub

Table 4: What each side sends the other. The two folders are served by separate servers and cannot link to each other, so a widget reads across instead.

So a note added, renamed or deleted over there shows up here on the next reload, and a step ticked here shows up over there. If the notes folder's status.json has not been told about a new file, §2 says so in red rather than quietly under-reporting.

What the widget cannot know. It carries facts, not judgement. The course table (§1), the gap table (§3) and the next step (§4) are hand-written and still need updating when the situation changes — as does headline in status.json. The rule is written into Machine Learning Notes/CLAUDE.md.

Two things would break the link: a widgets/ folder inside either served folder (the engine replaces its default widgets rather than merging, so the placeholder would render as "Widget not found"), and the Pi, where ml_status.py has to be copied into ~/mdengine/engine/widgets/ by hand like any engine file — mdpush carries notes, not the engine. Open the notes with their own server:

mddash ~/Documents/md-engine/"Machine Learning Notes"

Open question — interactive plots

Can a note carry a slider-driven figure? The pieces exist: the engine runs Python widgets, renders Chart.js configs, and style.css ships a .slider-grid class. What is unsettled is whether the slider should trigger a server round-trip to Python or simply re-plot in JavaScript.

The client answer is cheaper and almost certainly right for a teaching figure — and it is the only one that survives an exported bundle, where there is no server at all. A Python round-trip only earns its cost when the browser genuinely cannot do the recomputation.

Decide: client-side re-plot, or Python widget round-trip.
Prototype one slider-driven figure — learning rate $\alpha$ against the cost curve, using data that already exists in note 1.

Projects Tracker · AI / ML learning · track 1 of 5 ReMapping →