When Deep Learning Is Overkill

After three pages of neural fireworks, Iris asks the data team an innocent question: so are we replacing the churn model with a deep network? The answer is a flat no — and delivered with the confidence of people who have tried. Plateful's most valuable model is still the churn random forest from Chapter 7, trained on a table of engineered features, and nobody on the team is embarrassed about it.

This page is the chapter's honest epilogue. Deep learning did not make Chapters 4 through 7 obsolete; it conquered a specific territory — and outside that territory, the simple tools remain the professional choice. Topic 19 planted "try simple first" as a habit. This page closes the arc with the reasons, so you can tell where depth earns its cost and where it is a marvel pointed at the wrong problem.

Deep Learning's Home Turf

The last two pages drew the boundary without saying so. Pixels, audio, language — raw perception — is where deep learning lives, and it lives there for one precise reason: the features cannot be written by hand. Nobody can type out "noodle-ness" as columns, and no numbering of words carries their meaning. The telling facts must be manufactured from the raw material, layer by layer, and learned features (Topics 37 and 38) are the only method that has ever really worked.

On that turf, deep learning is not overkill — it is the only thing that works at all. If Plateful wants to check dish photos, transcribe support calls, or match reviews by meaning, the conversation starts and ends with a neural network. No honest engineer disputes the territory; the dispute starts when the fireworks get pointed at a table.

Tables Resist the Fireworks

Plateful's churn data is a table: orders per month, days since last order, discount usage, support contacts — columns that Topic 31's craft already distilled from raw logs. And here is the fact that surprises every newcomer: on data like this, ensembles of trees — Topic 20's random forest and its boosted cousins — routinely match or beat deep networks. Not occasionally. Routinely, across most business prediction problems on structured data.

The reason is almost anticlimactic. Deep learning's superpower is manufacturing features from raw perception — but in a good table, that work is already done. "Days since last order" is the distilled signal; there is nothing deeper for layers to discover under it. The network brings a feature-factory to data that arrives pre-featured, pays the full price for the machinery, and finds little left to earn with it.

The Bill for Depth

And the price is real, on every axis at once. A deep network is data-hungry — Topic 36's millions-need-millions logic — where a forest is content with the tens of thousands of rows a churn table actually has. It wants GPUs where a forest trains on an ordinary machine in minutes, which also means slower experiments and fewer ideas tested per week. And it surrenders something Chapter 10 will make weigh far more than aesthetics: Topic 19's tree could be read end to end, and even the forest can report which features drive it — while a million weights explain nothing on their own, and "why was this customer flagged?" may be a question regulators, auditors, or angry customers demand answered.

None of these costs is a reason never to go deep. They are the bill — and a bill is paid gladly when the purchase is face unlock or meaning-matched reviews. The failure mode is paying it for nothing: same accuracy as the forest, months later, at several times the cost, with the explainability gone. Complexity is a cost you pay, not a virtue you claim.

Photos, audio, or language — raw perception?→Deep learning — features must learn themselves

A table of engineered features — orders, customers, churn?→Simple first — linear, tree, random forest

Simple tried honestly — and measurably short of the bar?→Escalate — on evidence, never on fashion

The Professional Decision

Think of a helicopter and a bicycle, and a bakery two blocks away. The helicopter is a genuine marvel — faster, mightier, more sophisticated in every measurable way — and it is the wrong vehicle. It costs more to start than the errand is worth, needs a crew and a landing pad, and arrives no sooner. Nothing about its sophistication makes it better at two blocks. Match the vehicle to the errand, not to the brochure.

The professional decision is exactly that mapping: data kind and stakes first, tool second. Perception — go deep, it's the only road. Tables — simple first, and escalate only when an honest evaluation (Chapter 6's tools, not enthusiasm) shows the simple model falling short of the product bar. And when a vendor or a colleague says "we used deep learning" as if it settled the question of quality — you now know it settles nothing. It is a cost report, not a quality claim. Whether the cost bought anything is a separate question, answered with held-out data, the same as ever.