← All artifacts

// Artifact · Applied Research

Machine Learning vs. Deep Learning — A Decision Framework

Two real-world case studies where the "wrong" approach was actually tried and lost — and the one question that decides between them

Applied ResearchML vs DLCase StudiesTechnical Writing
Open full experience

// 01 · Introduction

Introduction

The usual answer to "when should I use deep learning?" is about data volume and compute cost — true, and almost never the real reason. This Workshop 2 report argues the distinction that actually decides the choice: whether the features that predict the outcome can be written down by a human in advance. It defends that principle with two production case studies where the losing approach wasn't dismissed on theory — it was genuinely attempted, and the evidence went the other way. Read the full report below.

// 02 · Description

Description

Case 1 — customer churn prediction (classical ML wins): on the canonical IBM/Kaggle Telco dataset (~7,000 subscribers), the features are already human-legible business facts — contract type, tenure, payment method — so there is no hierarchy left for a deep network to discover, and interpretability is a deployment requirement because retention teams must know why a customer is at risk. Frohböse (2020) benchmarked KNN, logistic regression, random forest, and SVM, then deliberately added a feed-forward neural network: it scored best (0.7996 accuracy, 0.5948 F1) but only marginally, at a heavy cost in interpretability and training complexity. Case 2 — diabetic retinopathy screening (deep learning wins): the predictive features (microaneurysms, haemorrhages, exudates) live in raw retinal pixels that no human can articulate as columns; hand-crafted classical approaches were attempted for roughly two decades before being conceded as theoretically incapable, and a CNN-based system, IDx-DR, became the first FDA-approved autonomous AI diagnostic in 2018. The report closes by generalizing the governing principle and its downstream consequences for data volume, interpretability, and cost.

// 03 · Objective

Objective

To move past the folklore answer on ML vs. deep learning and build a defensible decision framework for the Workshop 2 assignment — one grounded in documented deployments and honest about counter-evidence, including the churn author's own caveat about larger datasets, rather than argued from first principles alone.

// 04 · Process

Process

01Framed the report around one governing question — can a domain expert enumerate the predictive features in advance? — and committed to testing it against real deployments, not toy examples.
02Selected two case studies with an unusual property: in each, the losing approach was actually attempted on the same problem, so the comparison is evidence rather than assertion.
03Analyzed the Telco churn literature — including Frohböse's 2020 benchmark where a neural network was explicitly added to a classical lineup and won only marginally — and connected the interpretability requirement to how retention teams actually consume model output.
04Analyzed the diabetic retinopathy case: two decades of hand-crafted feature engineering, the theoretical reason it could not succeed, and the CNN-based IDx-DR system that earned autonomous FDA clearance in 2018.
05Addressed the strongest counter-arguments honestly, including why more data does not rescue deep learning for churn (the interpretability requirement does not shrink as data grows).
06Wrote and typeset the report, then published it here with the full PDF for scrutiny.

// 05 · Tools & Technologies

Tools & Technologies

Literature analysis (published benchmarks, FDA announcements, deployment studies)IBM/Kaggle Telco churn dataset context (~7,000 subscribers)Classical ML methods analyzed: logistic regression, KNN, random forest, SVMDeep learning methods analyzed: feed-forward networks, CNNs (IDx-DR)Technical writing & argumentation

// 06 · Value Proposition

Value Proposition

Unique Value

Most ML-vs-DL comparisons list pros and cons; this one advances a single falsifiable principle and stress-tests it against two deployments where the opposite approach was really tried. It also demonstrates intellectual honesty — the report includes the churn author's own caveat that cuts against its thesis, then answers it — which is the skill that separates engineering judgment from advocacy.

Relevance to My Audience

Choosing the wrong model family is one of the most expensive mistakes an AI team can make, and I make this call in practice on ingestion and ML systems at Amazon. For hiring managers this artifact shows I can reason about when NOT to use deep learning — a rarer and more valuable signal than knowing how to use it; for students and practitioners it is a reusable decision framework with citations to follow.

// 07 · References

References