Deep learning or classic machine learning? A practical way to choose
Tables of numbers usually want gradient boosting. Images, sound and free text usually want deep learning. The amount of labelled data and where the model has to run decide the rest.
Deep learning gets the headlines, and for good reason: it is what lets software see, listen and read. But it is not the right tool for every prediction. For many business problems a classic method is faster to build, cheaper to run and easier to explain, and it is just as accurate.
Start with the shape of the data
- Tables: rows of customers, orders, transactions, sensor readings with named columns. Classic methods such as gradient-boosted trees are usually the strongest choice.
- Images and video: photos, scans, camera feeds. Deep learning, almost always, usually starting from a pretrained model.
- Audio: calls, voice notes, machine sounds. Deep learning, often a pretrained speech or audio model.
- Free text: emails, reviews, chat messages. Pretrained language models, sometimes with a small classic model on top.
The type of data answers most of the question. Data size and where the model runs answer the rest.Then look at labelled data
Deep networks learn a lot from a lot of examples. With a few hundred labelled rows, a classic model or a pretrained network with a small amount of fine-tuning will do better than training anything large from scratch.
A rough guide to choosing an approach (rules of thumb, not hard limits).| Situation | Usually choose | Why |
|---|
| A few thousand table rows | Gradient-boosted trees | Strong, fast, explainable |
| A few hundred labelled images | Fine-tune a pretrained model | Reuses what it already learned |
| Tens of thousands of labelled images | Fine-tune a larger pretrained model | Room to learn your specifics |
| Free text, few labels | Pretrained language model | Understands language out of the box |
| Must run on a phone or small device | Small, compressed model | Speed and memory limits |
Where it runs matters as much as accuracy
- A model behind a website can use a server GPU. A model inside a phone app has to be small and fast.
- A model reading every frame from ten cameras needs very different hardware from one scoring a hundred loan applications a day.
- Training once on a rented GPU is cheap. Running a large model all day, every day, is where the cost really sits.
Explainability is a feature
When someone has to act on it
When a credit officer or a doctor has to act on a prediction, being able to say why matters. Classic models show which inputs pushed a score up or down quite directly. Deep models can be explained too, but with more effort and less certainty. That alone can decide the choice.
Choose the simplest model that meets the target, on the hardware it will actually run on.
Frequently asked questions
Do we need our own GPU?
For training, usually not: GPUs can be rented by the hour. For running a vision or speech model all day, a GPU on site or a dedicated server is often cheaper than paying per request.
Can we train on our own data without sharing it?
Yes. Models can be fine-tuned and run entirely on your own hardware, so the data never leaves your network.
Is deep learning more accurate?
On images, sound and text, by a wide margin. On ordinary business tables, often not; gradient-boosted trees remain hard to beat there.
Can you start with one approach and switch later?
Yes. We often start with a quick classic model as a baseline and move to a deep model only where the data and the gain justify it.