How deep learning works
Older machine learning needed a person to decide which features mattered. If you wanted software to judge whether an image showed a product, someone had to tell it to look at edges, colours or shapes. Deep learning removes that step. It stacks many layers of very simple calculations, each passing its result to the next, and lets the training process work out for itself what each layer should pay attention to.
Early layers end up detecting crude things — an edge, a change in brightness, a common letter pattern. Later layers combine those into something meaningful, such as a face, a logo or the likely next word in a sentence. No single layer understands anything. The behaviour comes from the arrangement, which is why the models are called neural networks and why the field needed both large datasets and a lot of computing power before it became practical.
Why deep learning matters
It is the reason software can now handle messy, unstructured input. Text, photographs, audio and video used to be things a computer could only store, not interpret. Deep learning is what put automatic captions, image search, voice assistants and language models within reach, and every generative tool a marketer uses today rests on it.
In day-to-day marketing you rarely touch it directly. You meet its output: ad platforms grouping creative by what it depicts, transcription of a recorded call, alt text suggestions, or an assistant that answers a question about a document. Understanding that these are all predictions from patterns — not comprehension — sets a sensible expectation of when to trust them.
Common mistakes with deep learning
The biggest is treating confident output as reasoning. A model that labels an image or answers a question is producing the most probable response given its training, not checking a fact. It has no way to signal that a case is unlike anything it has seen, so unusual inputs get an equally assured answer.
The second is expecting an explanation. Because the useful behaviour is spread across many layers, nobody can point to the reason for a particular output the way you can trace a rule in ordinary software. That matters when a decision affects a person — a loan, a job application, a refused claim — and it is why regulated work needs a human decision-maker on the record rather than a model’s verdict.
How to act on it
Judge these tools on your own material, not on demonstrations. Run a batch of your real product photos, your recorded calls in the accent your customers actually have, or your own documents through the tool and count the errors yourself. Performance on English written by a copywriter tells you little about performance on Nepali, Romanised Nepali or a mixed-language message.
Then decide where a wrong answer is cheap and where it is expensive, and only automate the cheap side. Suggested alt text, draft transcripts and creative tagging can be reviewed quickly. Anything customer-facing needs a person in front of it, which is the same principle that applies to generative AI generally and shapes how I approach AI consulting work.