Deep learning

What is deep learning?

Deep learning is machine learning with neural networks that have many layers between input and output. The word deep refers to those layers. Large modern models can have dozens or hundreds of them.

The layers are the point. Each layer learns patterns from the previous layer. In image recognition, early layers may detect edges and simple shapes. Later layers combine those into parts and objects: a face, number plate, crack, or product defect.

The same idea powers much of modern computer vision, speech recognition, text generation, translation, and large language models. A transformer is also a deep neural network architecture.

How is it different from classic machine learning?

In classic machine learning, people usually decide which features the model sees. If you predict late payments, you may create features such as payment history, open balance, customer age, and sector. This is feature engineering.

Deep learning can learn useful features from rawer data. A vision model receives pixels and learns edges, textures, shapes, and objects. A speech model receives audio and learns acoustic patterns. A language model receives tokens and learns statistical patterns in text.

That difference changes the trade-offs.

Data
Deep learning usually needs much more training data than simpler models, especially when trained from scratch.

Compute
Deep networks rely heavily on matrix operations, so GPUs and specialised accelerators matter.

Explainability
A deep model with millions or billions of parameters is harder to explain than a small tree model or linear model.

Data type
Deep learning shines on unstructured data: images, audio, video, and text. For structured business tables, simpler models often remain hard to beat.

The 2012 breakthrough

Deep neural networks were not invented in 2012, but 2012 made them impossible to ignore.

Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton trained a deep convolutional neural network on ImageNet, a dataset with more than a million labelled images across 1,000 classes. Their model, later known as AlexNet, used GPUs and had about 60 million parameters. It beat previous image-classification approaches by a large margin.

That result showed what happens when enough data, enough compute, and deep architectures meet. The same recipe, scaled up and adapted, underpins many later breakthroughs in vision, speech, and language.

Where deep learning is used

• Computer vision. Quality inspection, medical imaging, object detection, OCR, number-plate recognition, and image search.

• Speech. Speech-to-text, text-to-speech, call-centre transcription, voice commands, and meeting summaries.

• Language. Translation, chatbots, summarisation, retrieval-augmented generation, and large language models.

• Generative AI. Models that create text, images, audio, video, code, or structured data.

Most companies do not train deep networks from scratch. They use a pretrained model through an API, or fine-tune an existing model with their own examples. Transfer learning lowers the data and compute needed for many practical tasks.

When is deep learning overkill?

Many business datasets are structured tables: orders, customers, stock, invoices, visits, tickets, and payments. For those datasets, tree-based models such as gradient boosting and random forests often perform as well as or better than deep networks, especially on medium-sized tabular data.

A 2022 benchmark paper on tabular data found that tree-based models remained state of the art on many medium-sized datasets and were usually faster to train. That matches the practical experience of many business analytics teams.

So the rule of thumb is simple: use deep learning for images, speech, text, and other unstructured data, preferably through a pretrained model. For tabular business data, start with a simpler model and move to deep learning only when there is evidence it helps.

What to watch out for

Training from scratch is expensive
Large models need data, GPUs, time, and expertise. Buying access to a pretrained model is often cheaper.

More layers do not guarantee better results
Architecture, data quality, training setup, evaluation, and deployment all matter.

Explanations may be weak
If the decision must be justified to a customer, auditor, credit committee, or regulator, a simpler model may be safer.

Data rights matter
Training or fine-tuning on images, text, or customer data can raise privacy, copyright, and contractual questions.

Evaluation is still mandatory
A pretrained model can fail on your specific domain. Test it on representative examples before trusting it in a workflow.