Deep Learning
Deep learning is machine learning with neural networks of many layers, which learn useful representations of raw data — pixels, audio, text, sequences of events — layer by layer, instead of relying on features designed by hand; it is the technology behind image recognition, speech, the transformers and large language models, and generative AI. It needs large data and compute and wins wherever the input is unstructured; on the tabular data of most financial decisions, gradient-boosted trees still usually win.
- Also known as
- Deep neural networks, DL, Neural networks, Representation learning
- Used at
- computer vision (document capture, identity verification, deepfake detection), speech (voice assistants, call transcription), language (every large language model, document understanding, translation), sequences (transaction-sequence models for fraud, time series), recommendation, and the generative models for text, images, audio and code
- Binding
- none for the method; the systems built with it fall under the EU AI Act by use, model-risk rules by function, and the explainability expectations that deep models meet less easily than trees
- Typical history
- the perceptron (1958) and backpropagation (1986) predate it; the turning points were AlexNet's ImageNet win on GPUs (2012), sequence models for speech and translation (2014–2016), the transformer (2017), BERT and GPT (2018–2020), and the scaling to hundreds of billions of parameters (2020 onward); each turn was more data, more compute and an architecture that used them
- Typical requirements
- labelled or self-supervised data at scale (thousands of images for a fine-tuned classifier, trillions of tokens for a language model), GPU or accelerator compute for training, and engineering for serving; the costs are why most companies fine-tune or rent rather than train from scratch
- Parties
- the research groups and labs, the hardware makers whose GPUs made it feasible, the framework developers (PyTorch, TensorFlow, JAX), and the practitioners who apply it — in fintech, mostly through vendors and rented models, with in-house deep learning for sequence and document tasks at the larger firms
How deep learning works in practice
A document-capture model takes an image of an ID; its first layers respond to edges and textures, its middle layers to shapes and text regions, its later layers to the layout of a specific document type, and its final layer outputs the type and the field locations — none of it designed, all of it learned from labelled images by backpropagation. A transaction-sequence model reads a customer's last 500 events through attention layers and outputs a fraud probability, having learned that a password reset followed by a new payee followed by a large transfer is the pattern. A language model, the largest case, learned from trillions of tokens to predict the next word and, in doing so, learned grammar, facts, reasoning and code. In each case the representations are the model's own, which is the source of the accuracy and of the difficulty of explaining it.
The practical map for a fintech is clear by now. Deep learning is the right tool for unstructured inputs — documents, faces, voices, text, long event sequences — and is usually rented (a vision or language API) or fine-tuned from a pre-trained model rather than trained from scratch. For tabular decisions — credit, most fraud scoring on engineered features — gradient-boosted trees are as accurate or better, faster, cheaper and easier to explain, and they remain the default under model-risk scrutiny. The larger fintechs build deep models where their data is unique (transaction sequences, their own documents); the rest apply them through vendors and prompts. For a buyer, a target's deep-learning claims are read for what was trained on what data, and a fine-tuned open model is a capability, not a moat.
For a fintech, deep learning is the technology behind its perception — what it reads, sees and hears — and rarely behind its tabular decisions; for an acquirer, the question is which models the target trained on data nobody else has.
Example
A fintech's identity-verification flow runs on a vendor's deep-learning models: document classification, field extraction, face match and liveness. Its fraud team wants a model on its own transaction sequences, where it has 2 billion events and confirmed outcomes no vendor has. It builds a small transformer over event sequences — 12 million parameters, trained on one GPU for a day — and runs it as a second scorer on the 8% of transactions the gradient-boosted model marks uncertain. Detection on that segment rises from 61% to 79%; the attention weights show investigators which events mattered; the model card documents the training, the out-of-time validation and the explanation method. The company kept its tabular model for the bulk and built deep learning where its data was the asset — the pattern its acquirer's technology diligence, two years later, singles out as the reason the fraud stack is worth paying for.
Common mistakes
- Deep learning on tabular data where trees are better and explainable. - Training from scratch when fine-tuning or renting would do. - Underestimating the data a deep model needs, and overfitting a small set. - Deploying a deep model on a regulated decision without an explanation method the regulator accepts. - Reading a fine-tuned open model as proprietary technology.