Deep learning

A machine learning implementation technique that uses large quantities of data, or feedback from interactions with a simulation or an environment, as training sets for a network with multiple hidden layers, called a deep neural network, often employing an iterative optimization technique called gradient descent, to tune large numbers of parameters that describe weights given to connections among units.

Source: AIEOG AI Lexicon (Feb 2026), adapted from FSB P011117

Deep learning is a specialized subset of machine learning that uses artificial neural networks with many layers to learn complex patterns from large amounts of data. The “deep” in deep learning refers to the multiple hidden layers in the neural network, each of which learns increasingly abstract representations of the input data.

A standard machine learning model might learn straightforward relationships between inputs and outputs. A deep learning model can learn hierarchical, multi-level patterns: the first layers might detect basic features, middle layers combine those into more complex patterns, and later layers use those patterns to make predictions or decisions.

Deep learning powers many of the most prominent AI applications today: natural language processing, computer vision, speech recognition, and generative AI. These applications require the ability to process complex, unstructured data (text, images, audio) where deep learning excels.

However, deep learning’s strength is also its governance challenge. Models with millions or billions of parameters are inherently difficult to interpret. The decisions they make are driven by complex interactions among parameters that do not map neatly to human-understandable explanations.

Deep learning is entering financial services through multiple channels: fraud detection, natural language processing for document analysis, customer service chatbots, credit risk modeling, and market prediction. Its adoption brings both capability improvements and governance challenges.

• Performance advantages. Deep learning models can process complex, unstructured data and identify subtle patterns that simpler models miss. This can improve fraud detection accuracy, document processing speed, and customer experience.
• Explainability challenges. Deep learning models are often black boxes. In regulated contexts where decisions must be explained (lending, BSA/AML, insurance), this creates compliance friction.
• Data requirements. Deep learning requires significantly more training data than simpler techniques. Institutions must ensure they have sufficient, high-quality, representative data.
• Computational costs. Training and running deep learning models requires substantial computing resources, often provided by cloud-based AIaaS platforms, introducing third-party dependency.

1. Assess explainability needs. Before deploying deep learning, determine whether the use case requires explainable decisions. If it does, evaluate whether post-hoc explainability techniques are sufficient.
2. Apply model risk management. Deep learning models are subject to the same governance requirements as any other model. Their complexity may warrant additional validation rigor.
3. Validate with rigor. The complexity of deep learning models requires thorough validation, including performance testing across segments, bias assessment, and stress testing.
4. Monitor for drift. Deep learning models can be particularly susceptible to drift because they learn complex patterns that may be sensitive to data distribution changes.
5. Document design decisions. Maintain documentation of why deep learning was chosen over simpler alternatives, what architecture was selected, and how the model was trained.
6. Assess data sufficiency. Evaluate whether the available training data is large enough and representative enough to support the deep learning approach.

Machine learning, AI model, Foundation models, Black box, Training data, Supervised learning, Unsupervised learning