---
title: "Hidden Layer"
url: "https://mahouse.world/glossary/hidden-layer"
type: "glossary term"
topic: "Model architecture"
author: "Ihor Vlasov"
reviewed: "2026-09-20"
published: "2026-08-10"
---

# Hidden Layer

A hidden layer is any layer of a neural network between the input layer, which receives the data, and the output layer, which produces the prediction; it is where the network forms the intermediate representations it computes with — edges and shapes in a vision model, syntax in a language model, risk patterns in a fraud model. "Deep" learning means many hidden layers. No one specifies what they represent; training decides, the source of their power and their opacity.

**Also known as:** Hidden layers, Intermediate layer, Deep layers

## Key facts

- **Used at:** every neural network beyond a single-layer perceptron; a fraud model may have three hidden layers, a large language model around a hundred transformer blocks, each of which is several layers
- **Binding:** none directly; the explainability obligations for high-risk AI in the EU AI Act and model-risk guidance are largely about the fact that hidden layers do not come with meanings attached
- **Typical mechanics:** each hidden layer takes the previous layer's outputs, applies weights and a bias, and a non-linear activation function; without the non-linearity, any stack of layers would collapse to a single linear transformation and could not represent anything a line cannot; the number of layers (depth) and units per layer (width) are hyperparameters
- **Typical findings:** early layers learn general features (edges, word parts), later layers task-specific ones (objects, sentiment), which is why transfer learning reuses early layers; individual units sometimes correspond to human concepts and mostly do not; interpretability research reads hidden activations to find what a model represents
- **Parties:** the researchers who study what hidden layers learn, the engineers who size them, and the risk teams who need a decision explained when the representation that produced it has no name

## How hidden layers work in practice

A credit-risk network takes 300 features — income, history, behaviour — into a first hidden layer of 256 units. Each unit computes a weighted sum of all 300 inputs and passes it through ReLU, which zeroes negatives; after training, one unit may respond to "high utilisation with recent credit searches", another to "stable income with long tenure" — combinations no analyst wrote. The second hidden layer combines those into higher-order patterns; the output layer turns them into a probability of default. The network's accuracy comes from these learned combinations; its opacity comes from the same place — the 256 units have no labels, and the one that matters for a given decline is a direction in a 256-dimensional space.

That opacity is why regulated decisions handle hidden layers carefully. Tools exist to attribute an output to inputs (SHAP, integrated gradients), which give reasons in terms of the features the applicant recognises; interpretability research probes the hidden representations directly, with some success on language models; and many lenders keep the decision model to gradient-boosted trees, whose structure can be read, using neural networks upstream for feature extraction. Depth also brings the engineering questions: deeper networks learn richer representations but train harder (vanishing gradients, solved by skip connections and normalisation) and overfit small datasets; the fraud model with three hidden layers is the right size for its data, and the hundred-block language model is the right size for the internet.

For a fintech the hidden layer is where the model's value and its governance problem sit together: what it learned is why it works, and what it learned has to be surfaced in terms a regulator accepts.

## Example

A lender's model-risk team reviews a neural-network credit model with two hidden layers of 128 units, proposed to replace a logistic regression. Accuracy is better (Gini 0.62 versus 0.55). The team's conditions: SHAP-based reason codes for every decline, validated against the regression's on 10,000 cases (agreement on the top reason 84%); a probe of the hidden layers for proxies of protected characteristics, which finds one unit correlated 0.4 with postcode-derived ethnicity data and leads to the removal of two input features and retraining; and a monitoring rule that re-runs the probe quarterly. The model goes live four months after the first proposal; the regression stays as a challenger for a year.

## Common mistakes

- Adding layers to a model whose data cannot support them; more depth on a small dataset overfits.
- Deploying a deep model for regulated decisions without a reason-code method the regulator accepts.
- Assuming hidden units correspond to concepts; most are mixtures.
- Ignoring proxies: a hidden layer can reconstruct a protected characteristic from allowed inputs.
- Treating depth as a virtue in itself; the right depth is an empirical question.

## Related terms

- [Deep Learning](https://mahouse.world/glossary/deep-learning.md)
- [Forward Propagation](https://mahouse.world/glossary/forward-propagation.md)
- [Latent Space](https://mahouse.world/glossary/latent-space.md)
- [Model Parameters](https://mahouse.world/glossary/model-parameters.md)
- [Explainable AI](https://mahouse.world/glossary/explainable-ai.md)

— Ihor Vlasov, M&A FinTech Expert. Last reviewed 2026-09-20. https://mahouse.world/authors/ihor-vlasov