Chapter 22 of 25
Real-World Applications
Where MLPs actually show up in production
Motivation
Every idea covered so far — weighted sums, activations, backpropagation, initialization, regularization, the full training pipeline — exists to serve one final purpose: making predictions that matter, on problems that affect real people, in production systems running right now. This chapter is the payoff. Rather than one motivating scenario, this chapter is the scenario collection: a tour of where the humble multi-layer perceptron actually shows up across industries, either doing the whole job itself or quietly powering a critical piece of something much larger.
Building the Intuition
Two patterns repeat across almost every domain below:
- Standalone tabular classifier/regressor: the MLP takes structured, spreadsheet-like features directly — age, income, sensor readings, lab values — and outputs a prediction. This is the "classic" use case, and it's still extremely common.
- A component inside something bigger: the MLP takes already-extracted features (from a convolutional network, a Transformer, an embedding table) and turns them into a final prediction. In this role, the MLP is often called a "head," and it shows up inside nearly every modern deep learning architecture in some form.
Keeping these two roles in mind makes the following tour much easier to organize.
Healthcare
MLPs are widely used on structured, tabular clinical data: predicting patient risk scores from electronic health records (age, lab values, vital signs), estimating disease progression, and serving as the final classification layers in medical imaging pipelines, where a convolutional network extracts image features and an MLP head converts those features into a diagnosis.
Given tabular features — age, BMI, blood glucose, blood pressure, family history — an MLP with a sigmoid output can be trained with binary cross-entropy loss to output the probability that a patient will develop diabetes within a given time horizon, supporting clinical decision-making rather than replacing it.
Finance
- Credit scoring: predicting the probability of loan default from an applicant's financial history — a binary classification problem.
- Fraud detection: flagging anomalous transactions, typically framed as binary classification on highly imbalanced data — this is exactly where the earlier warning about accuracy being misleading on imbalanced datasets matters most.
- Algorithmic trading signals: predicting short-term price movement direction or magnitude — classification or regression — from engineered market features.
Cybersecurity
MLPs power network intrusion detection systems, trained on features extracted from network traffic — packet sizes, connection duration, protocol flags — to classify traffic as normal or malicious. Because attack patterns evolve continuously, such models are typically retrained regularly on updated traffic data rather than deployed once and left alone.
Recommendation Systems
While large-scale recommendation systems often use specialized architectures, MLPs are commonly used as one component: taking concatenated user and item embedding vectors as input and outputting a predicted rating or click-through probability, a regression or binary classification task respectively. This general pattern — an MLP applied on top of learned or engineered feature vectors — is one of the most common practical roles the MLP plays in modern industrial machine learning systems.
Manufacturing: Predictive Maintenance
Given sensor readings from industrial equipment over time — vibration, temperature, pressure — an MLP can be trained to predict the probability of failure within a given time window, allowing maintenance to be scheduled proactively rather than reactively. This reduces both unplanned downtime and unnecessary preventive maintenance performed on equipment that didn't actually need it.
Robotics
MLPs serve as function approximators inside robotic control systems: approximating a mapping from sensor readings to motor control signals, or serving as the value-function or policy-function approximator inside reinforcement learning algorithms.
Natural Language Processing
It's easy to overlook, but every Transformer-based language model — powering modern chatbots and translation systems — contains, within each of its layers, a position-wise feedforward network that is, structurally, exactly the multi-layer perceptron studied throughout this course: a linear layer, a nonlinear activation (typically GELU), and another linear layer. Understanding the MLP thoroughly is therefore directly relevant to understanding even the most advanced modern architectures, not just legacy tabular-data models.
Computer Vision
While convolutional neural networks dominate raw image processing, MLPs appear as the final classification "head" of most vision architectures, mapping extracted feature vectors to class probabilities via a softmax output layer.
Time Series and Tabular Data
For tabular data specifically — spreadsheet-like data with rows as examples and named columns as features — MLPs remain a strong, commonly used baseline, though well-tuned gradient-boosted tree methods frequently match or outperform MLPs on small-to-medium tabular datasets.
Summary Table
| Domain | Typical Task | Task Type |
|---|---|---|
| Healthcare | Disease risk prediction | Binary classification |
| Finance | Credit scoring, fraud detection | Binary classification |
| Cybersecurity | Intrusion detection | Binary/multi-class classification |
| Recommendation | Click-through / rating prediction | Classification / regression |
| Manufacturing | Predictive maintenance | Binary classification |
| Robotics | Control signal approximation | Regression |
| NLP | Feedforward sublayer in Transformers | Component within larger model |
| Computer Vision | Classification head after feature extraction | Multi-class classification |
Key Takeaways
- MLPs are widely used directly on tabular/structured data across healthcare, finance, cybersecurity, and manufacturing. - MLPs frequently serve as a component — particularly a classification or regression "head" — within larger architectures built on CNNs or Transformers. - The position-wise feedforward sublayer inside every Transformer is structurally an MLP, making this material directly relevant to understanding modern large language models. - Recognizing which of the two roles — standalone model or component — a problem calls for is a useful first step when designing a real system.
Common Mistakes
MLPs are far from obsolete — they remain the default choice for tabular data, and they are literally embedded inside every Transformer layer as the feedforward sublayer. "Not the architecture on the front page" is very different from "not used."
Several domains above — fraud detection, intrusion detection, predictive maintenance, rare disease prediction — are naturally imbalanced. Applying plain accuracy as the primary evaluation metric in any of these risks deploying a model that looks excellent on paper while catching almost none of the cases that actually matter.
Several of these applications — fraud patterns, network intrusion signatures, even market dynamics — shift over time. A model trained once and never revisited quietly degrades; production systems in these domains need a retraining cadence built in from the start.
Interview Corner
Where This Shows Up in Practice
Look closely at almost any production machine learning system — a bank's fraud pipeline, a hospital's risk dashboard, a recommendation engine, or the language model answering a question right now — and somewhere inside it, doing unglamorous but essential work, is a multi-layer perceptron: a stack of weighted sums, activations, and biases, trained by the exact backpropagation algorithm and pipeline covered throughout this course. That's the real takeaway of this entire course: the MLP isn't a historical stepping stone that got replaced. It's the working core that modern architectures are built around.