Neuro-Symbolic AI: Teaching AI Not Just to Learn, but to Reason

How combining neural networks with symbolic reasoning could make AI more explainable, reliable, and capable of reasoning about the world.

Source: https://link.springer.com/content/pdf/10.1007/s13369-025-10887-3.pdf

Artificial Intelligence has made remarkable progress over the last decade.

Deep learning systems can recognise objects in images, translate languages, generate text, detect diseases, forecast demand, and even write code.

But there is an important question: Can AI learn from data and also reason using knowledge, rules, relationships and logic?

This is where Neuro-Symbolic AI, often written as NeSy AI, becomes interesting

Neuro-Symbolic AI combines the learning ability of neural networks with the reasoning ability of symbolic AI.

This article gives you an idea from the ground up.

First, forget the complicated terminology

Let’s start with a simple example. Imagine showing a neural network thousands of images of cats and dogs.

After training, you give it a new image:

🐱 Image → “cat”

The neural network has learned a statistical pattern from the examples. It has discovered things such as

• shapes
• textures
• edges
• colours
• patterns
• spatial relationships

But now imagine asking:

“If every cat is an animal, and this image contains a cat, what can we conclude?”

A traditional neural network is naturally designed to perform this type of explicit logical reasoning.

We could tell an AI system:

Cat → Animal

and then

This image contains a Cat

Therefore:

This image contains an Animal

That is symbolic reasoning.

So we have two different capabilities:

Neural AI
    ↓
Learn patterns from data

Symbolic AI
    ↓
Reasoning using knowledge and rules

Neuro-Symbolic AI tries to bring these two worlds together.

What is Neuro-Symbolic AI?

Neural networks learn what the data is. Symbolic systems reason about what those things mean and how they are related.

A simplified architecture looks like this:

                 DATA
                  │
                  ↓
          ┌───────────────┐
          │ Neural Network│
          └───────────────┘
                  │
                  ↓
        Learned representations
                  │
                  ↓
          Concepts / Objects
                  │
                  ↓
        ┌──────────────────┐
        │ Symbolic Knowledge│
        │      + Logic      │
        └──────────────────┘
                  │
                  ↓
              Reasoning
                  │
                  ↓
              Decision

The neural component provides perception and learning. The symbolic component provides knowledge and reasoning.

The two worlds of AI

To properly understand Neuro-Symbolic AI, we first need to understand the two approaches it combines.

Neural AI: Most modern AI systems are based on neural networks.

Examples:

• CNNs
• Vision Transformers
• Transformers
• LSTMs
• diffusion models
• large language models

These models learn from examples.

Suppose we want to detect whether an image contains cancer.

We could provide

Image 1 --> Cancer
Image 2 --> Normal
Image 3 --> Cancer
Image 4 --> Normal
....

The model learns a function:

f(x)→y

where:

• x = input image
• f = learned model
• y = prediction

The model learns from data rather than us explicitly writing every rule. This is incredibly powerful.

The strength of neural networks

Neural networks are excellent at dealing with unstructured and noisy data.

For example

Image

Pixels
 ↓
Neural network
 ↓
Tumour probability = 0.91

Speech

Audio waveform
 ↓
Neural network
 ↓
"Hello, how are you?"

Sensor data

GHI
DHI
BNI
Voltage
 ↓
Temporal neural network
 ↓
Anomaly probability = 0.87

This is something traditional rule-based systems struggle with. We don’t want to manually write rules describing every possible pixel pattern or every possible voltage trajectory. The neural network can learn patterns from data.

But neural networks have limitations

Imagine a neural network detecting a tumour. It tells us:

Cancer probability = 94%

We might ask:

Why?

A neural network can provide explanations such as:

“These pixels contributed strongly to my prediction.”

But this is not necessarily the same as logical reasoning.

Get Gaurav Yadav’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Another problem is that neural networks can sometimes make predictions that contradicted know knowledge.

For example:

Neural network:

Tumour detected outside the liver
confidence = 96%

But suppose our domain knowledge tells us that the particular segmentation is anatomically impossible. A purely neural system may not automatically know that.

This is where symbolic reasoning can help.

What is Symbolic AI?

Symbolic AI represents knowledge explicitly. Instead of learning everything from raw data, we provide the system with concepts, relationships, facts, and rules.

For example:

Cat → Animal
Dog → Animal

We can represent relationships:

John → works_at → University
University → located_in → UK

Or rules:

If something is a Cat
AND
all Cats are Animals

THEN
that thing is an Animal.

This is called symbolic reasoning

A simple example of symbolic reasoning

Suppose we know:

Fact 1:
Socrates is a human.

Fact 2:
All humans are mortal.

The symbolic system can infer:

Socrates is mortal

The important point is that this conclusion wasn’t necessarily given directly as training data.

It was derived through reasoning.

This is one of the major differences between

Learning

and

Reasoning

Neural AI Vs Symbolic AI

A simple comparison:

| Neural AI                    | Symbolic AI                                    |
| ---------------------------- | ------------------------------------------ |
| Learns from data             | Uses explicit knowledge                    |
| -----------------------------|------------------------------------------- |
| Excellent with images        | Excellent with rules                       |
| -----------------------------|--------------------------------------------|
| Handles noisy data           | Handles logical relationships              |
|------------------------------|--------------------------------------------|                |
| Learns representations       | Represents concepts explicitly             |   
|------------------------------|--------------------------------------------|
| Statistical                  | Logical                                    |
|------------------------------|--------------------------------------------|                                    
| Often difficult to interpret | Usually easier to trace                    |
|------------------------------|--------------------------------------------|
| Can generalise from examples | Can reason from known rules                |
|------------------------------|--------------------------------------------|
| Often data-hungry            | Can work with limited data if knowledge    |
|                              | exists                                     |
|------------------------------|--------------------------------------------|

Neither approach is perfect. And that’s exactly why combining them is attractive.

Welcome to Neuro-Symbolic AI

Now imagine:

              REAL WORLD
                  │
                  ↓
               Raw Data
                  │
                  ↓
          ┌───────────────┐
          │ Neural Network│
          └───────────────┘
                  │
                  ↓
       Objects / concepts / states
                  │
                  ↓
          ┌─────────────────┐
          │Knowledge + Logic│
          └─────────────────┘
                  │
                  ↓
              Reasoning
                  │
                  ↓
              Decision

This is the basic idea of Neuro-Symbolic AI.

The neural network says:

What do I see?

The symbolic system says:

What does it mean, and what follows from it?

A real-world example: autonomous driving

Consider a self-driving car.

A camera sees:

Car
Pedestrian
Traffic light
Road

The neural network is responsible for perception:

Image
 ↓
Object detection
 ↓
Car
Pedestrian
Traffic light

But perception alone is not enough.

The system needs to reason:

Pedestrian detected
+
Pedestrian is crossing the road
+
Traffic light is red
        ↓
Slow down / stop

The neural network provides the observations.

The symbolic reasoning layer can encode relationships and constraints.

Together:

Seeing + Reasoning

Becomes more powerful than either alone.

Another example: medical AI

Suppose a CT scan contains a liver lesion.

A neural network might produce:

Lesion probability = 0.94

But a Neuro-Symbolic system could represent:

Lesion A
    ↓
Located in liver

Lesion A
    ↓
Irregular morphology

Lesion A
    ↓
Multiple similar lesions present

Then medical knowledge can provide relationships such as:

Liver lesion
+
certain imaging characteristics
+
clinical context
        ↓
increased suspicion of metastasis

The final decision, therefore, combines:

Visual evidence
+
Medical knowledge
+
Reasoning

This is one reason Neuro-Symbolic AI is particularly interesting for high-stakes AI applications.

I hope you got the basic understanding of the Neural-Symbolic AI. In the next article, we will discuss in detail the mathematical part of the neuro-symbolic AI. Thank you for reading this article.

################### End of the part 1###########################

Resources

1. https://arxiv.org/pdf/2303.04660
2. https://arxiv.org/pdf/2501.05435
3. https://link.springer.com/article/10.1007/s13369-025-10887-3