Try to write a perfect rule for recognizing a cat.
You might begin:
A cat has four legs.
Then you meet a three-legged cat.
A cat has fur.
Then you remember hairless cats.
A cat has pointed ears.
Not always.
Very quickly, the clean symbolic definition becomes a pile of exceptions.
Now try a different problem:
If $A>B$ and $B>C$, what follows?
Here the answer is exact:
\[ A>C. \]
No training set is needed.
These two examples capture an old tension in artificial intelligence.
Some knowledge is flexible, statistical, and hard to express as rules.
Other knowledge is crisp, relational, and naturally symbolic.
Science needs both.
Prefer to watch? Continue with the video below, or keep reading.
The Neural Approach: Learn from Examples
Modern AI has been dominated by neural networks.
Instead of writing explicit rules, we provide data.
The system learns patterns.
This is extraordinarily powerful when the world is messy.
Images are difficult to describe through hand-written rules.
Natural language contains ambiguity, context, exceptions, and subtle statistical structure.
Scientific data can be equally complex.
Spectra, microscopy images, molecular structures, and experimental signals often have patterns that are easier to learn than to specify explicitly.
Neural models thrive in these settings.
They can discover representations that humans did not know how to hand-code.
The Symbolic Approach: Make the Structure Explicit
Symbolic AI begins from a different philosophy.
Knowledge is represented through objects, rules, variables, logical relationships, or equations.
For example:
\[ F=ma. \]
Or:
\[ A\Rightarrow B, \qquad B\Rightarrow C \quad\Rightarrow\quad A\Rightarrow C. \]
The advantage is clarity.
A symbolic rule can often be inspected, proved, manipulated, and reused.
It supports exact reasoning.
This is especially attractive in mathematics and theoretical science.
But symbolic systems struggle when the rules are incomplete or when the concepts themselves are fuzzy.
Real laboratories are full of such situations.
Why Pure Neural Science Can Be Unsatisfying
Suppose a neural network predicts a molecular property with extraordinary accuracy.
A scientist asks:
Why?
The model returns a complex internal activation pattern.
That may not help.
If the model is used only as a screening tool, perhaps explanation is unnecessary.
But if we want a scientific theory, we often need a structure we can reason about.
We want to know which variables matter.
Which relationship is causal.
Which symmetry is present.
Which assumptions produce the result.
A purely statistical predictor may not provide those answers naturally.
Why Pure Symbolic Science Is Also Limited
Now imagine trying to build a complete symbolic model of a living cell.
We would need rules for an enormous number of interactions.
Many are uncertain.
Some depend on context.
Others are only approximately known.
The symbolic system becomes brittle or impossibly large.
The same issue appears in robotics and experimental science.
The real world does not present itself as a clean theorem.
Sensors are noisy.
Objects vary.
Definitions have edge cases.
That is why learned representations are so useful.
The Neuro-Symbolic Bet
Neuro-symbolic AI tries to combine the strengths of both approaches.
The neural component handles perception, uncertainty, and flexible pattern recognition.
The symbolic component handles explicit relations, constraints, and formal reasoning.
A simplified architecture might look like
\[ \text{raw data} \rightarrow \text{neural representation} \rightarrow \text{symbolic reasoning} \rightarrow \text{prediction or action}. \]
Or the flow may go in the opposite direction, with symbolic knowledge constraining a neural model.
For scientific AI, both directions are interesting.
A Chemistry Example
Imagine an AI analyzing molecular data.
A neural network may be excellent at learning which local structures are associated with a property.
But chemistry also contains explicit knowledge:
- conservation laws,
- allowed valences,
- symmetries,
- known reaction constraints,
- thermodynamic relationships.
Why force the neural model to rediscover every rule from data?
We could instead provide some constraints explicitly while allowing the model to learn what is difficult to encode.
This can reduce the search space and improve reliability.
Two Bridges Already in Use: PINNs and Symbolic Regression
Researchers are already building practical bridges between learned representations and explicit scientific structure.
Physics-Informed Neural Networks (PINNs) place known differential equations or other physical constraints directly into the training objective. A network fitting fluid-flow data, for example, can be penalized when its prediction violates the governing equations or conservation laws. The neural model still learns from data, but the physics constrains what counts as an acceptable solution.
Symbolic regression approaches the problem from the other direction. Instead of stopping with a black-box predictor, it searches for compact mathematical expressions that explain the observed data. Modern systems, including tools such as PySR, can use large searches over candidate expressions to produce equations that a scientist can inspect, test, and compare with known theory.
These approaches are not complete solutions to scientific reasoning. But they show two useful directions: make learning obey known laws, and make learned patterns return to human-readable symbolic form.
Mathematics Shows the Pattern Clearly
AI mathematics provides perhaps the cleanest example.
A language model may propose a proof in ordinary mathematical language.
But ordinary language is ambiguous.
A formal proof assistant can check each logical step under precise rules.
So one possible loop is
\[ \text{neural proposal} \rightarrow \text{symbolic formalization} \rightarrow \text{formal verification} \rightarrow \text{revision}. \]
The neural system provides breadth and creativity.
The symbolic system provides rigor.
Neither has to do everything.
This is a powerful design principle for AI4S.
Scientific Laws Are Symbolic Compression
There is another reason symbolic reasoning matters.
Science often aims to replace a large body of observations with a compact relationship.
Kepler's laws summarize planetary motion.
Maxwell's equations summarize a vast range of electromagnetic phenomena.
A learned predictor may reproduce the same outputs, but a symbolic law offers a different kind of compression.
It exposes structure.
It can be manipulated.
It can be combined with other laws.
It supports counterfactual reasoning.
This makes symbolic discovery an important goal even in an age of large neural models.
The Difficulty Is the Interface
Combining neural and symbolic systems sounds obvious.
Doing it well is difficult.
How does a continuous neural representation become a discrete symbol?
How do symbolic constraints influence learning without making the model too rigid?
What happens when the symbolic knowledge is wrong or incomplete?
How should uncertainty be represented in a logical system?
These are active research problems.
The interface between the two worlds may be more difficult than either side alone.
The Scientific Loop Needs Both Flexibility and Rigor
Return to our autonomous laboratory.
The AI observes a noisy experiment.
Neural methods help interpret the sensor data.
The system proposes a mechanism.
Symbolic relationships help express the mechanism clearly.
A simulator generates predictions.
A formal constraint checks whether the proposal violates known conservation laws.
The robot runs the next experiment.
The result updates the learned model.
This is not purely neural science.
It is not purely symbolic science.
It is a hybrid research process.
Could Symbols Help AI Invent Concepts?
There is also a deeper possibility.
In the previous blog, we asked whether AI can invent new scientific concepts.
A neural system may discover useful latent structure.
A symbolic layer may help turn that structure into something explicit.
For example, the model may notice that a certain combination of measured quantities repeatedly behaves as one unit.
The system then proposes a new variable:
\[ z = f(x_1,x_2,x_3). \]
If $z$ simplifies predictions across many experiments, it may become a candidate scientific concept.
The symbolic representation allows the abstraction to become portable.
It can be tested, communicated, and incorporated into other theories.
We May Be Rebuilding the Scientific Method in Software
Human science already combines both modes.
Scientists use intuition and pattern recognition.
They also use equations and formal reasoning.
They recognize a shape in a graph before proving why it appears.
They make an analogy before writing the mathematical model.
They move constantly between flexible and explicit representations.
A mature AI scientist may need to do the same.
The future may therefore be less about choosing between neural and symbolic AI than about deciding where each belongs in the scientific loop.
And if these systems become extremely productive, another problem will appear.
What happens when AI can generate more proofs, explanations, and candidate discoveries than humans can possibly read?
Next in the series: When Answers Become Abundant: What Is Left for Scientists?
0 comments