Symbolic AI and deep learning: two ways to model data
Both learn from data, and both can predict. They differ in what they hand back. Deep learning returns a neural network: a fixed architecture with thousands to billions of trained numbers, which you can call and measure. Symbolic AI returns explicit structure, such as a formula or a short program, which you can also read. Where deep learning fits opaque parameters, Symbolic AI looks for the underlying structure.
Neither is the better method in general. They answer different needs, and the useful question is which need you have.
What each produces
A neural network is chosen as an architecture first: layers, units, connections. Training then adjusts its numbers until the predictions match the examples. The knowledge it gains sits in those numbers together, and no single one of them means anything to a reader.
A symbolic model is searched for rather than trained. With Symbolic Regression, the method Diafunc is built on, candidate formulas compete on how well they predict and how small they are, and the form of the model is part of what is discovered. The result is the formula itself. This one was discovered from 308 measurements of yacht hulls:
A discovered model
residuary_resistance = 93.8 × tanh(-1.65 × prismatic_coefficient
+ 0.01 × longitudinal_position_of_buoyancy
+ 10.73 × froude_number
− 0.027 × beam_draught_ratio
− 4.25) + 94.3 You can see which inputs it uses and which it leaves out, which one carries most of the weight, and in which direction each pushes the result. It comes from a public analysis, with its quality measured on rows the search never saw.
How they differ
| Deep learning | Symbolic AI | |
|---|---|---|
| What you get | A trained network: an architecture and its numbers | An explicit model: a formula or a short program |
| How you check it | By measuring its predictions, and with explanation methods applied afterwards | By measuring its predictions, and by reading it |
| What an explanation is | A second analysis laid over the model | The model |
| Where it is at home | Images, sound, language, and other data without named columns, in large amounts | Tables: named columns and a target, from about a hundred rows per target |
| Outside the data it saw | It returns a number, and little tells you whether to believe it | It returns a number, and the formula shows how it got there, so you can judge it against what you know |
| Moving it elsewhere | The network and a runtime for it | A few lines of code in the language you need |
When deep learning is the right tool
When the input is an image, a recording or free text, a neural network is the established choice, and nothing on this page argues otherwise. The same holds when prediction accuracy is all that counts and nobody will ever ask why: a model you cannot read can be the right tool, and it may predict better than a model small enough to read.
When a readable model is the right tool
When the data is a table and someone will ask what the answer rests on. That someone may be a colleague who has to act on the prediction, a customer, or you in a year. A readable model lets you check it against what you know about your subject, notice a term that makes no sense, and say where it should not be used.
Readable has a price, and an honest report states it. An automated analysis on Diafunc scores the discovered model against simple baselines on the same unseen rows, so you see what its structure adds over a straight line. It also lists its concerns about its own model, the weak regions included.
Using both
The two are not rivals inside one project. A network can read the images and turn them into measurements; a readable model can then relate the measurements to the outcome. A readable model can also serve as the check on a larger one: if a short formula explains nearly as much, you have learned what the larger model mostly does.
Where Diafunc stands
Diafunc is a Symbolic AI Platform. You upload a table and choose what to explain; an automated analysis discovers one readable model per target and writes the report around it. The technology underneath is Ufinq. It is benchmark-tested: evaluated on SRBench, the standard Symbolic Regression benchmark, with the results public at ufinq.com/benchmarks.