If you want AI to predict next week's stock market or analyze a complex Excel spreadsheet, do you choose a powerful neural network or a gradient boosting tree (GBDT) like XGBoost?
In the data science community, this has long been a problem of 'one-sided strengths': neural networks excel at unstructured data like images and speech, but on structured data in spreadsheets they are often beaten by simpler decision trees. A new study, MANN (2026), proposes an elegant idea: since each paradigm has its specialty, why not fuse them into an all-rounder?
---
#### 1. Neural Networks' 'Aristocratic Problem': Why Do They Struggle with Excel?
Neural networks (NN) are like aristocrats raised in art school—they have remarkable intuition for pixels and sound waves. But when facing discrete, seemingly irregular data columns in Excel, they often underperform: they easily 'overthink' (overfit) and are extremely sensitive to hyperparameters.
Gradient boosting trees (GBDT), by contrast, are like seasoned veteran accountants—unschooled in art, but highly precise about the logical relationships among data, and highly efficient.
#### 2. MANN: Putting Off-Road Tires on a Supercar
The core innovation of MANN (Multiple Additive Neural Networks): it retains the gradient boosting framework, but swaps out all the internal 'parts' for neural networks.
- Base learner overhaul: In traditional gradient boosting, many 'decision trees' are summed together. In MANN, those trees are replaced with nearly-shallow neural network units.
- Continual learning: Decision trees have a fatal flaw—when data updates, they often must be rebuilt from scratch. But the neural units in MANN can be 'fine-tuned'. This means your model can evolve in production like a human, instead of re-taking the exam every time.
- Capsule network boost: MANN also introduces capsule networks for feature extraction, letting it understand not just numbers in a spreadsheet but also accompanying images and audio.
- Fearless tuning: It is insensitive to learning rate and iteration count. What used to take three days of hyperparameter tuning now converges on its own.
- Automatic braking: A built-in early stopping mechanism automatically prevents the model from overfitting.
#### 3. Benchmarks: More Accurate, and More 'Stable'
Experimental results show that across multiple datasets, MANN not only beats the long-reigning champion XGBoost in accuracy—more importantly, it offers robustness:
#### Editor's Take:
The significance of MANN is that it breaks down the 'silo mentality' between algorithms.
It tells us: the most powerful intelligence often comes not from inventing an entirely new, complex structure, but from combining the strengths of existing excellent paradigms. When neural networks don the armor of gradient boosting, they transform from a one-sided 'artist' into an all-around 'commander'.
What kind of 'one-sided' data gives you the most trouble in your business scenarios? Leave a comment and let's discuss whether MANN might be your ultimate remedy.
---
*Note: This article is based on the paper 'Multiple Additive Neural Networks', arXiv:2604.26888.*