SmallMLP

Adaptive neural networks for small nonlinear data, with calibrated conformal intervals.

SmallMLP is a library for small, nonlinear datasets ($n < 500$). It replaces manual architecture search with adaptive hyperparameters derived from dataset parameters ($n$, $d$, $K$), and produces calibrated prediction intervals through weighted conformal prediction.

• Regression: 25 / 45 wins, best mean rank (1.82) among 5 models — no tuning.
• Intervals: 19% narrower than split conformal at equal coverage (31 / 32 wins).
• Classification: best mean rank among 10 models on 18 datasets; significantly better than LogReg, KNN, SVC; statistically indistinguishable from MLP-128 and RF.

---

Why SmallMLP?

Standard MLPs need architecture search. Kernel methods need bandwidth tuning. Bayesian methods scale poorly. SmallMLP removes all three problems:

Property	Standard MLP	GP	SmallMLP
Works on small data ($n < 500$)	✗	✓	✓
No hyperparameter tuning	✗	✗	✓
Prediction intervals / sets	✗	✓	✓
Trains in seconds on $n < 500$	✓	✗	✓
Multiclass-aware width	✗	—	✓

Adaptive by construction. Layer width, dropout, and weight decay are computed from dataset parameters — not searched.

---

Installation

pip install smallmlp

Requires Python 3.10+, torch>=2.0, scikit-learn>=1.3.

---

Quick start — regression

import numpy as np
from sklearn.datasets import load_diabetes
from sklearn.model_selection import train_test_split
from smallmlp import SmallMLPRegressor

X, y = load_diabetes(return_X_y=True)

# 60 / 20 / 20 split: train / calibration / validation
X_tr, X_tmp, y_tr, y_tmp = train_test_split(X, y, test_size=0.4, random_state=0)
X_cal, X_val, y_cal, y_val = train_test_split(X_tmp, y_tmp, test_size=0.5, random_state=0)

model = SmallMLPRegressor()
model.fit(X_tr, y_tr)
model.fit_conformal(X_cal, y_cal, X_val, y_val, alpha=0.1)

y_hat = model.predict(X_val)
lo, hi = model.predict_interval_conformal(X_val, alpha=0.1)

coverage = np.mean((y_val >= lo) & (y_val <= hi))
print(f"Coverage: {coverage:.3f}  Width: {np.mean(hi - lo):.2f}")

---

Quick start — classification

import numpy as np
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from smallmlp import SmallMLPClassifier

X, y = load_iris(return_X_y=True)
X_tr, X_tmp, y_tr, y_tmp = train_test_split(X, y, test_size=0.4, stratify=y, random_state=0)
X_cal, X_val, y_cal, y_val = train_test_split(X_tmp, y_tmp, test_size=0.5, stratify=y_tmp, random_state=0)

clf = SmallMLPClassifier(class_weight=None)
clf.fit(X_tr, y_tr)
clf.fit_conformal(X_cal, y_cal, X_val, y_val, alpha=0.1)

probs = clf.predict_proba(X_val)
sets = clf.predict_set(X_val, alpha=0.1)

coverage = np.mean([y_val[i] in sets[i] for i in range(len(y_val))])
avg_size = np.mean([len(s) for s in sets])
print(f"Coverage: {coverage:.3f}  Avg set size: {avg_size:.3f}")

---

Method

SmallMLP is an adaptive MLP backbone with weighted conformal uncertainty on top.

1. Adaptive width

Layer width is computed from dataset parameters. First layer:

$$ w_1 = \max\left(2K, ; \min\left(\left\lfloor \alpha \cdot \sqrt{n \cdot K \cdot \sqrt{d}} \right\rfloor, ; 4n\right)\right) $$

Subsequent layers:

$$ w_l = \max\left(K, ; \left\lfloor w_1 \cdot \beta^{,l-1} \right\rfloor\right) $$

with $\alpha = 4.0$ and $\beta = 0.7$ chosen by empirical sweep.

Properties:

• Grows sublinearly with $n$, $K$ (as $\sqrt{\cdot}$), and $d$ (as $d^{1/4}$).
• First layer bounded below by $2K$; later layers by $K$.
• Capped at $4n$ to prevent blow-up on tiny datasets.
• Depth decay $\beta^{l-1}$ shrinks later layers.

For regression, we fall back to a fixed width $\min(2n, 128)$.

2. Adaptive dropout and weight decay

Weight decay is derived from $n$:

$$ \lambda = \frac{1}{n} $$

Dropout uses a clamped function of $d/n$:

$$ p_{\text{drop}} = \min\left(\max\left(3 \cdot \frac{d}{n}, ; 0.1\right), ; 0.5\right) $$

Raw $d/n$ systematically under-estimates required regularization on small data (sweep: optimum in $[0.1, 0.2]$, raw ratio gives $0.02$–$0.08$).

3. Training

Adam optimizer, early stopping on validation loss (20% of train). Class weights are applied only when the imbalance ratio exceeds $2{:}1$.

4. Weighted conformal intervals and sets

The backbone produces an embedding $h(x)$. Given calibration residuals $R_j = |y_j - \hat{y}(x_j)|$ (regression) or scores $s_j = 1 - \hat{p}_{y_j}(x_j)$ (classification), the interval or set for a new $x^*$ uses a weighted quantile of the calibration scores, with weights derived from the kernel:

$$ w_j(x^{\ast}) = \exp\left(-\frac{|h(x^{\ast}) - h(x_j)|^2}{2 h_{\text{cal}}^2}\right) $$

This gives:

• Finite-sample coverage guarantee under exchangeability.
• Locally adaptive widths — narrow where data is dense, wide in empty regions.
• Tuning-free calibration — $h_{\text{cal}}$ selected by grid search on held-out validation.

---

Results

Regression — point prediction

45 small regression datasets ($n < 500$), 5-fold CV, mean absolute error.

Model	Mean rank ↓	Mean MAE ↓	Wins / 45
SmallMLP	1.82	8.43	25
RF (100)	2.33	13.58	14
MLP (256, 128)	3.22	11.57	4
KNN ($k=5$)	3.67	18.89	1
MLP (100,)	3.96	19.07	1

SmallMLP beats MLP (256, 128) by 27% on MAE without any tuning.

Regression — prediction intervals

45 datasets, $\alpha = 0.1$ (target coverage ≥ 0.90), 60/20/20 split.

Method	Valid (cov ≥ 0.88)	Mean width among valid ↓
Weighted conformal	38 / 45	51.8
Split conformal	32 / 45	64.0
Heuristic $\hat{y} \pm z\sigma$	17 / 45	145.7

Among 32 datasets where both weighted and split conformal are valid, weighted conformal produces narrower intervals on 31 (mean width ratio 0.81).

Classification — full benchmark

18 small classification datasets, 5-fold CV, 10 models. Metric: mean rank (lower is better, rank 1 = best on that dataset).

Model	Mean rank ↓	Mean acc ↑	Wins / 18
SmallMLP (classic)	4.19	0.8120	3
SmallMLP (formula)	4.50	0.8117	2
MLP (128,)	4.64	0.8132	3
MLP (100,)	5.25	0.8101	1
SVC (RBF)	5.31	0.7990	2
MLP (100, 100)	5.56	0.8037	3
RF (100)	5.92	0.8035	6
LogReg	6.17	0.7920	4
KNN ($k=5$)	8.17	0.7715	2

Wilcoxon signed-rank vs SmallMLP (formula), paired across 18 datasets:

Baseline	p-value	Significant?
MLP (128,)	1.00	no
MLP (100,)	0.62	no
MLP (100,100)	0.26	no
RF (100)	0.37	no
SmallMLP classic	0.62	no
SVC (RBF)	0.047	yes (SmallMLP better)
LogReg	0.013	yes (SmallMLP better)
KNN ($k=5$)	0.002	yes (SmallMLP better)

Summary: SmallMLP achieves the best mean rank among 10 models, is statistically indistinguishable from tuned MLPs and RF, and significantly better than LogReg, KNN, and SVC.

Classification — per-K-bucket accuracy

Bucket	SmallMLP (formula)	MLP (128,)	RF (100)
Binary ($K=2$)	0.8675	0.8697	0.8521
Multiclass ($2<K\le5$)	0.8537	0.8531	0.8210
Hard multiclass ($K>5$)	0.6776	0.6798	0.7019

SmallMLP competitive on binary and moderate multiclass; RF dominates on hard multiclass.

Classification — adaptive dropout ablation

9 small classification datasets, 10-fold CV, accuracy.

Dropout mode	Mean accuracy ↑
Unclamped $\min(d/n, ; 0.5)$	0.8214
Clamped $\min(\max(3d/n, ; 0.1), ; 0.5)$	0.8269
Fixed $p = 0.2$	0.8278

Clamped dropout matches fixed $p = 0.2$ within noise; advantage concentrated on multiclass ($K \geq 6$) and high-dimensional ($d \geq 30$) datasets.

Classification — conformal prediction sets

15 small datasets (binary and multiclass), 60/20/20 split, $\alpha = 0.1$.

Dataset	$K$	Coverage	Avg set size
iris	3	1.00	1.07
wine	3	0.97	1.00
breast cancer	2	0.97	1.00
hepatitis	2	0.90	1.26
parkinsons	2	1.00	1.05
ionosphere	2	0.93	1.06
banknote	2	0.99	1.00
vowel ($K=26$)	26	0.92	2.54

Coverage ≥ 0.90 on the majority of datasets; set size close to 1 on easy problems, grows on hard multiclass ones.

---

When to use SmallMLP

Good fit:

• Small datasets ($n < 500$) with nonlinear structure.
• Scientific instruments: multi-sensor arrays, medical cohorts, chemistry.
• When uncertainty quantification matters (intervals with coverage guarantee).
• When you don't want to tune architecture or regularization.

Not a good fit:

• Large datasets ($n > 10{,}000$) — more data favors standard MLPs.
• Linear problems — logistic regression or ridge is better.
• Very high dimensions ($d > 200$) with tiny $n$ — kernel methods suffer.

---

Use cases

SmallMLP is designed for scientific instruments that produce many features per observation but few observations overall:

• Telescopes with multi-sensor arrays: 30 sensors × 100–200 observations per campaign.
• Particle detectors: 10–100 features per event, 200–1000 events per analysis.
• Medical cohorts: 10–50 clinical variables, 50–300 patients.
• Chemistry / catalysis: 10–40 descriptors, 50–200 experiments.

In all these settings, standard MLPs overfit, Gaussian Processes scale poorly, and hyperparameter tuning is impractical. SmallMLP provides adaptive regression and classification with calibrated conformal intervals — no tuning required.

---

Limitations

• $n < 500$. Fit cost grows with $n$, width, and epochs.
• Heteroscedastic residuals. On ~15% of regression datasets, weighted conformal undercovers. Attributed to strong heteroscedasticity under non-exchangeability.
• Linear problems. SmallMLP loses to logistic regression and ridge on linear or near-linear tasks.
• Raw classification accuracy. SmallMLP matches but does not beat tuned MLPs and RF. Its advantages are tuning-free operation, calibrated uncertainty, and competitive mean rank.
• Adaptive dropout is only marginally better than fixed $p = 0.2$ on average. Benefit is regime-specific.

Negative results

We document what we tried that did not work, to guide future work:

• Adaptive depth $L = f(n, d, K)$: no significant improvement over fixed $L = 2$ on $n < 500$.
• PCA-based effective dimensionality $d_{\text{eff}}$ in the width formula: no improvement over raw $d$.
• Per-class sample count $\log_2(n/K)$ and softened depth decay $\sqrt{d}/\sqrt{l}$ in the binary width formula: within noise.
• Bias initialization sweep (zeros, positive, uniform): none beat PyTorch Kaiming default. Correlation between bias negativity and accuracy was spurious.
• Activation sweep (tanh, GELU, SiLU, algebraic sigmoid, softsign): ReLU wins on 18 small classification datasets; smooth/saturating activations underperform.
• Sqrt-balanced class weights: improve over linear weights under extreme imbalance, but neither beats no weighting on accuracy.
• Reweighting / soft-median in conformal scores: does not improve coverage.
• Heteroscedastic head: degrades point predictions.

What we tried that did work

• Clamped adaptive dropout min(max(3d/n, 0.1), 0.5) — better than raw d/n.
• Sqrt-balanced class weights — safer than linear under extreme imbalance (though not better than no weights for accuracy).
• Weighted conformal — 19% narrower intervals at equal coverage.
• Multi-class-aware width — competitive with fixed-width baselines without tuning.

---

API overview

SmallMLPRegressor

SmallMLPRegressor(
    activation="relu",
    lr=1e-3,
    weight_decay=None,   # default 1/n
    max_epochs=500,
    patience=30,
    batch_size=None,     # default min(n, 32)
    val_frac=0.2,
    random_state=42,
    verbose=False,
)

Methods: fit, predict, fit_conformal, predict_interval_conformal.

SmallMLPClassifier

SmallMLPClassifier(
    activation="relu",          # 'relu' | 'tanh' | 'gelu' | 'silu' | 'algsig' | 'softsign'
    lr=1e-3,
    weight_decay=None,          # default 1/n
    max_epochs=500,
    patience=30,
    batch_size=None,
    val_frac=0.2,
    class_weight=None,          # None or "balanced"
    random_state=42,
    verbose=False,
    width_mode="formula",       # "formula" or "classic"
    alpha=4.0,                  # width scale
    beta=0.7,                   # depth decay
    bias_init="kaiming",        # "kaiming" | "zeros" | "positive" | "uniform"
)

Methods: fit, predict, predict_proba, fit_conformal, predict_set, predict_set_labels.

---

License

MIT License. See LICENSE for details.

---

Acknowledgments

Built independently during undergraduate studies at Beijing Institute of Technology. Inspired by the author's earlier work on robust gradient boosting for small data (SmallGBM).

Проверь, что в Results таблица согласована с реальными числами — если после обновления кода числа изменились, обнови таблицу.