Demoren/nes-surrogate-architectures

NES Surrogate Dataset

Overview

This dataset contains trained neural architectures, their predictions, and validation performance, designed for studying:

• surrogate modeling of neural architectures
• diversity estimation between models
• ensemble construction strategies

Each architecture is associated with:

• its structure (DARTS-like cell)
• model weights
• validation predictions
• validation accuracy

---

Dataset Structure

CIFAR10/
CIFAR100/
FashionMNIST/

Each dataset directory contains:

architectures/
weights/

architectures/

Each JSON file contains:

• architecture definition (DARTS-like DAG)
• validation predictions
• validation accuracy

Predictions are computed on a shared validation split, enabling construction of pairwise similarity matrices.

weights/

Contains trained model weights corresponding to each architecture in safetensors format.

---

Example Entry

{
  "architecture": {
    "normal/op_2_0": "sep_conv_3x3",
    "normal/input_2_0": [1],
    "normal/op_2_1": "sep_conv_5x5",
    "normal/input_2_1": [0]
  },
  "valid_predictions": [6, 2, 5, 6, 3],
  "valid_accuracy": 0.76
}

---

Dataset Size

• ~3,000 models per dataset
• Total: architectures, predictions, and weights for three benchmarks

---

Data Split

For each dataset, the original training set is split into:

• 20% training subset
• 80% validation subset

The split is performed using:

• fixed random seed: 42
• torch.utils.data.Subset

The validation subset is used to:

• compute model accuracy
• generate prediction vectors for diversity estimation

---

Training Setup

Architectures are trained under a reduced configuration to limit computational cost:

• optimizer: SGD
• learning rate: cosine schedule from 0.025 → 1e-3
• weight decay: 3e-4
• batch size: 96
• auxiliary loss weight: 0.4

Training Statistics

Dataset	Num. Cells	Initial Width	Num. Epochs	Avg. Accuracy (%)	Avg. Top-1 Agreement
FashionMNIST	3	16	125	89.6 ± 0.5	0.900 ± 0.004
CIFAR-10	8	16	200	75.8 ± 0.6	0.693 ± 0.006
CIFAR-100	8	16	200	37.6 ± 1.1	0.324 ± 0.008

Note: Models are not trained to full convergence. They are trained for a fixed number of epochs sufficient to obtain reliable relative performance estimates.

---

Key Properties

• DARTS-like architecture search space
• Graph-based representation (DAGs)
• Aligned predictions across models
• Supports diversity estimation via prediction similarity
• Suitable for surrogate-based ranking and selection

---

Intended Use

This dataset enables:

• training accuracy surrogate models
• learning diversity embeddings (e.g., via triplet loss)
• constructing similarity matrices between models
• analyzing relationships between architecture and predictions

Example similarity metric:

similarity = (y_i == y_j).mean()

---

Relation to Paper

This dataset accompanies the paper:

"Surrogate Assisted Diversity Estimation in Neural Ensemble Search"

It is used to:

• train surrogate models for accuracy and diversity
• guide ensemble construction
• study scaling behavior with respect to dataset size

---

Notes

• Similarity between models is computed from prediction agreement
• The dataset is designed for relative ranking, not absolute accuracy
• Can be used with alternative diversity metrics (e.g., correlation, divergence)

---