SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2103.16483 · CVPR · 2021

Benchmarking Representation Learning for Natural World Image Collections

Serge Belongie, Elijah Cole, Grant Van Horn, Sara Beery, Oisin Mac Aodha, Kimberly Wilber

arXiv · PDF · Open in the Atlas

Code that ran

7 of 12 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 6 failures of our instrument; 5 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2103.16483.json · badge: /api/ran/2103.16483.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
visipedia/newt canonical (classifier label)
classifier basis: census
7 ran · 5 no recorded run
FunctionStatusWhere it lives
conv_block Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
visipedia/newt/benchmark/tf_resnet.py
code served (permissive licence) · get_code("cbe50e0989ae1e3f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned EagerTensor, ndim 4, dtype family float, shape (2, 4, 4, 128) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tensorflow as tf
import numpy as np

# Set random seed for reproducibility
tf.random.set_seed(0)
np.random.seed(0)

# Create a small input tensor
# Shape: [batch=2, height=8, width=8, channels=64]
# Using channels_last format (default for TensorFlow)
input_tensor = tf.random.normal([2, 8, 8, 64], dtype=tf.float32)

# Define arguments for conv_block
ARGS = {
    'input_tensor': input_tensor,
    'kernel_size': 3,
    'filters': [64, 64, 128],  # filters1, filters2, filters3
    'stage': 2,
    'block': 'a',
    'strides': (2, 2),
    'use_l2_regularizer': True,
    'batch_norm_decay': 0.9,
    'batch_norm_epsilon': 1e-05,
    'batch_norm_trainable': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, ndim 4, dtype family float, shape (2, 4, 4, 128) — shape (2, 4, 4, 128) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
evalute_features Ran
ran; our contract draft was wrong, not the code
visipedia/newt/benchmark/evaluate_linear_models.py
code served (permissive licence) · get_code("47c40e00f535da70")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
import pandas as pd

torch.manual_seed(0)

# Create a simple evaluator function that returns a dict of results
def simple_evaluator_fn(X_train, y_train, X_test, y_test):
    """
    Simple evaluator that takes train/test splits and returns results dict.
    """
    return {
        'accuracy': 0.95,
        'loss': 0.05,
        'f1': 0.93
    }

# Create sample data for multiple datasets
datasets = []

# Dataset 1
X_train_1 = np.random.randn(10, 4)
X_test_1 = np.random.randn(5, 4)
y_train_1 = np.random.randint(0, 2, 10)
y_test_1 = np.random.randint(0, 2, 5)

datasets.append({
    'name': 'dataset_1',
    'X_train': X_train_1,
    'X_test': X_test_1,
    'y_train': y_train_1,
    'y_test': y_test_1
})

# Dataset 2
X_train_2 = np.random.randn(10, 4)
X_test_2 = np.random.randn(5, 4)
y_train_2 = np.random.randint(0, 2, 10)
y_test_2 = np.random.randint(0, 2, 5)

datasets.append({
    'name': 'dataset_2',
    'X_train': X_train_2,
    'X_test': X_test_2,
    'y_train': y_train_2,
    'y_test': y_test_2
})

# Create features dataframe
features_df = pd.DataFrame(datasets)

ARGS = {
    'features_df': features_df,
    'evaluator_fn': simple_evaluator_fn
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas, torch, tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
fixed_padding Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
visipedia/newt/benchmark/tf_resnet.py
code served (permissive licence) · get_code("43006b1579cc0b34")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned EagerTensor, ndim 4, dtype family float, shape (2, 6, 5, 5) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tensorflow as tf
import numpy as np

# Create a small test tensor with shape [batch=2, height=4, width=3, channels=5]
# Using channels_last format (default)
inputs_channels_last = tf.constant(
    np.random.randn(2, 4, 3, 5).astype(np.float32)
)

# Create a small test tensor with shape [batch=2, channels=5, height=4, width=3]
# Using channels_first format
inputs_channels_first = tf.constant(
    np.random.randn(2, 5, 4, 3).astype(np.float32)
)

# Define ARGS dict with valid arguments
ARGS = {
    'inputs': inputs_channels_last,
    'kernel_size': 3,
    'data_format': 'channels_last'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, ndim 4, dtype family float, shape (2, 6, 5, 5) — shape (2, 6, 5, 5) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
linearsvc Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
visipedia/newt/benchmark/linear_evaluation.py
code served (permissive licence) · get_code("d9c3f2f2c048d6a2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
from sklearn.svm import LinearSVC
from sklearn.model_selection import GridSearchCV, PredefinedSplit
from sklearn.metrics import accuracy_score
import sklearn.preprocessing

# Create small synthetic data
np.random.seed(0)
n_train = 10
n_test = 4
n_features = 6
n_classes = 2

X_train = np.random.randn(n_train, n_features).astype(np.float32)
y_train = np.random.randint(0, n_classes, n_train)

X_test = np.random.randn(n_test, n_features).astype(np.float32)
y_test = np.random.randint(0, n_classes, n_test)

# Create predefined validation indices (for PredefinedSplit)
# -1 means training, 0 means validation fold
predefined_val_indices = np.array([-1, -1, -1, 0, 0, -1, -1, -1, -1, 0])

ARGS = {
    'X_train': X_train,
    'y_train': y_train,
    'X_test': X_test,
    'y_test': y_test,
    'max_iter': 1000,
    'grid_search': False,
    'predefined_val_indices': None,
    'standardize': False,
    'normalize': True,
    'dual': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, sklearn — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_newt_task Ran
ran; our contract draft was wrong, not the code
visipedia/newt/benchmark/dataset_utils.py
code served (permissive licence) · get_code("11489c1ebee3c555")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 2, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import os
import tempfile
import shutil

# Create a temporary directory structure that mimics a NeWT task
task_dir = tempfile.mkdtemp()

# Create train.json
train_dataset = {
    "images": [
        {"id": 1, "filename": "train_img_1.jpg"},
        {"id": 2, "filename": "train_img_2.jpg"}
    ],
    "annotations": [
        {"image_id": 1, "category_id": 0},
        {"image_id": 2, "category_id": 1}
    ]
}

# Create test.json
test_dataset = {
    "images": [
        {"id": 3, "filename": "test_img_1.jpg"},
        {"id": 4, "filename": "test_img_2.jpg"}
    ],
    "annotations": [
        {"image_id": 3, "category_id": 0},
        {"image_id": 4, "category_id": 1}
    ]
}

# Write JSON files
with open(os.path.join(task_dir, 'train.json'), 'w') as f:
    json.dump(train_dataset, f)

with open(os.path.join(task_dir, 'test.json'), 'w') as f:
    json.dump(test_dataset, f)

# Create dummy image files
for img_info in train_dataset["images"] + test_dataset["images"]:
    img_path = os.path.join(task_dir, img_info["filename"])
    with open(img_path, 'w') as f:
        f.write("dummy image content")

ARGS = {
    "task_dir": task_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 2, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
loader Ran
ran, honoured the contract we drafted
visipedia/newt/benchmark/pt_resnet_feature_extractor.py
code served (permissive licence) · get_code("65dab949aa3e1b87")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Image. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from PIL import Image
import os

# Create a temporary directory and a simple RGB image file
temp_dir = tempfile.mkdtemp()
img_path = os.path.join(temp_dir, "test_image.png")

# Create a small test image and save it
test_img = Image.new('RGB', (4, 3), color=(255, 0, 0))
test_img.save(img_path)

ARGS = {
    'path': img_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Image — a Image output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Image output the recorded ndim 3 and dtype family int describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
sgd Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
visipedia/newt/benchmark/linear_evaluation.py
code served (permissive licence) · get_code("cb90fce2472de4f8")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
from sklearn.preprocessing import StandardScaler

# Create small synthetic dataset
np.random.seed(0)
n_train = 10
n_test = 5
n_features = 4
n_classes = 2

X_train = np.random.randn(n_train, n_features).astype(np.float32)
y_train = np.random.randint(0, n_classes, n_train)

X_test = np.random.randn(n_test, n_features).astype(np.float32)
y_test = np.random.randint(0, n_classes, n_test)

# Create predefined validation indices for the PredefinedSplit
# -1 means training fold, 0 means validation fold
predefined_val_indices = np.array([-1] * 7 + [0] * 3)

ARGS = {
    'X_train': X_train,
    'y_train': y_train,
    'X_test': X_test,
    'y_test': y_test,
    'max_iter': 1000,
    'loss_type': 'hinge',
    'grid_search': False,
    'predefined_val_indices': None,
    'standardize': False,
    'normalize': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, sklearn — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
identity_block Not yet run
unverified (harvested, no recorded run)
visipedia/newt/benchmark/tf_resnet.py
code served (permissive licence) · get_code("2c7bf74214627a61")
How it ran

no invocation recorded.

load_cub Not yet run
unverified (harvested, no recorded run)
visipedia/newt/benchmark/dataset_utils.py
code served (permissive licence) · get_code("26d13a1ab4665887")
How it ran

no invocation recorded.

load_feature_extractor Not yet run
unverified (harvested, no recorded run)
visipedia/newt/benchmark/pt_resnet_feature_extractor.py
code served (permissive licence) · get_code("a4f41438107ab2a7")
How it ran

no invocation recorded.

load_oxford_flowers Not yet run
unverified (harvested, no recorded run)
visipedia/newt/benchmark/dataset_utils.py
code served (permissive licence) · get_code("bfe9138310b695d3")
How it ran

no invocation recorded.

logreg Not yet run
unverified (harvested, no recorded run)
visipedia/newt/benchmark/linear_evaluation.py
code served (permissive licence) · get_code("7bde24b55ef593b2")
How it ran

no invocation recorded.

Also present in other papers' code

Of 12 samples lifted from this paper's linked repositories, 12 carry a function identity and 0 do not. Of the 12 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 12 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Recent progress in self-supervised learning has resulted in models that are capable of extracting rich representations from image collections without requiring any explicit label supervision. However, to date the vast majority of these approaches have restricted themselves to training on standard benchmark datasets such as ImageNet. We argue that fine-grained visual categorization problems, such as plant and animal species classification, provide an informative testbed for self-supervised learning. In order to facilitate progress in this area we present two new natural world visual classification datasets, iNat2021 and NeWT. The former consists of 2.7M images from 10k different species uploaded by users of the citizen science application iNaturalist. We designed the latter, NeWT, in collaboration with domain experts with the aim of benchmarking the performance of representation learning algorithms on a suite of challenging natural world binary classification tasks that go beyond standard species classification. These two new datasets allow us to explore questions related to large-scale representation and transfer learning in the context of finegrained categories. We provide a comprehensive analysis of feature extractors trained with and without supervision on ImageNet and iNat2021, shedding light on the strengths and weaknesses of different learned features across a diverse set of tasks. We find that features produced by standard supervised methods still outperform those produced by self-supervised approaches such as SimCLR. However, improved self-supervised learning methods are constantly being released and the iNat2021 and NeWT datasets are a valuable resource for tracking their progress.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2103.16483")
get_code_for_paper("2103.16483")
have("2103.16483")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2103.16483.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2103.16483.svg)](https://syntology.ai/paper/2103.16483)

Connect an agent — have() is free.