SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2106.11872 · 2021

Randomness In Neural Network Training: Characterizing The Impact of Tooling

arXiv · PDF · Open in the Atlas

Code that ran

9 of 14 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; 0 honoured the contract we drafted, 0 violated it and 9 have no contract verdict; 0 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/2106.11872.json · badge: /api/ran/2106.11872.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
usyd-fsalab/NeuralNetworkRandomness canonical (classifier label)
classifier basis: census
9 ran · 5 no recorded run
FunctionStatusWhere it lives
MediumCNN Ran
ran on a synthesized input
usyd-fsalab/NeuralNetworkRandomness/src/common/cnns.py
code served (permissive licence) · get_code("b587e238a34a5ea6")
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 Functional. 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 tensorflow as tf

# The function MediumCNN takes a single parameter: kernel_size
# kernel_size is used as the kernel size for Conv2D layers
# It should be an integer representing the height/width of the convolutional kernel

ARGS = {
    "kernel_size": 3
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Functional — a Functional output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Functional 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
ResidualBlock Ran
ran on a synthesized input
usyd-fsalab/NeuralNetworkRandomness/src/common/resnet.py
code served (permissive licence) · get_code("11da740fe6395a4a")
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 KerasTensor. 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 tensorflow as tf
from tensorflow.keras import layers, Input
import numpy as np

# Create a small input tensor for the residual block
# Using shape (batch_size, height, width, channels) = (2, 4, 4, 3)
input_tensor = Input(shape=(4, 4, 3))

ARGS = {
    'x': input_tensor,
    'filters': 16,
    'kernel_size': 3,
    'weight_decay': 0.0001,
    'seed': 42,
    'downsample': 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 KerasTensor — a KerasTensor output has no array shape of its own; none is 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
celebA_transform Ran
ran on a synthesized input
usyd-fsalab/NeuralNetworkRandomness/src/training_script/resnet_celeba.py
code served (permissive licence) · get_code("7ceef7834e740164")
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. 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 tensorflow as tf
import numpy as np

# Create a minimal data dict that satisfies the function's requirements
# The function expects:
# - data['image']: a tensor that can be resized to (128, 128)
# - data['attributes']: a dict where values can be converted to tensor elements

# Create a small image tensor (tiny dimensions for testing)
image_array = np.random.rand(64, 64, 3).astype(np.float32)
image_tensor = tf.constant(image_array)

# Create attributes dict with a few sample attributes
attributes_dict = {
    'gender': 1.0,
    'age': 0.5,
    'smile': 1.0,
    'glasses': 0.0
}

# Construct the data dict
data = {
    'image': image_tensor,
    'attributes': attributes_dict
}

ARGS = {
    'data': data
}
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 tuple — a tuple output has no array shape of its own; none is 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
cifar_random_crop_and_flip Ran
ran on a synthesized input
usyd-fsalab/NeuralNetworkRandomness/src/common/input_pipeline.py
code served (permissive licence) · get_code("6ec7fa06967e3f07")
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 ndarray, ndim 3, dtype family float, shape (32, 32, 3) 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 torch
import numpy as np

torch.manual_seed(0)

# Create a small test image tensor
# The function expects a tensor with .numpy() method
# CIFAR images are 32x32x3, so we'll create that shape
image = torch.randint(0, 256, (32, 32, 3), dtype=torch.uint8).float()

ARGS = {
    "image": image
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 3, dtype family float, shape (32, 32, 3) — shape (32, 32, 3) 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
conv2d_bn Ran
ran on a synthesized input
usyd-fsalab/NeuralNetworkRandomness/src/common/resnet.py
code served (permissive licence) · get_code("98b375ca4e1053fc")
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, 16) 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 input tensor with shape (batch=2, height=4, width=4, channels=3)
x = tf.random.normal((2, 4, 4, 3))

ARGS = {
    'x': x,
    'filters': 16,
    'kernel_size': (3, 3),
    'seed': 42,
    'weight_decay': 0.01,
    'strides': (1, 1)
}
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, 16) — shape (2, 4, 4, 16) 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
conv2d_bn_relu Ran
ran on a synthesized input
usyd-fsalab/NeuralNetworkRandomness/src/common/resnet.py
code served (permissive licence) · get_code("be3aecbe3c13892c")
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, 16) 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 seeds for reproducibility
np.random.seed(0)
tf.random.set_seed(0)

# Create a small input tensor (batch_size=2, height=4, width=4, channels=3)
x = tf.random.normal(shape=(2, 4, 4, 3))

ARGS = {
    'x': x,
    'filters': 16,
    'kernel_size': (3, 3),
    'seed': 42,
    'weight_decay': 0.0001,
    'strides': (1, 1)
}
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, 16) — shape (2, 4, 4, 16) 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
get_standard_deviation_of_accuracy Ran
ran on a synthesized input
usyd-fsalab/NeuralNetworkRandomness/run_ae.py
code served (permissive licence) · get_code("e120ab8759b2766a")
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 float64, ndim 0, 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 os
import csv
import tempfile
import numpy as np
from pathlib import Path

# Create a temporary directory structure with mock CSV files
temp_dir = tempfile.mkdtemp()

# Define the ModelConfig and related enums
class DeterministicSetting:
    def __init__(self, value):
        self.value = value

class Models:
    def __init__(self, value):
        self.value = value

class Datasets:
    def __init__(self, value):
        self.value = value

class ModelConfig:
    def __init__(self, model, dataset, determinism, num_models, additional_args=None):
        self.model = model
        self.dataset = dataset
        self.determinism = determinism
        self.num_models = num_models
        self.additional_args = additional_args
    
    def get_ckpt_dir_prefix(self):
        prefix = f'logs_{self.model.value}_{self.dataset.value}_{self.determinism.value}_'
        if self.additional_args and '--bn' in self.additional_args:
            prefix += '_bn'
        return os.path.join(temp_dir, prefix)

# Create mock CSV files with validation accuracy data
num_models = 3
for idx in range(num_models):
    model_dir = os.path.join(temp_dir, f'logs_smallcnn_cifar10_default_{idx}')
    os.makedirs(model_dir, exist_ok=True)
    
    csv_path = os.path.join(model_dir, 'log.csv')
    with open(csv_path, 'w', newline='') as f:
        writer = csv.writer(f)
        writer.writerow(['epoch', 'loss', 'val_accuracy'])
        writer.writerow(['1', '0.5', '0.80'])
        writer.writerow(['2', '0.3', '0.85'])
        writer.writerow(['3', '0.2', '0.90'])

# Create the ModelConfig instance
model = Models('smallcnn')
dataset = Datasets('cifar10')
determinism = DeterministicSetting('default')

config = ModelConfig(
    model=model,
    dataset=dataset,
    determinism=determinism,
    num_models=num_models,
    additional_args=None
)

# Override get_ckpt_dir_prefix to return correct test path
original_get_ckpt_dir_prefix = config.get_ckpt_dir_prefix

def mock_get_ckpt_dir_prefix():
    return os.path.join(temp_dir, 'logs_smallcnn_cifar10_default_')

config.get_ckpt_dir_prefix = mock_get_ckpt_dir_prefix

ARGS = {
    'config': config
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
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
horizontal_flip Ran
ran on a synthesized input
usyd-fsalab/NeuralNetworkRandomness/src/common/input_pipeline.py
code served (permissive licence) · get_code("eafcda6bb13a0e7c")
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 ndarray, ndim 2, dtype family int, shape (4, 3) 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 numpy as np

# Create a small sample image as a numpy array
# Using shape (4, 3) to have distinct dimensions
image = np.random.randint(0, 256, size=(4, 3), dtype=np.uint8)

# axis can be 0 or 1 based on the function's assertion
axis = 0

ARGS = {
    "image": image,
    "axis": axis
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family int, shape (4, 3) — shape (4, 3) 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
preprocess_image Ran
ran on a synthesized input
usyd-fsalab/NeuralNetworkRandomness/src/training_script/smallcnn.py
code served (permissive licence) · get_code("0c8bdadaba87adfd")
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. 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 tensorflow as tf
import numpy as np

# Create a sample data dictionary with an image and label
# The function expects data['image'] and data['label']
image_array = np.random.randint(0, 256, size=(4, 3, 3), dtype=np.uint8)  # 4x3x3 image
label_value = np.array(5, dtype=np.int32)  # A label between 0-9

data = {
    'image': image_array,
    'label': label_value
}

ARGS = {
    'data': data
}
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 tuple — a tuple output has no array shape of its own; none is 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
cifar_data_preprocessing Not yet run
unverified (harvested, no recorded run)
usyd-fsalab/NeuralNetworkRandomness/src/common/input_pipeline.py
code served (permissive licence) · get_code("52b7a2ba9ec2f57c")
How it ran

no invocation recorded.

get_standard_deviation_of_fnr Not yet run
unverified (harvested, no recorded run)
usyd-fsalab/NeuralNetworkRandomness/run_ae.py
code served (permissive licence) · get_code("7e9d8b95a3c7afeb")
How it ran

no invocation recorded.

get_standard_deviation_of_fpr Not yet run
unverified (harvested, no recorded run)
usyd-fsalab/NeuralNetworkRandomness/run_ae.py
code served (permissive licence) · get_code("b990fe5855fdceff")
How it ran

no invocation recorded.

lr_schedule Not yet run
unverified (harvested, no recorded run)
usyd-fsalab/NeuralNetworkRandomness/src/training_script/smallcnn.py
code served (permissive licence) · get_code("9f3fb2653030c316")
How it ran

no invocation recorded.

lr_scheduler Not yet run
unverified (harvested, no recorded run)
usyd-fsalab/NeuralNetworkRandomness/src/training_script/resnet_celeba.py
code served (permissive licence) · get_code("384218efc7e9fdd8")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 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 14 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

The quest for determinism in machine learning has disproportionately focused on characterizing the impact of noise introduced by algorithmic design choices. In this work, we address a less well understood and studied question: how does our choice of tooling introduce randomness to deep neural network training. We conduct large scale experiments across different types of hardware, accelerators, state of art networks, and open-source datasets, to characterize how tooling choices contribute to the level of non-determinism in a system, the impact of said non-determinism, and the cost of eliminating different sources of noise. Our findings are surprising, and suggest that the impact of non-determinism in nuanced. While top-line metrics such as top-1 accuracy are not noticeably impacted, model performance on certain parts of the data distribution is far more sensitive to the introduction of randomness. Our results suggest that deterministic tooling is critical for AI safety. However, we also find that the cost of ensuring determinism varies dramatically between neural network architectures and hardware types, e.g., with overhead up to $746\%$, $241\%$, and $196\%$ on a spectrum of widely used GPU accelerator architectures, relative to non-deterministic training. The source code used in this paper is available at https://github.com/usyd-fsalab/NeuralNetworkRandomness.

For agents

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

get_harvested_code_for_paper("2106.11872")
get_code_for_paper("2106.11872")
have("2106.11872")

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

curl https://syntology.ai/api/ran/2106.11872.json

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

[![Syntology run record](https://syntology.ai/api/ran/2106.11872.svg)](https://syntology.ai/paper/2106.11872)

Connect an agent — have() is free.