SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2008.03703 · 2020

What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation

arXiv · PDF · Open in the Atlas

Code that ran

4 of 9 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 3 have no contract verdict; 1 failure 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/2008.03703.json · badge: /api/ran/2008.03703.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
google-research/heldout-influence-estimation community (classifier label)
classifier basis not recorded on the link
4 ran · 5 no recorded run
FunctionStatusWhere it lives
distorted_bounding_box_crop Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
google-research/heldout-influence-estimation/libdata/imagenet_preprocessing.py
code served (permissive licence) · get_code("b54e5d5d7eccda19")
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 SymbolicTensor. 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.compat.v1 as tf
import numpy as np

tf.disable_v2_behavior()

# Create a minimal valid JPEG image as bytes
# A minimal 4x3 JPEG (arbitrary small dimensions)
minimal_jpeg = b'\xff\xd8\xff\xe0\x00\x10JFIF\x00\x01\x01\x00\x00\x01\x00\x01\x00\x00\xff\xdb\x00C\x00\x08\x06\x06\x07\x06\x05\x08\x07\x07\x07\t\t\x08\n\x0c\x14\r\x0c\x0b\x0b\x0c\x19\x12\x13\x0f\x14\x1d\x1a\x1f\x1e\x1d\x1a\x1c\x1c $.\' ",#\x1c\x1c(7),01444\x1f\'9=82<.342\xff\xc0\x00\x0b\x08\x00\x03\x00\x04\x01\x01\x11\x00\xff\xc4\x00\x1f\x00\x00\x01\x05\x01\x01\x01\x01\x01\x01\x00\x00\x00\x00\x00\x00\x00\x00\x01\x02\x03\x04\x05\x06\x07\x08\t\n\x0b\xff\xc4\x00\xb5\x10\x00\x02\x01\x03\x03\x02\x04\x03\x05\x05\x04\x04\x00\x00\x01}\x01\x02\x03\x00\x04\x11\x05\x12!1A\x06\x13Qa\x07"q\x142\x81\x91\xa1\x08#B\xb1\xc1\x15R\xd1\xf0$3br\x82\t\n\x16\x17\x18\x19\x1a%&\'()*456789:CDEFGHIJSTUVWXYZcdefghijstuvwxyz\x83\x84\x85\x86\x87\x88\x89\x8a\x92\x93\x94\x95\x96\x97\x98\x99\x9a\xa2\xa3\xa4\xa5\xa6\xa7\xa8\xa9\xaa\xb2\xb3\xb4\xb5\xb6\xb7\xb8\xb9\xba\xc2\xc3\xc4\xc5\xc6\xc7\xc8\xc9\xca\xd2\xd3\xd4\xd5\xd6\xd7\xd8\xd9\xda\xe1\xe2\xe3\xe4\xe5\xe6\xe7\xe8\xe9\xea\xf1\xf2\xf3\xf4\xf5\xf6\xf7\xf8\xf9\xfa\xff\xda\x00\x08\x01\x01\x00\x00?\x00\xfb\xd4\xff\xd9'

# image_bytes: a tensor containing binary JPEG data
image_bytes = tf.constant(minimal_jpeg)

# bbox: tensor of shape [1, num_boxes, 4] with coordinates in [0, 1)
# Using 2 boxes as an example
bbox = tf.constant([
    [
        [0.1, 0.1, 0.5, 0.5],  # [ymin, xmin, ymax, xmax]
        [0.2, 0.2, 0.8, 0.8]
    ]
], dtype=tf.float32)

# min_object_covered: fraction of bounding box that must be in crop
min_object_covered = 0.1

# aspect_ratio_range: (min_aspect, max_aspect) where aspect = width / height
aspect_ratio_range = (0.75, 1.33)

# area_range: (min_area, max_area) as fractions of original image
area_range = (0.05, 1.0)

# max_attempts: number of attempts to generate valid crop
max_attempts = 100

# scope: optional name scope
scope = None

ARGS = {
    'image_bytes': image_bytes,
    'bbox': bbox,
    'min_object_covered': min_object_covered,
    'aspect_ratio_range': aspect_ratio_range,
    'area_range': area_range,
    'max_attempts': max_attempts,
    'scope': scope
}
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 SymbolicTensor — a SymbolicTensor 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
one_hot Ran
ran on a synthesized input
google-research/heldout-influence-estimation/mnist-example/mnist_infl_mem.py
code served (permissive licence) · get_code("a8f998b9ec355770")
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 float, 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

# Construct arguments for one_hot function
x = np.array([0, 1, 2, 1])  # Small array of class indices
k = 3  # Number of classes for one-hot encoding
dtype = np.float32  # Data type for output

ARGS = {
    "x": x,
    "k": k,
    "dtype": dtype,
}
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 float, 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_for_eval Ran
ran on a synthesized input
google-research/heldout-influence-estimation/libdata/imagenet_preprocessing.py
code served (permissive licence) · get_code("46fa247f2c3e6e4d")
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 SymbolicTensor. 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.compat.v1 as tf
import numpy as np

# Disable v2 behavior for TF 1.x compatibility
tf.disable_v2_behavior()

# Create a small valid JPEG image as bytes
# We'll create a minimal valid JPEG programmatically
def create_test_jpeg_bytes():
    """Create a minimal valid JPEG image as bytes."""
    # Create a simple RGB image using numpy
    image_array = np.random.randint(0, 256, (100, 120, 3), dtype=np.uint8)
    
    # Use TensorFlow to encode it as JPEG
    with tf.Session() as sess:
        jpeg_bytes = sess.run(tf.image.encode_jpeg(image_array))
    return jpeg_bytes

# Generate the JPEG bytes
image_bytes_value = create_test_jpeg_bytes()

# Create the ARGS dictionary with all parameters for preprocess_for_eval
ARGS = {
    "image_bytes": image_bytes_value,
    "dtype": tf.float32,
    "image_size": 224
}
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 SymbolicTensor — a SymbolicTensor 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
preprocess_for_train Ran
ran on a synthesized input
google-research/heldout-influence-estimation/libdata/imagenet_preprocessing.py
code served (permissive licence) · get_code("c46504ca7d0285e1")
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 SymbolicTensor. 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.compat.v1 as tf
import numpy as np

tf.disable_v2_behavior()

# Create a small valid JPEG image in memory
def create_jpeg_bytes():
    """Create a minimal valid JPEG image as bytes."""
    # Create a small 256x256 RGB image
    img_array = np.random.randint(0, 256, (256, 256, 3), dtype=np.uint8)
    # Encode as JPEG
    jpeg_encoded = tf.image.encode_jpeg(img_array)
    with tf.Session() as sess:
        jpeg_bytes = sess.run(jpeg_encoded)
    return jpeg_bytes

# Generate the JPEG bytes
image_bytes_value = create_jpeg_bytes()

# Create the ARGS dictionary
ARGS = {
    'image_bytes': tf.constant(image_bytes_value),
    'dtype': tf.float32,
    'image_size': 224,
}
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 SymbolicTensor — a SymbolicTensor 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
batch_correctness Not yet run
unverified (harvested, no recorded run)
google-research/heldout-influence-estimation/mnist-example/mnist_infl_mem.py
code served (permissive licence) · get_code("84d9720661af92de")
How it ran

no invocation recorded.

load_data Not yet run
unverified (harvested, no recorded run)
google-research/heldout-influence-estimation/libdata/indexed_tfrecords.py
code served (permissive licence) · get_code("223b796c7c48a3c1")
How it ran

no invocation recorded.

loss Not yet run
unverified (harvested, no recorded run)
google-research/heldout-influence-estimation/mnist-example/mnist_infl_mem.py
code served (permissive licence) · get_code("8f1d8f0c8fede07a")
How it ran

no invocation recorded.

parse_record Not yet run
unverified (harvested, no recorded run)
google-research/heldout-influence-estimation/libdata/indexed_tfrecords.py
code served (permissive licence) · get_code("dd68508cbccc4bf8")
How it ran

no invocation recorded.

process_record_dataset Not yet run
unverified (harvested, no recorded run)
google-research/heldout-influence-estimation/libdata/indexed_tfrecords.py
code served (permissive licence) · get_code("c00febd1ff3e9ed5")
How it ran

no invocation recorded.

Also present in other papers' code

Of 9 samples lifted from this paper's linked repositories, 9 carry a function identity and 0 do not. Of the 9 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 8 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.

FunctionAlso present inEarliest-datedOthers named
distorted_bounding_box_crop 2 other papers
2 dated earlier
1904.02877 (2019-04)
Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4 Hours
2006.14536 (2020-06)
Smooth Adversarial Training

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

Deep learning algorithms are well-known to have a propensity for fitting the training data very well and often fit even outliers and mislabeled data points. Such fitting requires memorization of training data labels, a phenomenon that has attracted significant research interest but has not been given a compelling explanation so far. A recent work of Feldman (2019) proposes a theoretical explanation for this phenomenon based on a combination of two insights. First, natural image and data distributions are (informally) known to be long-tailed, that is have a significant fraction of rare and atypical examples. Second, in a simple theoretical model such memorization is necessary for achieving close-to-optimal generalization error when the data distribution is long-tailed. However, no direct empirical evidence for this explanation or even an approach for obtaining such evidence were given. In this work we design experiments to test the key ideas in this theory. The experiments require estimation of the influence of each training example on the accuracy at each test example as well as memorization values of training examples. Estimating these quantities directly is computationally prohibitive but we show that closely-related subsampled influence and memorization values can be estimated much more efficiently. Our experiments demonstrate the significant benefits of memorization for generalization on several standard benchmarks. They also provide quantitative and visually compelling evidence for the theory put forth in (Feldman, 2019).

For agents

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

get_harvested_code_for_paper("2008.03703")
get_code_for_paper("2008.03703")
have("2008.03703")

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

curl https://syntology.ai/api/ran/2008.03703.json

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

[![Syntology run record](https://syntology.ai/api/ran/2008.03703.svg)](https://syntology.ai/paper/2008.03703)

Connect an agent — have() is free.