SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 1901.04592 · 2019

Interpretable machine learning: definitions, methods, and applications

arXiv · PDF · Open in the Atlas

Code that ran

11 of 27 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; 2 honoured the contract we drafted, 0 violated it and 9 have no contract verdict; 0 failures of our instrument; 16 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/1901.04592.json · badge: /api/ran/1901.04592.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
csinva/hierarchical-dnn-interpretations canonical (classifier label)
classifier basis: census
11 ran · 16 no recorded run
FunctionStatusWhere it lives
calc_pad Ran
ran, honoured the contract we drafted
csinva/hierarchical-dnn-interpretations/acd/util/conv2dnp.py
pointer only (licence: MIT); the stored copy get_code serves, from nsaphra/acd, is pointer only (licence: NONE) · get_code("54e4782533ba8866")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from nsaphra/acd, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

calc_size Ran
ran, honoured the contract we drafted
csinva/hierarchical-dnn-interpretations/acd/util/conv2dnp.py
pointer only (licence: MIT); the stored copy get_code serves, from nsaphra/acd, is pointer only (licence: NONE) · get_code("4100b072ce597ce6")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from nsaphra/acd, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

extract_sliding_windows_gradw Ran
ran on a synthesized input
csinva/hierarchical-dnn-interpretations/acd/util/conv2dnp.py
code served (permissive licence) · get_code("9c3f257866bb5fb0")
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 6, dtype family float, shape (2, 4, 4, 3, 3, 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

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

# Create small test inputs
# x: [N, H, W, C] - batch of 2, height 6, width 6, channels 3
x = np.random.randn(2, 6, 6, 3).astype(np.float32)

# ksize: [KH, KW] - kernel size
ksize = [3, 3]

# pad: padding method - can be 'SAME', 'VALID', or a number
pad = 'SAME'

# stride: [SH, SW] - stride values
stride = [2, 2]

# orig_size: [H', W'] - original/output size
orig_size = [4, 4]

# floor_first: boolean flag
floor_first = True

ARGS = {
    'x': x,
    'ksize': ksize,
    'pad': pad,
    'stride': stride,
    'orig_size': orig_size,
    'floor_first': floor_first
}
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 6, dtype family float, shape (2, 4, 4, 3, 3, 3) — shape (2, 4, 4, 3, 3, 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
load_data Ran
ran on a synthesized input
csinva/hierarchical-dnn-interpretations/dsets/imagenet/dset.py
code served (permissive licence) · get_code("46f245f91c8c2e4a")
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 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 tempfile
import os
from pathlib import Path

# Create temporary directory structure to mock ImageNet
temp_dir = tempfile.mkdtemp()
train_dir = os.path.join(temp_dir, 'train')
val_dir = os.path.join(temp_dir, 'val')

os.makedirs(train_dir, exist_ok=True)
os.makedirs(val_dir, exist_ok=True)

# Create minimal dummy image files for ImageFolder to work
from PIL import Image
import numpy as np

torch.manual_seed(0)

# Create a few dummy classes with dummy images
for split_dir in [train_dir, val_dir]:
    for class_idx in range(2):
        class_dir = os.path.join(split_dir, f'class_{class_idx}')
        os.makedirs(class_dir, exist_ok=True)
        # Create 2 dummy images per class
        for img_idx in range(2):
            img_array = np.random.randint(0, 256, (224, 224, 3), dtype=np.uint8)
            img = Image.fromarray(img_array)
            img.save(os.path.join(class_dir, f'image_{img_idx}.jpg'))

ARGS = {
    'imagenet_dir': temp_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, torch, torchvision — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, 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
propagate_batchnorm2d Ran
ran on a synthesized input
csinva/hierarchical-dnn-interpretations/acd/scores/cd_propagate.py
code served (permissive licence) · get_code("2c90ac9573576d86")
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 torch
import torch.nn as nn

torch.manual_seed(0)

# Create a simple BatchNorm2d module
module = nn.BatchNorm2d(num_features=4)

# Create small tensors with distinct dimensions
# batch_size=2, channels=4, height=3, width=5
relevant = torch.randn(2, 4, 3, 5)
irrelevant = torch.randn(2, 4, 3, 5)

ARGS = {
    "relevant": relevant,
    "irrelevant": irrelevant,
    "module": module
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — 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
propagate_conv_linear Ran
ran on a synthesized input
csinva/hierarchical-dnn-interpretations/acd/scores/cd_propagate.py
code served (permissive licence) · get_code("9e337dcd6fa467df")
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 torch
import torch.nn as nn

torch.manual_seed(0)

# Create a simple linear module that the function will call
class SimpleLinearModule(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)
    
    def forward(self, x):
        return self.linear(x)

# Create tensors with batch size 2 and feature dim 4
relevant = torch.randn(2, 4)
irrelevant = torch.randn(2, 4)

# Create the module
module = SimpleLinearModule()

ARGS = {
    "relevant": relevant,
    "irrelevant": irrelevant,
    "module": module
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — 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
propagate_pooling Ran
ran on a synthesized input
csinva/hierarchical-dnn-interpretations/acd/scores/cd_propagate.py
code served (permissive licence) · get_code("1b19ddbd7737cc58")
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 5, 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 torch.nn as nn

torch.manual_seed(0)

# Create a minimal pooler class that supports return_indices
class MinimalPooler(nn.Module):
    def __init__(self):
        super().__init__()
        self.return_indices = False
    
    def forward(self, x):
        # x shape: (batch, channels, height, width)
        # Perform max pooling with kernel=2, stride=2
        if self.return_indices:
            return nn.functional.max_pool2d(x, kernel_size=2, stride=2, return_indices=True)
        else:
            return nn.functional.max_pool2d(x, kernel_size=2, stride=2)

# Create small tensors
# relevant shape: (batch=2, channels=4, height=4, width=3)
relevant = torch.randn(2, 4, 4, 3)

# irrelevant shape: (batch=2, channels=4, height=4, width=3)
irrelevant = torch.randn(2, 4, 4, 3)

# Create pooler
pooler = MinimalPooler()

ARGS = {
    "relevant": relevant,
    "irrelevant": irrelevant,
    "pooler": pooler
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 5, 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
refine_scores Ran
ran on a synthesized input
csinva/hierarchical-dnn-interpretations/acd/agglomeration/agg_2d.py
code served (permissive licence) · get_code("35f2273c5de00174")
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 Tensor, ndim 1, dtype family float, shape (2,) 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)

# The function indexes into scores along dimension 1 at index lab_num
# scores needs to be a 2D tensor/array
# lab_num needs to be an integer index valid for the second dimension

scores = torch.randn(2, 4)  # batch size 2, 4 labels
lab_num = 2  # valid index into dimension 1

ARGS = {
    "scores": scores,
    "lab_num": lab_num,
}
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 Tensor, ndim 1, dtype family float, shape (2,) — shape (2,) 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
threshold_scores Ran
ran on a synthesized input
csinva/hierarchical-dnn-interpretations/acd/agglomeration/agg_1d.py
code served (permissive licence) · get_code("6b39e8b9e90c1d43")
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 1, dtype family float, shape (8,) 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 sample scores array with various values and some NaNs
scores = np.array([1.5, -2.3, 0.8, np.nan, 3.2, -1.1, 2.5, 0.3])

# Define percentile_include parameter
percentile_include = 75

# Define absolute parameter
absolute = False

# Construct ARGS dictionary
ARGS = {
    "scores": scores,
    "percentile_include": percentile_include,
    "absolute": absolute
}
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 1, dtype family float, shape (8,) — shape (8,) 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
threshold_scores Ran
ran on a synthesized input
csinva/hierarchical-dnn-interpretations/acd/agglomeration/agg_2d.py
code served (permissive licence) · get_code("2ae7daa6e619d55a")
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 1, dtype family float, shape (10,) 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 scores array with some NaN values
scores = np.array([1.5, 2.3, np.nan, 4.1, 5.6, np.nan, 3.2, 7.8, 6.1, 8.9], dtype=np.float32)

# Set percentile_include to a reasonable value
percentile_include = 50

ARGS = {
    "scores": scores,
    "percentile_include": percentile_include
}
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 1, dtype family float, shape (10,) — shape (10,) 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
visualize_batch_preds Ran
ran on a synthesized input
csinva/hierarchical-dnn-interpretations/visualization/viz_2d.py
code served (permissive licence) · get_code("e02745f340467d1d")
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 (28, 28) 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
import matplotlib.pyplot as plt

# Create small prediction array
preds = np.array([0.1, 0.5, 0.3, 0.9, 0.2, 0.8])

# Create a previous image mask (boolean array of same shape as reshaped grid)
prev_im = np.zeros((28, 28), dtype=bool)
prev_im[0:5, 0:5] = True

ARGS = {
    "preds": preds,
    "prev_im": prev_im,
    "N": 28,
    "im_num_start": 0
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
matplotlib, 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 (28, 28) — shape (28, 28) 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
agglomerate Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/agglomeration/agg_1d.py
code served (permissive licence) · get_code("33d9d5e45e5d68c0")
How it ran

no invocation recorded.

agglomerate Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/agglomeration/agg_2d.py
code served (permissive licence) · get_code("7eae009e30183429")
How it ran

no invocation recorded.

cd Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/scores/cd.py
code served (permissive licence) · get_code("be288b6f3d3d398a")
How it ran

no invocation recorded.

cd_generic Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/scores/cd.py
code served (permissive licence) · get_code("8becf4c93688ea33")
How it ran

no invocation recorded.

cd_propagate_mnist Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/scores/cd_architecture_specific.py
code served (permissive licence) · get_code("7930da855e30bb01")
How it ran

no invocation recorded.

cd_propagate_resnet Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/scores/cd_architecture_specific.py
code served (permissive licence) · get_code("5120f7225c37624b")
How it ran

no invocation recorded.

cd_text Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/scores/cd.py
code served (permissive licence) · get_code("d5f59990e6fe26e2")
How it ran

no invocation recorded.

cd_track_vgg Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/scores/cd_architecture_specific.py
code served (permissive licence) · get_code("9f35efb88f5d2085")
How it ran

no invocation recorded.

collapse_tree Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/agglomeration/agg_1d.py
code served (permissive licence) · get_code("3094108dbe03f52e")
How it ran

no invocation recorded.

discrete_cmap Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/visualization/viz_2d.py
code served (permissive licence) · get_code("0fbfcd3ecb36c8bf")
How it ran

no invocation recorded.

get_im_and_label Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/dsets/imagenet/dset.py
code served (permissive licence) · get_code("f4beec75da6f8e0d")
How it ran

no invocation recorded.

gradient_times_input_scores Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/scores/score_funcs.py
code served (permissive licence) · get_code("557ddf3dd7b26d8b")
How it ran

no invocation recorded.

ig_scores_1d Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/scores/score_funcs.py
code served (permissive licence) · get_code("b5ed9dfff1bbb2d1")
How it ran

no invocation recorded.

ig_scores_2d Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/acd/scores/score_funcs.py
code served (permissive licence) · get_code("d589fb3829819544")
How it ran

no invocation recorded.

pred_ims Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/dsets/imagenet/dset.py
code served (permissive licence) · get_code("170ff6204f66b748")
How it ran

no invocation recorded.

word_heatmap Not yet run
unverified (harvested, no recorded run)
csinva/hierarchical-dnn-interpretations/visualization/viz_1d.py
code served (permissive licence) · get_code("afaa71c3c23899ba")
How it ran

no invocation recorded.

Also present in other papers' code

Of 27 samples lifted from this paper's linked repositories, 27 carry a function identity and 0 do not. Of the 27 distinct functions with an identity, 2 are also present in other papers whose code we hold (2 with at least one paper dated earlier, 0 the same month, 0 later) and 25 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
calc_pad 1 other paper
1 dated earlier
1703.01365 (2017-03)
Axiomatic Attribution for Deep Networks
calc_size 1 other paper
1 dated earlier
1703.01365 (2017-03)
Axiomatic Attribution for Deep Networks

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

Machine-learning models have demonstrated great success in learning complex patterns that enable them to make predictions about unobserved data. In addition to using models for prediction, the ability to interpret what a model has learned is receiving an increasing amount of attention. However, this increased focus has led to considerable confusion about the notion of interpretability. In particular, it is unclear how the wide array of proposed interpretation methods are related, and what common concepts can be used to evaluate them. We aim to address these concerns by defining interpretability in the context of machine learning and introducing the Predictive, Descriptive, Relevant (PDR) framework for discussing interpretations. The PDR framework provides three overarching desiderata for evaluation: predictive accuracy, descriptive accuracy and relevancy, with relevancy judged relative to a human audience. Moreover, to help manage the deluge of interpretation methods, we introduce a categorization of existing techniques into model-based and post-hoc categories, with sub-groups including sparsity, modularity and simulatability. To demonstrate how practitioners can use the PDR framework to evaluate and understand interpretations, we provide numerous real-world examples. These examples highlight the often under-appreciated role played by human audiences in discussions of interpretability. Finally, based on our framework, we discuss limitations of existing methods and directions for future work. We hope that this work will provide a common vocabulary that will make it easier for both practitioners and researchers to discuss and choose from the full range of interpretation methods.

For agents

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

get_harvested_code_for_paper("1901.04592")
get_code_for_paper("1901.04592")
have("1901.04592")

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

curl https://syntology.ai/api/ran/1901.04592.json

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

[![Syntology run record](https://syntology.ai/api/ran/1901.04592.svg)](https://syntology.ai/paper/1901.04592)

Connect an agent — have() is free.