SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 1704.02685 · 2017

Learning Important Features Through Propagating Activation Differences

Peyton Greenside, Anshul Kundaje, Avanti Shrikumar

arXiv · PDF · Open in the Atlas

Code that ran

8 of 15 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (2 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 1 violated it and 4 have no contract verdict; 3 failures of our instrument; 7 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/1704.02685.json · badge: /api/ran/1704.02685.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
shaoshanglqy/shap-shapley community
Papers with Code archive
3 ran · 4 no recorded run
vlue-c/Visual-Explanation-Methods-PyTorch community
Papers with Code archive
2 ran · 0 no recorded run
austinbrown34/shap community
Papers with Code archive
1 ran · 2 no recorded run
copy not recorded community
not in the paper's declared repositories
1 ran · 1 no recorded run
saivarunr/xshap community
Papers with Code archive
1 ran · 0 no recorded run
FunctionStatusWhere it lives
DeepLift Ran
ran on a synthesized input · constructed an object; no result was computed
vlue-c/Visual-Explanation-Methods-PyTorch/torchvex/contrib/deeplift.py
code served (permissive licence) · get_code("f70e444e11ee8d3a")
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 an instance of its own class, DeepLift: the run constructed the object and did not call it, so no result was computed. 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
from numbers import Number

torch.manual_seed(0)

# Create a simple model that takes inputs and produces class logits
class SimpleModel(nn.Module):
    def __init__(self, input_dim, num_classes):
        super().__init__()
        self.linear = nn.Linear(input_dim, num_classes)
    
    def forward(self, x):
        return self.linear(x)

# Create model with input_dim=4, num_classes=3
model = SimpleModel(input_dim=4, num_classes=3)

# Create simple preprocessing function
def preprocess_fn(x):
    return x * 2.0

# Create simple postprocessing function
def postprocess_fn(x):
    return torch.abs(x)

# Create baseline tensor (batch_size=2, dim=4)
baseline = torch.randn(2, 4)

ARGS = {
    "model": model,
    "baseline": baseline,
    "preprocess": preprocess_fn,
    "postprocess": postprocess_fn,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
captum, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DeepLift — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a DeepLift 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
ExplanationMethod Ran
ran on a synthesized input · constructed an object; no result was computed
vlue-c/Visual-Explanation-Methods-PyTorch/torchvex/contrib/deeplift.py
code served (permissive licence) · get_code("baff092b503bcfbe")
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 an instance of its own class, ExplanationMethod: the run constructed the object and did not call it, so no result was computed. 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
from numbers import Number

torch.manual_seed(0)

# Define a simple model that takes inputs and outputs logits
class SimpleModel(nn.Module):
    def __init__(self, input_dim, num_classes):
        super().__init__()
        self.linear = nn.Linear(input_dim, num_classes)
    
    def forward(self, x):
        return self.linear(x)

# Define a simple preprocess function
def preprocess_fn(x):
    return x * 0.5

# Define a simple postprocess function
def postprocess_fn(x):
    return torch.clamp(x, 0, 1)

# Define a concrete ExplanationMethod subclass for testing
class ConcreteExplanationMethod(torch.nn.Module):
    def __init__(self, model, preprocess, postprocess):
        super().__init__()
        self.model = model
        self.preprocess = preprocess
        self.postprocess = postprocess
    
    def process(self, inputs, target):
        # Simple implementation: return inputs as-is
        return inputs
    
    def _forward(self, inputs, target=None):
        if target is None:
            target = self.model(inputs).argmax(1)
        if isinstance(target, Number):
            target = [target]
        if not isinstance(target, torch.Tensor):
            target = torch.as_tensor(target)
        out = inputs
        if self.preprocess is not None:
            out = self.preprocess(inputs)
        if out.device != target.device:
            target = target.to(out.device)
        out = self.process(out, target)
        if self.postprocess is not None:
            out = self.postprocess(out)
        return out
    
    forward = _forward

# Create instances
model = SimpleModel(input_dim=4, num_classes=3)
preprocess = preprocess_fn
postprocess = postprocess_fn

ARGS = {
    "model": model,
    "preprocess": preprocess,
    "postprocess": postprocess,
}
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 ExplanationMethod — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ExplanationMethod 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
PyTorchDeepExplainer Ran
ran on a synthesized input
austinbrown34/shap/shap/explainers/deep/deep_pytorch.py
code served (permissive licence) · get_code("17c52f66fb9f49e0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned list, ndim 3, 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
import numpy as np

# Create a simple model for testing
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(5, 8)
        self.relu = nn.ReLU()
        self.fc2 = nn.Linear(8, 3)
    
    def forward(self, x):
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        return x

# Create model and background data
model = SimpleModel()
model.eval()

# Background data for the explainer (batch of 4 samples, 5 features each)
background_data = torch.randn(4, 5)

# Initialize the explainer
INIT_ARGS = {
    'model': model,
    'data': background_data
}

# Data to explain (batch of 2 samples, 5 features each)
test_data = torch.randn(2, 5)

# The computation method is shap_values
METHOD = 'shap_values'

ARGS = {
    'X': test_data
}
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
distutils, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 3, 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
PyTorchDeepExplainer Ran
ran on a synthesized input
saivarunr/xshap/shap/explainers/deep/deep_pytorch.py
code served (permissive licence) · get_code("315e69a6733bc503")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned list, ndim 3, 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
import numpy as np

# Create a simple PyTorch model
class SimpleModel(nn.Module):
    def __init__(self):
        super(SimpleModel, self).__init__()
        self.fc1 = nn.Linear(4, 8)
        self.fc2 = nn.Linear(8, 3)
    
    def forward(self, x):
        x = torch.relu(self.fc1(x))
        x = self.fc2(x)
        return x

# Create model and data
model = SimpleModel()

# Create background data for the explainer
background_data = torch.randn(2, 4)

INIT_ARGS = {
    'model': model,
    'data': background_data
}

# For the shap_values method, we need input data X
test_data = torch.randn(2, 4)

ARGS = {
    'X': test_data
}

METHOD = 'shap_values'

# Import the Explainer class (assuming it's defined above)
from distutils.version import LooseVersion
import warnings

class Explainer(object):
    """ This is the superclass of all explainers.
    """

    def shap_values(self, X):
        raise Exception('SHAP values not implemented for this explainer!')

    def attributions(self, X):
        return self.shap_values(X)

op_handler = {}

torch_module = None

class PyTorchDeepExplainer(Explainer):

    def __init__(self, model, data):
        global torch_module
        if torch_module is None:
            torch_module = torch
            if LooseVersion(torch.__version__) < LooseVersion('0.4'):
                warnings.warn('Your PyTorch version is older than 0.4 and not supported.')
        self.multi_input = False
        if type(data) == list:
            self.multi_input = True
        if type(data) != list:
            data = [data]
        self.data = data
        self.layer = None
        self.input_handle = None
        self.interim = False
        self.interim_inputs_shape = None
        self.expected_value = None
        if type(model) == tuple:
            self.interim = True
            model, layer = model
            model = model.eval()
            self.layer = layer
            self.add_target_handle(self.layer)
            with torch.no_grad():
                _ = model(*data)
                interim_inputs = self.layer.target_input
                if type(interim_inputs) is tuple:
                    self.interim_inputs_shape = [i.shape for i in interim_inputs]
                else:
                    self.interim_inputs_shape = [interim_inputs.shape]
            self.target_handle.remove()
            del self.layer.target_input
        self.model = model.eval()
        self.multi_output = False
        self.num_outputs = 1
        with torch.no_grad():
            outputs = model(*data)
            if outputs.shape[1] > 1:
                self.multi_output = True
                self.num_outputs = outputs.shape[1]

    def add_target_handle(self, layer):
        input_handle = layer.register_forward_hook(self.get_target_input)
        self.target_handle = input_handle

    def add_handles(self, model, forward_handle, backward_handle):
        handles_list = []
        for child in model.children():
            if 'nn.modules.container' in str(type(child)):
                handles_list.extend(self.add_handles(child, forward_handle, backward_handle))
            else:
                handles_list.append(child.register_forward_hook(forward_handle))
                handles_list.append(child.register_backward_hook(backward_handle))
        return handles_list

    def remove_attributes(self, model):
        for child in model.children():
            if 'nn.modules.container' in str(type(child)):
                self.remove_attributes(child)
            else:
                try:
                    del child.x
                except AttributeError:
                    pass
                try:
                    del child.y
                except AttributeError:
                    pass

    @staticmethod
    def get_target_input(module, input, output):
        try:
            del module.target_input
        except AttributeError:
            pass
        setattr(module, 'target_input', input)

    @staticmethod
    def add_interim_values(module, input, output):
        try:
            del module.x
        except AttributeError:
            pass
        try:
            del module.y
        except AttributeError:
            pass
        module_type = module.__class__.__name__
        if module_type in op_handler:
            func_name = op_handler[module_type].__name__
            if func_name == 'passthrough':
                pass
            else:
                for i in range(len(input)):
                    if i != 0 and type(output) is tuple:
                        assert input[i] == output[i], 'Only the 0th input may vary!'
                if func_name in ['maxpool', 'nonlinear_1d']:
                    if type(input) is tuple:
                        setattr(module, 'x', torch.nn.Parameter(input[0].detach()))
                    else:
                        setattr(module, 'x', torch.nn.Parameter(input.detach()))
                    if type(output) is tuple:
                        setattr(module, 'y', torch.nn.Parameter(output[0].detach()))
                    else:
                        setattr(module, 'y', torch.nn.Parameter(output.detach()))

    def shap_values(self, X):
        raise Exception('SHAP values not implemented for this explainer!')


# Construct and call
explainer = PyTorchDeepExplainer(**INIT_ARGS)
try:
    result = getattr(explainer, METHOD)(**ARGS)
except Exception as e:
    print(f"Expected exception: {e}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
distutils, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 3, 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
backward_walk_ops Ran
ran; our contract draft was wrong, not the code
shaoshanglqy/shap-shapley/shap/explainers/deep/deep_tf.py
code served (permissive licence) · get_code("174664b8521b7367")
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 torch.nn as nn

torch.manual_seed(0)

class MockTensor:
    def __init__(self, op):
        self.op = op

class MockOp:
    def __init__(self, op_type, inputs=None):
        self.type = op_type
        self.inputs = inputs if inputs is not None else []

# Create a simple computation graph
# op1 has no inputs
op1 = MockOp("conv")
op1.inputs = []

# op2 depends on op1
tensor1 = MockTensor(op1)
op2 = MockOp("relu")
op2.inputs = [tensor1]

# op3 depends on op2
tensor2 = MockTensor(op2)
op3 = MockOp("pool")
op3.inputs = [tensor2]

# op4 is a blacklisted type, depends on op3
tensor3 = MockTensor(op3)
op4 = MockOp("dropout")
op4.inputs = [tensor3]

# op5 depends on op4
tensor4 = MockTensor(op4)
op5 = MockOp("linear")
op5.inputs = [tensor4]

# Create a blacklisted tensor that should stop traversal
blacklisted_tensor = MockTensor(None)
op6 = MockOp("batch_norm")
op6.inputs = [blacklisted_tensor]

ARGS = {
    "start_ops": [op5, op6],
    "tensor_blacklist": [blacklisted_tensor],
    "op_type_blacklist": ["dropout"]
}
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 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
deeplift_grad Ran
ran, violated the contract we drafted
this paper's copy was not recorded; identical code first harvested from heidelberg-nlp/cc-shap
pointer only; the stored copy get_code serves, from heidelberg-nlp/cc-shap, is pointer only (licence: NONE) · get_code("2f372a731d99a85b")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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.

forward_walk_ops Ran
ran; our contract draft was wrong, not the code
shaoshanglqy/shap-shapley/shap/explainers/deep/deep_tf.py
code served (permissive licence) · get_code("556ef4411f7dfff0")
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 torch.nn as nn

torch.manual_seed(0)

class MockTensor:
    def __init__(self, name):
        self.name = name
        self.consumers_list = []
    
    def consumers(self):
        return self.consumers_list
    
    def add_consumer(self, op):
        self.consumers_list.append(op)


class MockOp:
    def __init__(self, op_id, op_type, outputs):
        self.id = op_id
        self.type = op_type
        self.outputs = outputs


# Create mock tensors
tensor_a = MockTensor("tensor_a")
tensor_b = MockTensor("tensor_b")
tensor_c = MockTensor("tensor_c")
tensor_d = MockTensor("tensor_d")

# Create mock ops
op1 = MockOp(1, "Conv2d", [tensor_a, tensor_b])
op2 = MockOp(2, "ReLU", [tensor_c])
op3 = MockOp(3, "Linear", [tensor_d])
op4 = MockOp(4, "BatchNorm", [])

# Set up consumer relationships
tensor_a.add_consumer(op2)
tensor_a.add_consumer(op3)
tensor_b.add_consumer(op4)
tensor_c.add_consumer(op4)

# Define arguments for forward_walk_ops
ARGS = {
    "start_ops": [op1],
    "tensor_blacklist": [tensor_d],
    "op_type_blacklist": ["BatchNorm"],
    "within_ops": [op1, op2, op3, op4]
}
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 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
tensors_blocked_by_false Ran
ran; our contract draft was wrong, not the code
shaoshanglqy/shap-shapley/shap/explainers/deep/deep_tf.py
code served (permissive licence) · get_code("42c14f4de4f2d1ba")
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 torch.nn as nn

torch.manual_seed(0)

class MockTensor:
    def __init__(self, name):
        self.name = name
        self.consumers_list = []
    
    def consumers(self):
        return self.consumers_list

class MockOp:
    def __init__(self, op_type, outputs=None):
        self.type = op_type
        self.outputs = outputs if outputs is not None else []

# Create a simple computation graph structure
# We need ops that can have outputs, and those outputs can have consumers

# Create tensors
tensor1 = MockTensor("tensor1")
tensor2 = MockTensor("tensor2")
tensor3 = MockTensor("tensor3")
tensor4 = MockTensor("tensor4")

# Create ops
# A non-Switch op that will be recursed into
regular_op = MockOp('MatMul', outputs=[tensor1, tensor2])

# A Switch op - its second output should be blocked
switch_op = MockOp('Switch', outputs=[tensor3, tensor4])

# Wire up consumers to create the graph
# tensor1 consumers will lead to switch_op
tensor1.consumers_list = [switch_op]

# tensor2 consumers will lead to another regular op
another_op = MockOp('ReLU', outputs=[])
tensor2.consumers_list = [another_op]

ops = [regular_op]

ARGS = {
    'ops': ops
}
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 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
Explainer Not yet run
unverified (harvested, no recorded run)
shaoshanglqy/shap-shapley/shap/explainers/deep/deep_tf.py
code served (permissive licence) · get_code("6cddc3a37d231206")
How it ran

no invocation recorded.

TFDeepExplainer Not yet run
unverified (harvested, no recorded run)
shaoshanglqy/shap-shapley/shap/explainers/deep/deep_tf.py
code served (permissive licence) · get_code("b802ce233152291a")
How it ran

no invocation recorded.

add_interim_values Not yet run
unverified (harvested, no recorded run)
austinbrown34/shap/shap/explainers/deep/deep_pytorch.py
code served (permissive licence) · get_code("034debadf477812f")
How it ran

no invocation recorded.

break_dependence Not yet run
unverified (harvested, no recorded run)
shaoshanglqy/shap-shapley/shap/explainers/deep/deep_tf.py
code served (permissive licence) · get_code("d30fb55d50e88ecc")
How it ran

no invocation recorded.

deeplift_tensor_grad Not yet run
unverified (harvested, no recorded run)
this paper's copy was not recorded; identical code first harvested from heidelberg-nlp/cc-shap
pointer only; the stored copy get_code serves, from heidelberg-nlp/cc-shap, is pointer only (licence: NONE) · get_code("10adb065b8e203e6")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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. No invocation recorded.

get_target_input Not yet run
unverified (harvested, no recorded run)
austinbrown34/shap/shap/explainers/deep/deep_pytorch.py
code served (permissive licence) · get_code("52628a5c5ffac3aa")
How it ran

no invocation recorded.

passthrough Not yet run
unverified (harvested, no recorded run)
shaoshanglqy/shap-shapley/shap/explainers/deep/deep_tf.py
code served (permissive licence) · get_code("70789f676e5840c3")
How it ran

no invocation recorded.

Also present in other papers' code

Of 15 samples lifted from this paper's linked repositories, 15 carry a function identity and 0 do not. Of the 15 distinct functions with an identity, 3 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 2 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.

FunctionAlso present inEarliest-datedOthers named
Explainer
brick of this paper's assembly: helper
1 other paper
1 dated earlier · 1 only as an assembly brick
1703.01365 (2017-03)
Axiomatic Attribution for Deep Networks
deeplift_grad
brick of this paper's assembly: helper
1 other paper
1 later
2311.07466 (2023-11)
On Measuring Faithfulness or Self-consistency of Natural Language Explanations
deeplift_tensor_grad
brick of this paper's assembly: helper
1 other paper
1 later
2311.07466 (2023-11)
On Measuring Faithfulness or Self-consistency of Natural Language Explanations

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 purported "black box" nature of neural networks is a barrier to adoption in applications where interpretability is essential. Here we present DeepLIFT (Deep Learning Important FeaTures), a method for decomposing the output prediction of a neural network on a specific input by backpropagating the contributions of all neurons in the network to every feature of the input. DeepLIFT compares the activation of each neuron to its 'reference activation' and assigns contribution scores according to the difference. By optionally giving separate consideration to positive and negative contributions, DeepLIFT can also reveal dependencies which are missed by other approaches. Scores can be computed efficiently in a single backward pass. We apply DeepLIFT to models trained on MNIST and simulated genomic data, and show significant advantages over gradient-based methods.

For agents

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

get_harvested_code_for_paper("1704.02685")
get_code_for_paper("1704.02685")
have("1704.02685")

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

curl https://syntology.ai/api/ran/1704.02685.json

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

[![Syntology run record](https://syntology.ai/api/ran/1704.02685.svg)](https://syntology.ai/paper/1704.02685)

Connect an agent — have() is free.