SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2203.08120 · ICLR · 2022

Deep Learning without Shortcuts: Shaping the Kernel with Tailored Rectifiers

Aleksandar Botev, Guodong Zhang, James Martens

arXiv · PDF · Open in the Atlas

Code that ran

14 of 24 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; 7 honoured the contract we drafted, 2 violated it and 1 has no contract verdict; 4 failures of our instrument; 10 no recorded run), measured 2026-11-22; 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/2203.08120.json · badge: /api/ran/2203.08120.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
deepmind/kfac_jax canonical (classifier label)
classifier basis not recorded on the link
3 ran · 3 no recorded run
deepmind/dks alias (classifier label)
classifier basis: classifier_2025
11 ran · 7 no recorded run
FunctionStatusWhere it lives
_bentid Ran
ran on a synthesized input
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("b803f9baa24bfd4f")
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 (4,) 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.

from autograd import numpy as np

def _bentid(x):
    return (np.sqrt(x ** 2 + 1.0) - 1.0) / 2.0 + x

# Create a small numpy array as input
x = np.array([0.5, 1.0, -0.5, 2.0], dtype=float)

ARGS = {
    "x": x
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
autograd — 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 (4,) — shape (4,) 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
_calc_c_curv Ran
ran, honoured the contract we drafted
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("4a9bd4aba951f2ee")
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 numpy as np

def simple_activation(x):
    """Simple smooth activation function (ReLU-like but differentiable)."""
    return np.maximum(0.1 * x, x)

def smooth_activation(x):
    """Smooth activation: tanh."""
    return np.tanh(x)

# Pre-compute q_output for efficiency
def compute_q_output(activation):
    """Estimate q_output value for the activation function."""
    from scipy.integrate import fixed_quad
    from scipy.special import roots_legendre
    fn_weighted = lambda x: np.exp(-x ** 2 / 2) * (activation(x) ** 2)
    integral, _ = fixed_quad(fn_weighted, -10.0, 10.0, n=100)
    return integral / np.sqrt(2 * np.pi)

q_output_value = compute_q_output(smooth_activation)

ARGS = {
    'activation': smooth_activation,
    'c': 0.5,
    'q_output': q_output_value,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
absl, autograd, numpy, scipy — 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
_calc_c_slope Ran
ran, honoured the contract we drafted
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("af6685abaca2e383")
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 numpy as np
from autograd import numpy as anp
from autograd import elementwise_grad as egrad

# Define a simple activation function for testing
def relu_activation(x):
    return anp.maximum(0.0, x)

def tanh_activation(x):
    return anp.tanh(x)

# Mock the cache object that the function expects
class MockRootsCache:
    def __init__(self):
        self.cache = {}

# Patch the module-level cache (simulating what would exist in the real module)
_cached_roots_legendre = MockRootsCache()

# Pre-compute q_output for the activation function
# q_output = E[activation(x)^2] where x ~ N(0,1)
def estimate_gaussian_mean_simple(fn):
    """Simple estimation using quadrature."""
    from scipy import integrate as sp_int
    fn_weighted = lambda x: np.exp(-x ** 2 / 2) * fn(x)
    integral, _ = sp_int.quad(fn_weighted, -10.0, 10.0)
    return integral / np.sqrt(2 * np.pi)

q_output_relu = estimate_gaussian_mean_simple(lambda x: relu_activation(x) ** 2)
q_output_tanh = estimate_gaussian_mean_simple(lambda x: tanh_activation(x) ** 2)

# Construct the ARGS dictionary with all required parameters for _calc_c_slope
ARGS = {
    'activation': tanh_activation,
    'c': 0.5,
    'q_output': q_output_tanh
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
absl, autograd, numpy, scipy — 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
_elu Ran
ran, violated the contract we drafted
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("66703fda3e8cfbe3")
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 (4,) 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.

from autograd import numpy as np

# The function _elu takes three parameters: x, a, and l
# x is the input array/tensor
# a is the alpha parameter (default 1.0)
# l is the lambda/scale parameter (default 1.0)

# Create a small input array with both negative and positive values
x = np.array([-2.0, -1.0, 0.5, 1.5])

# Set alpha parameter
a = 1.0

# Set lambda/scale parameter
l = 1.0

ARGS = {
    "x": x,
    "a": a,
    "l": l,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
autograd — 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 (4,) — shape (4,) 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
_estimate_2d_gaussian_mean Ran
ran, honoured the contract we drafted
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("b4a306c2c7f07155")
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 float, ndim 0, dtype family float, shape () 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.

from autograd import numpy as np
import scipy.integrate as sp_int

# Define a simple function that takes x and y as arguments
# The function signature expects fn to be callable with two arguments
def sample_fn(x, y):
    """A simple test function: returns x^2 + y^2"""
    return x**2 + y**2

ARGS = {
    "fn": sample_fn
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
autograd, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () 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
_gelu Ran
ran, honoured the contract we drafted
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("c8c394e2c6a1a2ce")
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 (2, 4) 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.

from autograd import numpy as np

# The function _gelu takes a single parameter x
# It performs mathematical operations on x: tanh, sqrt, multiplication, power
# x should be a numpy array (since the function uses autograd.numpy)

# Create a small test input - a simple numpy array
x = np.array([[-2.0, -1.0, 0.0, 1.0],
              [2.0, 0.5, -0.5, 1.5]])

ARGS = {
    'x': x
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
autograd — 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 (2, 4) — shape (2, 4) 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
_gelu_exact Ran
ran, honoured the contract we drafted
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("45c504b4c4b5cb39")
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 (2, 4) 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
from autograd import numpy as anp
from autograd import scipy as sp

# Create a small numpy array to test _gelu_exact
# The function takes x and applies: x * (erf(x / sqrt(2)) + 1) / 2
x = anp.array([[-1.0, 0.0, 0.5, 2.0],
               [1.0, -0.5, 1.5, 0.1]], dtype=anp.float32)

ARGS = {
    "x": x
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
autograd, 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 (2, 4) — shape (2, 4) 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_roots_cache_file Ran
ran; our contract draft was wrong, not the code
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("fecba24c85a8383c")
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 str. 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 os

def _get_roots_cache_file(order):
    """Returns the expected path to the roots cache file."""
    current_dir = os.path.dirname(os.path.abspath(__file__))
    filename = os.path.join(current_dir, f'roots_{order}.npy')
    return filename

ARGS = {
    'order': 5
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
_softplus Ran
ran, honoured the contract we drafted
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("8e18417d18243196")
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 (2, 4) 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

# The function _softplus takes a single parameter 'x' which should be a numpy array
# The function applies numerical operations: log, exp, abs, and maximum
# So x should be a numeric array that supports these operations

x = np.array([[-2.0, -1.0, 0.5, 2.0],
              [1.5, -0.5, 3.0, -1.5]], dtype=np.float32)

ARGS = {
    'x': x
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
autograd, 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 (2, 4) — shape (2, 4) 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
_swish Ran
ran, honoured the contract we drafted
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("2c59e2c9e3970f5a")
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 (2, 4) 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
from autograd import scipy as sp

torch.manual_seed(0)

_sigmoid = sp.special.expit

def _swish(x):
    return x * _sigmoid(x)

# Create a small numpy array as input
# _swish takes a single parameter x and applies the swish activation
x_input = np.array([[-1.0, 0.0, 1.0, 2.0],
                     [0.5, -0.5, 1.5, -1.5]], dtype=np.float32)

ARGS = {
    'x': x_input
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
autograd, 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 2, dtype family float, shape (2, 4) — shape (2, 4) 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
_transform_activation Ran
ran; our contract draft was wrong, not the code
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("5afc010dc0e842f8")
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 function. 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)

# Define a simple activation function that takes x and optional keyword arguments
def simple_relu(x, **kwargs):
    return torch.relu(x)

def simple_sigmoid(x, **kwargs):
    return torch.sigmoid(x)

# Create test parameters dictionary with transformation parameters
params = {
    'input_scale': 2.0,
    'input_shift': 0.5,
    'output_shift': 1.0,
    'output_scale': 0.5
}

ARGS = {
    'phi': simple_relu,
    'params': params
}
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 function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
broadcast_in_dim_equivalent Ran
ran, violated the contract we drafted
deepmind/kfac_jax/kfac_jax/_src/tag_graph_matcher.py
code served (permissive licence) · get_code("987e113518340066")
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 bool, ndim 0, dtype family float, shape () 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 jax
import jax.numpy as jnp
import jax.extend as jex
from jax import core

# Create minimal JaxprEqn objects for broadcast_in_dim primitive
def create_broadcast_in_dim_eqn():
    """Create a minimal JaxprEqn for broadcast_in_dim primitive."""
    # Get the broadcast_in_dim primitive
    broadcast_in_dim_prim = core.Primitive.new_name('broadcast_in_dim')
    
    # Create input and output abstractions
    input_shape = (2, 3)
    output_shape = (2, 3, 4)
    
    # Create abstract values
    in_aval = core.ShapedArray(input_shape, jnp.float32)
    out_aval = core.ShapedArray(output_shape, jnp.float32)
    
    # Create variables for the equation
    in_var = core.Var(in_aval)
    out_var = core.Var(out_aval)
    
    # Create JaxprEqn with broadcast_in_dim primitive
    eqn = core.JaxprEqn(
        primitive=core.get_aval(jnp.broadcast_to(jnp.zeros(input_shape), output_shape)).__class__.__bases__[0],
        invars=[in_var],
        outvars=[out_var],
        params={}
    )
    
    return eqn

# Simpler approach: directly access existing primitives and create equations
def create_simple_jaxpr_eqn():
    """Create a JaxprEqn using actual JAX primitives."""
    # Use the actual broadcast_in_dim primitive from JAX
    from jax.interpreters import xla
    from jax.lib import xla_client
    
    # Create abstract input/output values
    in_aval = core.ShapedArray((2, 3), jnp.float32)
    out_aval = core.ShapedArray((2, 3, 4), jnp.float32)
    
    in_var = core.Var(in_aval)
    out_var = core.Var(out_aval)
    
    # Get the actual broadcast_in_dim primitive
    from jax.experimental.key_reuse import _check_key_reuse_count
    from jax._src.dtypes import preinit
    from jax.experimental import io_callback
    
    # Access primitive directly through core module
    prim = core.Primitive.new_name('broadcast_in_dim')
    prim.multiple_results = False
    
    eqn = core.JaxprEqn(
        primitive=prim,
        invars=[in_var],
        outvars=[out_var],
        params={'shape': (2, 3, 4), 'broadcast_dimensions': (0, 1)}
    )
    
    return eqn

# Most direct approach: construct minimal mock objects
class MockPrimitive:
    def __init__(self, name):
        self.name = name

class MockJaxprEqn:
    def __init__(self, prim_name):
        self.primitive = MockPrimitive(prim_name)

# Create the arguments
equation1 = MockJaxprEqn('broadcast_in_dim')
equation2 = MockJaxprEqn('broadcast_in_dim')

ARGS = {
    'equation1': equation1,
    'equation2': equation2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () 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
compatible_sum Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
deepmind/kfac_jax/kfac_jax/_src/curvature_blocks/utils.py
code served (permissive licence) · get_code("dc1b986b1be917d4")
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 ArrayImpl, ndim 4, dtype family float, shape (2, 1, 3, 5) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax.numpy as jnp

# Create test tensors and parameters for compatible_sum function
tensor = jnp.ones((2, 4, 3, 5))  # 4D tensor with distinct dimensions
target_shape = (1, 3, 5)  # Target shape compatible with tensor
skip_axes = [0]  # Axes to skip during summation

ARGS = {
    "tensor": tensor,
    "target_shape": target_shape,
    "skip_axes": skip_axes
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 4, dtype family float, shape (2, 1, 3, 5) — shape (2, 1, 3, 5) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
eval_jaxpr_eqn Ran
ran; our contract draft was wrong, not the code
deepmind/kfac_jax/kfac_jax/_src/tag_graph_matcher.py
code served (permissive licence) · get_code("42474a82097165d1")
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 2, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
from jax import core
from jax.extend import core as jex_core
import jax.extend as jex

# Create a minimal Jaxpr equation for testing
# We'll use a simple primitive operation like addition

# Get the add primitive
add_prim = jax.lax.add_p

# Create minimal source info
class MockSourceInfo:
    def __init__(self):
        self.traceback = None

# Create a minimal JaxprEqn-like object
class MockJaxprEqn:
    def __init__(self):
        self.primitive = add_prim
        self.params = {}
        self.source_info = MockSourceInfo()
        self.invars = []
        self.outvars = []

# Construct the equation
eqn = MockJaxprEqn()

# Create input values - simple JAX arrays for addition
in_values = [
    jnp.array([1.0, 2.0, 3.0, 4.0]),
    jnp.array([5.0, 6.0, 7.0, 8.0])
]

ARGS = {
    'eqn': eqn,
    'in_values': in_values
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 2, 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
_calc_c_map Not yet run
unverified (harvested, no recorded run)
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("3e9f86b459eecc31")
How it ran

no invocation recorded.

_calc_q_slope_1 Not yet run
unverified (harvested, no recorded run)
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("95c482d63878f89f")
How it ran

no invocation recorded.

_compute_output_params Not yet run
unverified (harvested, no recorded run)
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("518f8f86235bcdf2")
How it ran

no invocation recorded.

_estimate_gaussian_mean Not yet run
unverified (harvested, no recorded run)
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("17cfc190249b5d4f")
How it ran

no invocation recorded.

_leaky_relu Not yet run
unverified (harvested, no recorded run)
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("8bd5c8a0631e8f17")
How it ran

no invocation recorded.

_precompute_or_load_roots Not yet run
unverified (harvested, no recorded run)
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("9c3d8825878c6315")
How it ran

no invocation recorded.

_solve_for_activation_params Not yet run
unverified (harvested, no recorded run)
deepmind/dks/dks/base/activation_transform.py
code served (permissive licence) · get_code("95b55a581b4926c8")
How it ran

no invocation recorded.

get_and_verify_loss_meta Not yet run
unverified (harvested, no recorded run)
deepmind/kfac_jax/kfac_jax/_src/layers_and_loss_tags.py
code served (permissive licence) · get_code("14e10e479569a361")
How it ran

no invocation recorded.

get_loss_outputs Not yet run
unverified (harvested, no recorded run)
deepmind/kfac_jax/kfac_jax/_src/layers_and_loss_tags.py
code served (permissive licence) · get_code("b4a5e9703e6605e1")
How it ran

no invocation recorded.

reshape_equivalent Not yet run
unverified (harvested, no recorded run)
deepmind/kfac_jax/kfac_jax/_src/tag_graph_matcher.py
code served (permissive licence) · get_code("ee339b16cfba7ff0")
How it ran

no invocation recorded.

Also present in other papers' code

Of 24 samples lifted from this paper's linked repositories, 24 carry a function identity and 0 do not. Of the 24 distinct functions with an identity, 6 are also present in other papers whose code we hold (6 with at least one paper dated earlier, 0 the same month, 6 later) and 18 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
broadcast_in_dim_equivalent 2 other papers
1 dated earlier, 1 later
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2605.26248 (2026-05)
Unified Neural Scaling Laws
compatible_sum 2 other papers
1 dated earlier, 1 later
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2605.26248 (2026-05)
Unified Neural Scaling Laws
eval_jaxpr_eqn 2 other papers
1 dated earlier, 1 later
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2605.26248 (2026-05)
Unified Neural Scaling Laws
get_and_verify_loss_meta 2 other papers
1 dated earlier, 1 later
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2605.26248 (2026-05)
Unified Neural Scaling Laws
get_loss_outputs 2 other papers
1 dated earlier, 1 later
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2605.26248 (2026-05)
Unified Neural Scaling Laws
reshape_equivalent 2 other papers
1 dated earlier, 1 later
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2605.26248 (2026-05)
Unified Neural Scaling Laws

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

Training very deep neural networks is still an extremely challenging task. The common solution is to use shortcut connections and normalization layers, which are both crucial ingredients in the popular ResNet architecture. However, there is strong evidence to suggest that ResNets behave more like ensembles of shallower networks than truly deep ones. Recently, it was shown that deep vanilla networks (i.e. networks without normalization layers or shortcut connections) can be trained as fast as ResNets by applying certain transformations to their activation functions. However, this method (called Deep Kernel Shaping) isn't fully compatible with ReLUs, and produces networks that overfit significantly more than ResNets on ImageNet. In this work, we rectify this situation by developing a new type of transformation that is fully compatible with a variant of ReLUs -Leaky ReLUs. We show in experiments that our method, which introduces negligible extra computational cost, achieves validation accuracies with deep vanilla networks that are competitive with ResNets (of the same width/depth), and significantly higher than those obtained with the Edge of Chaos (EOC) method. And unlike with EOC, the validation accuracies we obtain do not get worse with depth.

For agents

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

get_harvested_code_for_paper("2203.08120")
get_code_for_paper("2203.08120")
have("2203.08120")

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

curl https://syntology.ai/api/ran/2203.08120.json

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

[![Syntology run record](https://syntology.ai/api/ran/2203.08120.svg)](https://syntology.ai/paper/2203.08120)

Connect an agent — have() is free.