Spaces:
Running
Running
File size: 11,063 Bytes
beea5e8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 | """Five-million-sample ReLU and LeakyReLU diagonal NTK experiment."""
import math
import time
import jax
jax.config.update("jax_enable_x64", True)
jax.config.update("jax_platform_name", "cpu")
import jax.numpy as jnp
import numpy as np
WIDTHS = (20, 80)
N_NETWORKS = 5_000_000
N_PROBES = 2
BATCH_SIZE = 1_000
BLOCK_SIZE = 50_000
DEPTH = 4
C_W = 2.0
LEAK = 0.1
SEED = 325081522
INPUTS = np.array(
[
[-0.9895229339599609, -0.5992491841316223],
[-0.17877478897571564, 2.253682851791382],
],
dtype=np.float64,
)
def init_params(key, width):
keys = jax.random.split(key, DEPTH)
shapes = (
(width, INPUTS.shape[1]),
(width, width),
(width, width),
(2, width),
)
return tuple(
jnp.sqrt(C_W) * jax.random.normal(layer_key, shape, dtype=jnp.float64)
for layer_key, shape in zip(keys, shapes, strict=True)
)
def activation(value, alpha):
return jnp.where(value >= 0.0, value, alpha * value)
def forward(params, inputs, alpha):
z = inputs
for layer, weights in enumerate(params):
activations = z if layer == 0 else activation(z, alpha)
z = activations @ weights.T / jnp.sqrt(jnp.float64(activations.shape[-1]))
return z
def tangent(key, params):
keys = jax.random.split(key, len(params))
return tuple(
jax.random.rademacher(probe_key, weights.shape, dtype=jnp.float64)
for probe_key, weights in zip(keys, params, strict=True)
)
def activation_estimate(params, probe_keys, alpha):
def one_probe(probe_key):
_, directional = jax.jvp(
lambda p: forward(p, jnp.asarray(INPUTS), alpha),
(params,),
(tangent(probe_key, params),),
)
return jnp.array(
(
jnp.mean(directional[0] ** 2),
jnp.mean(directional[0] * directional[1]),
)
)
return jnp.mean(jax.vmap(one_probe)(probe_keys), axis=0)
def one_network(key, width):
parameter_key, probe_root = jax.random.split(key)
params = init_params(parameter_key, width)
probe_keys = jax.random.split(probe_root, N_PROBES)
relu = activation_estimate(params, probe_keys, 0.0)
leaky = activation_estimate(params, probe_keys, LEAK)
return jnp.concatenate((relu, leaky))
def relu_moments(q1, q2, covariance):
scale = math.sqrt(q1 * q2)
correlation = float(np.clip(covariance / scale, -1.0, 1.0))
angle = math.acos(correlation)
relu_covariance = scale * (
math.sin(angle) + (math.pi - angle) * math.cos(angle)
) / (2.0 * math.pi)
positive_probability = (math.pi - angle) / (2.0 * math.pi)
return relu_covariance, positive_probability
def activation_moments(q1, q2, covariance, alpha):
positive_covariance, same_sign_probability = relu_moments(q1, q2, covariance)
opposite_covariance, opposite_sign_probability = relu_moments(
q1, q2, -covariance
)
covariance_value = (
(1.0 + alpha**2) * positive_covariance
- 2.0 * alpha * opposite_covariance
)
derivative_value = (
(1.0 + alpha**2) * same_sign_probability
+ 2.0 * alpha * opposite_sign_probability
)
return covariance_value, derivative_value
def infinite_ntk(alpha):
covariance = INPUTS @ INPUTS.T / INPUTS.shape[1]
theta = covariance.copy()
covariance = C_W * covariance
for _ in range(1, DEPTH):
sigma_covariance = np.empty((2, 2), dtype=np.float64)
derivative_covariance = np.empty((2, 2), dtype=np.float64)
for left in range(2):
for right in range(2):
values = activation_moments(
covariance[left, left],
covariance[right, right],
covariance[left, right],
alpha,
)
sigma_covariance[left, right] = values[0]
derivative_covariance[left, right] = values[1]
theta = sigma_covariance + C_W * derivative_covariance * theta
covariance = C_W * sigma_covariance
return theta
def verify_rows(rows):
diagonal_checks = []
offdiagonal_controls = []
for row in rows:
infinite = np.asarray(row["infinite_ntk"])
mean = np.asarray(row["mean"])
standard_error = np.asarray(row["standard_error"])
diagonal_relative_shift = abs(mean[0] - infinite[0]) / abs(infinite[0])
diagonal_relative_99_upper = (
abs(mean[0] - infinite[0]) + 2.576 * standard_error[0]
) / abs(infinite[0])
diagonal_z = abs(mean[0] - infinite[0]) / standard_error[0]
offdiagonal_relative_shift = abs(mean[1] - infinite[1]) / abs(infinite[1])
offdiagonal_z = abs(mean[1] - infinite[1]) / standard_error[1]
row["comparison"] = {
"diagonal_relative_shift": float(diagonal_relative_shift),
"diagonal_relative_99pct_upper": float(diagonal_relative_99_upper),
"diagonal_z": float(diagonal_z),
"offdiagonal_relative_shift": float(offdiagonal_relative_shift),
"offdiagonal_z": float(offdiagonal_z),
}
diagonal_checks.append(diagonal_z <= 4.0 and diagonal_relative_99_upper <= 0.01)
offdiagonal_controls.append(
offdiagonal_z >= 5.0 and offdiagonal_relative_shift >= 0.01
)
checks = {
"all_diagonal_means_within_4_standard_errors": bool(
all(row["comparison"]["diagonal_z"] <= 4.0 for row in rows)
),
"all_diagonal_99pct_intervals_inside_1pct_equivalence_margin": bool(
all(
row["comparison"]["diagonal_relative_99pct_upper"] <= 0.01
for row in rows
)
),
"all_offdiagonal_controls_detect_at_least_1pct_correction_at_5se": bool(
all(offdiagonal_controls)
),
"every_activation_width_pair_passes": bool(all(diagonal_checks)),
}
return {"checks": checks, "passed": all(checks.values())}
def run_five_million_scale():
started = time.perf_counter()
references = {
"ReLU": infinite_ntk(0.0),
"LeakyReLU(alpha=0.1)": infinite_ntk(LEAK),
}
rows = []
for width_index, width in enumerate(WIDTHS):
width_started = time.perf_counter()
seed = SEED + 100_000 * width_index
root_key = jax.random.PRNGKey(seed)
batched = jax.jit(jax.vmap(lambda key: one_network(key, width)))
total_sum = np.zeros(4, dtype=np.float64)
total_sum_of_squares = np.zeros(4, dtype=np.float64)
block_sum = np.zeros(4, dtype=np.float64)
block_count = 0
block_means = []
for lower in range(0, N_NETWORKS, BATCH_SIZE):
upper = min(lower + BATCH_SIZE, N_NETWORKS)
indices = jnp.arange(lower, upper, dtype=jnp.uint32)
keys = jax.vmap(lambda index: jax.random.fold_in(root_key, index))(indices)
samples = np.asarray(batched(keys))
total_sum += samples.sum(axis=0)
total_sum_of_squares += np.square(samples).sum(axis=0)
block_sum += samples.sum(axis=0)
block_count += samples.shape[0]
if block_count == BLOCK_SIZE:
block_means.append((block_sum / block_count).tolist())
block_sum.fill(0.0)
block_count = 0
if upper % 500_000 == 0:
print(
f"CLAIM3_EMPIRICAL_PROGRESS width={width} "
f"networks={upper}/{N_NETWORKS} "
f"seconds={time.perf_counter() - width_started:.1f}",
flush=True,
)
mean = total_sum / N_NETWORKS
variance = (
total_sum_of_squares - N_NETWORKS * np.square(mean)
) / (N_NETWORKS - 1)
standard_deviation = np.sqrt(np.maximum(variance, 0.0))
standard_error = standard_deviation / math.sqrt(N_NETWORKS)
width_runtime_seconds = time.perf_counter() - width_started
for activation_index, name in enumerate(references):
lower = 2 * activation_index
upper = lower + 2
rows.append(
{
"activation": name,
"width": width,
"seed": seed,
"mean": mean[lower:upper].tolist(),
"standard_deviation": standard_deviation[lower:upper].tolist(),
"standard_error": standard_error[lower:upper].tolist(),
"infinite_ntk": [
float(references[name][0, 0]),
float(references[name][0, 1]),
],
"block_means_100_blocks_of_50000": [
block[lower:upper] for block in block_means
],
"width_runtime_seconds": width_runtime_seconds,
}
)
verification = verify_rows(rows)
checks = {
"paper_four_layer_bias_free_architecture": DEPTH == 4,
"paper_weight_variance": C_W == 2.0,
"paper_two_output_trace_average": True,
"exact_five_million_initializations_per_width": N_NETWORKS == 5_000_000,
"all_widths_at_least_20": all(width >= 20 for width in WIDTHS),
"relu_and_leakyrelu_alpha_point_one": LEAK == 0.1,
"one_hundred_raw_block_means_per_activation_width": all(
len(row["block_means_100_blocks_of_50000"]) == 100 for row in rows
),
"statistical_contract_passes": verification["passed"],
}
return {
"claim": (
"At exact five-million-initialization scale, four-layer bias-free "
"ReLU and LeakyReLU diagonal NTK means equal their infinite-width values"
),
"architecture": {
"depth": DEPTH,
"hidden_widths": list(WIDTHS),
"output_width": 2,
"bias": False,
"weight_variance": C_W,
"parameterization": "paper Appendix B standard/book convention",
},
"inputs": INPUTS.tolist(),
"columns": ["theta_00", "theta_01"],
"network_initializations_per_activation_width": N_NETWORKS,
"hutchinson_probes_per_network": N_PROBES,
"estimator_deviation": (
"Unbiased two-probe parameter-space Hutchinson estimator replaces "
"the paper implementation's exact recursive Jacobian trace"
),
"common_random_numbers": (
"ReLU and LeakyReLU share each Gaussian initialization and probe; "
"networks remain iid within each activation"
),
"source_pdf_sha256": {
"relu": "9dcccaca6bf62d5d07ffd24e8b1c346e3e25b07bef099dc85b2f0eb88c187c0a",
"leakyrelu": "85a5e639eda539957ddcf835e734a8d4c37b7dcb10d9c7f6f13379be27fb52bd",
},
"rows": rows,
"verification": verification,
"checks": checks,
"runtime_seconds": time.perf_counter() - started,
"passed": all(checks.values()),
}
|