File size: 9,490 Bytes
17373f3 | 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 | #!/usr/bin/env python3
"""
The Impossible Classification Test
=====================================
If 9 methods x 4 layers = 36 features can't distinguish visual
from gibberish, the subspace projections contain ZERO content info.
CPU only. Loads data from scaled gibberish GPU checkpoint.
Setup:
!pip install -q scikit-learn scipy xgboost
"""
import json
import numpy as np
from pathlib import Path
from sklearn.ensemble import GradientBoostingClassifier, RandomForestClassifier
from sklearn.svm import SVC
from sklearn.linear_model import LogisticRegression
from sklearn.neural_network import MLPClassifier
from sklearn.model_selection import cross_val_score, StratifiedKFold
from sklearn.metrics import roc_auc_score, classification_report
from sklearn.preprocessing import StandardScaler
from scipy import stats as sp
from google.colab import drive
drive.mount("/content/drive", force_remount=False)
print("=" * 65)
print("The Impossible Classification Test")
print("=" * 65)
# ---- Load data from scaled gibberish checkpoint ----
CHECKPOINT = Path("/content/drive/MyDrive/topohd_scaled_gib/gpu_checkpoint.json")
assert CHECKPOINT.exists(), "Run scaled_gibberish_gpu.py first!"
with open(CHECKPOINT) as f:
raw = json.load(f)
# Parse keys: format is "prompttype|method|layer"
TARGET_LAYERS = [8, 16, 24, 32]
methods = set()
for key in raw:
if key.startswith("_"): continue
parts = key.split("|")
if len(parts) == 3:
methods.add(parts[1])
methods.discard("random")
methods = sorted(methods)
print(f"\n Methods: {methods}")
print(f" Layers: {TARGET_LAYERS}")
print(f" Feature dimensions: {len(methods)} x {len(TARGET_LAYERS)} = {len(methods)*len(TARGET_LAYERS)}")
# ---- Build feature matrix ----
# For each prompt index, build a feature vector from all (method, layer) combinations
feature_names = [f"{m}_L{l}" for m in methods for l in TARGET_LAYERS]
N_FEATURES = len(feature_names)
# Get prompt counts
n_visual = raw.get("_progress_visual", 0)
n_gibberish = raw.get("_progress_gibberish", 0)
print(f" Visual prompts: {n_visual}")
print(f" Gibberish prompts: {n_gibberish}")
N = min(n_visual, n_gibberish)
print(f" Using {N} per class (balanced)")
# Build matrices
X_visual = np.zeros((N, N_FEATURES))
X_gibberish = np.zeros((N, N_FEATURES))
for fi, (m, l) in enumerate([(m, l) for m in methods for l in TARGET_LAYERS]):
v_key = f"visual|{m}|{l}"
g_key = f"gibberish|{m}|{l}"
v_vals = raw.get(v_key, [])
g_vals = raw.get(g_key, [])
for i in range(min(N, len(v_vals))):
X_visual[i, fi] = v_vals[i]
for i in range(min(N, len(g_vals))):
X_gibberish[i, fi] = g_vals[i]
X = np.vstack([X_visual, X_gibberish])
y = np.array([1]*N + [0]*N) # 1=visual, 0=gibberish
# Remove any rows with all zeros (missing data)
valid = X.sum(axis=1) != 0
X = X[valid]
y = y[valid]
print(f" Valid samples: {len(X)} ({sum(y)} visual, {len(y)-sum(y)} gibberish)")
# ---- Train classifiers ----
print(f"\n Training 5 classifiers with 10-fold stratified CV ...")
print(f" (Chance level = 50%)")
print(f"\n {'Classifier':<30} {'Accuracy':>10} {'Std':>8} {'AUROC':>8}")
print(f" {'-'*56}")
cv = StratifiedKFold(n_splits=10, shuffle=True, random_state=42)
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)
classifiers = {
"Logistic Regression": LogisticRegression(max_iter=1000, random_state=42),
"SVM (RBF kernel)": SVC(kernel='rbf', probability=True, random_state=42),
"Random Forest (100 trees)": RandomForestClassifier(n_estimators=100, random_state=42),
"Gradient Boosted Trees": GradientBoostingClassifier(
n_estimators=200, max_depth=4, random_state=42),
"MLP (128-64-32)": MLPClassifier(
hidden_layer_sizes=(128, 64, 32), max_iter=500, random_state=42),
}
all_results = {}
for name, clf in classifiers.items():
# Accuracy
acc_scores = cross_val_score(clf, X_scaled, y, cv=cv, scoring='accuracy')
# AUROC
auc_scores = cross_val_score(clf, X_scaled, y, cv=cv, scoring='roc_auc')
mean_acc = acc_scores.mean()
std_acc = acc_scores.std()
mean_auc = auc_scores.mean()
all_results[name] = dict(accuracy=float(mean_acc), std=float(std_acc),
auroc=float(mean_auc))
marker = " <<<" if mean_acc > 0.55 else ""
print(f" {name:<30} {mean_acc*100:>9.1f}% {std_acc*100:>7.1f}% "
f"{mean_auc:>7.3f}{marker}")
# ---- Statistical test: is the best classifier better than chance? ----
print(f"\n Statistical test: best classifier vs chance (50%)")
best_name = max(all_results, key=lambda k: all_results[k]["accuracy"])
best_acc = all_results[best_name]["accuracy"]
best_std = all_results[best_name]["std"] if "std" in all_results[best_name] else 0
# Re-run best classifier to get per-fold accuracies
best_clf = classifiers[best_name]
fold_accs = cross_val_score(best_clf, X_scaled, y, cv=cv, scoring='accuracy')
# One-sample t-test: is mean accuracy > 0.50?
t_stat, p_val = sp.ttest_1samp(fold_accs, 0.50)
p_one_sided = p_val / 2 if t_stat > 0 else 1.0
print(f" Best: {best_name} ({best_acc*100:.1f}%)")
print(f" Per-fold: {[f'{a*100:.1f}%' for a in fold_accs]}")
print(f" t-test vs 50%: t={t_stat:.3f}, p={p_one_sided:.4f} (one-sided)")
if p_one_sided > 0.05:
print(f" >>> NOT SIGNIFICANT: best classifier ≈ chance <<<")
else:
print(f" >>> SIGNIFICANT: classifier beats chance (but check effect size) <<<")
print(f" Effect: {(best_acc - 0.50)*100:+.1f}pp above chance")
# ---- Feature importance (from gradient boosted) ----
print(f"\n Feature Importance (Gradient Boosted Trees):")
gb = GradientBoostingClassifier(n_estimators=200, max_depth=4, random_state=42)
gb.fit(X_scaled, y)
importances = gb.feature_importances_
# Top 10 features
top_idx = np.argsort(importances)[::-1][:10]
print(f" {'Feature':<25} {'Importance':>12}")
print(f" {'-'*37}")
for idx in top_idx:
print(f" {feature_names[idx]:<25} {importances[idx]:>12.4f}")
# Check if any single feature is useful
print(f"\n Max single-feature importance: {importances.max():.4f}")
print(f" (Uniform = {1/N_FEATURES:.4f})")
if importances.max() < 2/N_FEATURES:
print(f" No feature is more important than chance → no signal exists")
# ---- Also test with factual and math ----
print(f"\n Bonus: Can classifier distinguish visual from factual?")
n_factual = raw.get("_progress_factual", 0)
if n_factual > 0:
X_factual = np.zeros((min(N, n_factual), N_FEATURES))
for fi, (m, l) in enumerate([(m, l) for m in methods for l in TARGET_LAYERS]):
f_key = f"factual|{m}|{l}"
f_vals = raw.get(f_key, [])
for i in range(min(N, len(f_vals))):
X_factual[i, fi] = f_vals[i]
X_vf = np.vstack([X_visual[:min(N, n_factual)], X_factual])
y_vf = np.array([1]*min(N, n_factual) + [0]*min(N, n_factual))
valid_vf = X_vf.sum(axis=1) != 0
X_vf, y_vf = X_vf[valid_vf], y_vf[valid_vf]
X_vf_s = scaler.transform(X_vf)
gb_vf = GradientBoostingClassifier(n_estimators=200, max_depth=4, random_state=42)
vf_scores = cross_val_score(gb_vf, X_vf_s, y_vf, cv=cv, scoring='accuracy')
print(f" Visual vs Factual: {vf_scores.mean()*100:.1f}% ± {vf_scores.std()*100:.1f}%")
print(f"\n Bonus: Can classifier distinguish visual from math?")
n_math = raw.get("_progress_math", 0)
if n_math > 0:
X_math = np.zeros((min(N, n_math), N_FEATURES))
for fi, (m, l) in enumerate([(m, l) for m in methods for l in TARGET_LAYERS]):
mk = f"math|{m}|{l}"
m_vals = raw.get(mk, [])
for i in range(min(N, len(m_vals))):
X_math[i, fi] = m_vals[i]
X_vm = np.vstack([X_visual[:min(N, n_math)], X_math])
y_vm = np.array([1]*min(N, n_math) + [0]*min(N, n_math))
valid_vm = X_vm.sum(axis=1) != 0
X_vm, y_vm = X_vm[valid_vm], y_vm[valid_vm]
X_vm_s = scaler.transform(X_vm)
gb_vm = GradientBoostingClassifier(n_estimators=200, max_depth=4, random_state=42)
vm_scores = cross_val_score(gb_vm, X_vm_s, y_vm, cv=cv, scoring='accuracy')
print(f" Visual vs Math: {vm_scores.mean()*100:.1f}% ± {vm_scores.std()*100:.1f}%")
# ---- Verdict ----
print(f"\n{'='*65}")
print("VERDICT")
print(f"{'='*65}")
all_at_chance = all(r["accuracy"] < 0.55 for r in all_results.values())
if all_at_chance:
print(f"""
>>> ZERO DISCRIMINATIVE INFORMATION <<<
Five classifiers (logistic regression, SVM, random forest,
gradient boosted trees, neural network) trained on the full
{N_FEATURES}-dimensional subspace projection profile
({len(methods)} methods x {len(TARGET_LAYERS)} layers) cannot distinguish
visual prompts from gibberish above chance level.
The 'visual subspace' projections contain ZERO information
about whether the input describes visual content or is
random character sequences. This is not a limitation of
any individual method — it is a fundamental property of
how PCA/SVD extracts directions from transformer hidden states.
""")
else:
above = {k: v for k, v in all_results.items() if v["accuracy"] >= 0.55}
print(f"\n {len(above)} classifiers achieved >55% accuracy.")
print(f" Some weak signal exists in the projection profiles.")
for k, v in above.items():
print(f" {k}: {v['accuracy']*100:.1f}%")
# Save
OUT = Path("/content/drive/MyDrive/topohd_classification")
OUT.mkdir(exist_ok=True, parents=True)
with open(OUT / "classification_results.json", "w") as f:
json.dump(all_results, f, indent=2)
print(f"\n Saved to {OUT}/")
|