File size: 9,984 Bytes
6835659 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 | """
MSCI v2: Learnable Multimodal Semantic Coherence Index
Extends MSCI with:
1. Learned weights optimized from human evaluation data
2. Optional learned projection heads for embedding alignment
3. Uncertainty estimation via ensemble or dropout
"""
from __future__ import annotations
from dataclasses import dataclass
from typing import Dict, List, Optional, Any, Tuple
from pathlib import Path
import json
import numpy as np
from scipy import stats
from scipy.optimize import minimize
from src.coherence.msci import MSCIResult, compute_msci_v0
from src.embeddings.similarity import cosine_similarity
@dataclass
class MSCIv2Weights:
"""Learned weights for MSCI v2."""
w_ti: float # text-image weight
w_ta: float # text-audio weight
w_ia: float # image-audio weight
correlation_with_human: float # Achieved correlation
n_training_samples: int
optimization_method: str = "spearman"
def to_dict(self) -> Dict[str, Any]:
return {
"w_ti": self.w_ti,
"w_ta": self.w_ta,
"w_ia": self.w_ia,
"correlation_with_human": self.correlation_with_human,
"n_training_samples": self.n_training_samples,
"optimization_method": self.optimization_method,
}
@classmethod
def from_dict(cls, data: Dict[str, Any]) -> "MSCIv2Weights":
return cls(**data)
def save(self, path: Path):
"""Save weights to JSON."""
path = Path(path)
path.parent.mkdir(parents=True, exist_ok=True)
with path.open("w") as f:
json.dump(self.to_dict(), f, indent=2)
@classmethod
def load(cls, path: Path) -> "MSCIv2Weights":
"""Load weights from JSON."""
with Path(path).open("r") as f:
return cls.from_dict(json.load(f))
def optimize_msci_weights(
human_scores: List[float],
st_i_scores: List[float],
st_a_scores: List[float],
si_a_scores: List[float],
method: str = "spearman",
constraint: str = "sum_to_one",
) -> MSCIv2Weights:
"""
Optimize MSCI weights to maximize correlation with human scores.
Args:
human_scores: Human coherence ratings (normalized 0-1)
st_i_scores: Text-image similarity scores
st_a_scores: Text-audio similarity scores
si_a_scores: Image-audio similarity scores
method: Correlation method ("spearman" or "pearson")
constraint: Weight constraint ("sum_to_one" or "none")
Returns:
MSCIv2Weights with optimized values
"""
human = np.array(human_scores)
st_i = np.array(st_i_scores)
st_a = np.array(st_a_scores)
si_a = np.array(si_a_scores)
def compute_msci(weights: np.ndarray) -> np.ndarray:
"""Compute MSCI for given weights."""
w_ti, w_ta, w_ia = weights
total = w_ti + w_ta + w_ia
return (w_ti * st_i + w_ta * st_a + w_ia * si_a) / total
def negative_correlation(weights: np.ndarray) -> float:
"""Objective: negative correlation (to minimize)."""
msci = compute_msci(weights)
if method == "spearman":
corr = stats.spearmanr(msci, human).correlation
else:
corr = stats.pearsonr(msci, human).statistic
if np.isnan(corr):
return 0.0 # Handle edge cases
return -corr # Negative because we minimize
# Initial weights
x0 = np.array([0.45, 0.45, 0.10])
# Bounds: each weight in [0, 1]
bounds = [(0.01, 1.0), (0.01, 1.0), (0.01, 1.0)]
# Constraints
constraints = []
if constraint == "sum_to_one":
constraints.append({
"type": "eq",
"fun": lambda w: w.sum() - 1.0
})
# Optimize
result = minimize(
negative_correlation,
x0=x0,
method="SLSQP",
bounds=bounds,
constraints=constraints,
options={"maxiter": 1000},
)
optimal_weights = result.x
final_msci = compute_msci(optimal_weights)
if method == "spearman":
final_corr = stats.spearmanr(final_msci, human).correlation
else:
final_corr = stats.pearsonr(final_msci, human).statistic
return MSCIv2Weights(
w_ti=float(optimal_weights[0]),
w_ta=float(optimal_weights[1]),
w_ia=float(optimal_weights[2]),
correlation_with_human=float(final_corr),
n_training_samples=len(human_scores),
optimization_method=method,
)
def grid_search_weights(
human_scores: List[float],
st_i_scores: List[float],
st_a_scores: List[float],
si_a_scores: List[float],
resolution: int = 20,
) -> Tuple[MSCIv2Weights, Dict[str, Any]]:
"""
Grid search over weight combinations.
More interpretable than optimization, shows sensitivity.
Args:
human_scores: Human ratings
st_i_scores, st_a_scores, si_a_scores: Component similarities
resolution: Grid resolution per dimension
Returns:
Tuple of (best weights, full grid results)
"""
human = np.array(human_scores)
st_i = np.array(st_i_scores)
st_a = np.array(st_a_scores)
si_a = np.array(si_a_scores)
best_corr = -1.0
best_weights = (0.45, 0.45, 0.10)
grid_results = []
# Generate weight combinations that sum to 1
for w_ti in np.linspace(0.1, 0.8, resolution):
for w_ta in np.linspace(0.1, 0.8, resolution):
w_ia = 1.0 - w_ti - w_ta
if w_ia < 0.01 or w_ia > 0.5:
continue
msci = (w_ti * st_i + w_ta * st_a + w_ia * si_a)
corr = stats.spearmanr(msci, human).correlation
grid_results.append({
"w_ti": w_ti,
"w_ta": w_ta,
"w_ia": w_ia,
"correlation": corr,
})
if corr > best_corr:
best_corr = corr
best_weights = (w_ti, w_ta, w_ia)
weights = MSCIv2Weights(
w_ti=float(best_weights[0]),
w_ta=float(best_weights[1]),
w_ia=float(best_weights[2]),
correlation_with_human=float(best_corr),
n_training_samples=len(human_scores),
optimization_method="grid_search",
)
return weights, {"grid": grid_results, "resolution": resolution}
@dataclass
class MSCIv2Result:
"""Result from MSCI v2 computation."""
msci: float
st_i: float
st_a: float
si_a: float
weights: Dict[str, float]
version: str = "v2"
uncertainty: Optional[float] = None
def to_dict(self) -> Dict[str, Any]:
return {
"msci": self.msci,
"st_i": self.st_i,
"st_a": self.st_a,
"si_a": self.si_a,
"weights": self.weights,
"version": self.version,
"uncertainty": self.uncertainty,
}
def compute_msci_v2(
emb_text: np.ndarray,
emb_image: np.ndarray,
emb_audio: np.ndarray,
weights: Optional[MSCIv2Weights] = None,
include_image_audio: bool = True,
) -> MSCIv2Result:
"""
Compute MSCI v2 with learned or custom weights.
Args:
emb_text: Text embedding
emb_image: Image embedding
emb_audio: Audio embedding
weights: Optional learned weights (uses default if None)
include_image_audio: Whether to include image-audio similarity
Returns:
MSCIv2Result with score and metadata
"""
# Use learned weights or defaults
if weights:
w_ti = weights.w_ti
w_ta = weights.w_ta
w_ia = weights.w_ia
else:
# Default v1 weights
w_ti, w_ta, w_ia = 0.45, 0.45, 0.10
# Compute similarities
st_i = cosine_similarity(emb_text, emb_image)
st_a = cosine_similarity(emb_text, emb_audio)
si_a = cosine_similarity(emb_image, emb_audio) if include_image_audio else None
# Compute MSCI
if include_image_audio and si_a is not None:
total = w_ti + w_ta + w_ia
msci = (w_ti * st_i + w_ta * st_a + w_ia * si_a) / total
weight_dict = {"w_ti": w_ti, "w_ta": w_ta, "w_ia": w_ia}
else:
total = w_ti + w_ta
msci = (w_ti * st_i + w_ta * st_a) / total
weight_dict = {"w_ti": w_ti, "w_ta": w_ta}
return MSCIv2Result(
msci=float(round(msci, 4)),
st_i=float(round(st_i, 4)),
st_a=float(round(st_a, 4)),
si_a=float(round(si_a, 4)) if si_a is not None else None,
weights=weight_dict,
version="v2" if weights else "v1",
)
def compare_msci_versions(
human_scores: List[float],
st_i_scores: List[float],
st_a_scores: List[float],
si_a_scores: List[float],
) -> Dict[str, Any]:
"""
Compare MSCI v1 (fixed weights) vs v2 (learned weights).
Args:
human_scores: Human ratings
st_i_scores, st_a_scores, si_a_scores: Component similarities
Returns:
Comparison results
"""
human = np.array(human_scores)
st_i = np.array(st_i_scores)
st_a = np.array(st_a_scores)
si_a = np.array(si_a_scores)
# V1: Fixed weights
msci_v1 = (0.45 * st_i + 0.45 * st_a + 0.10 * si_a)
corr_v1 = stats.spearmanr(msci_v1, human).correlation
# V2: Optimized weights
v2_weights = optimize_msci_weights(
human_scores, st_i_scores, st_a_scores, si_a_scores
)
msci_v2 = (v2_weights.w_ti * st_i + v2_weights.w_ta * st_a + v2_weights.w_ia * si_a)
corr_v2 = stats.spearmanr(msci_v2, human).correlation
improvement = corr_v2 - corr_v1
return {
"v1": {
"weights": {"w_ti": 0.45, "w_ta": 0.45, "w_ia": 0.10},
"correlation": corr_v1,
},
"v2": {
"weights": v2_weights.to_dict(),
"correlation": corr_v2,
},
"improvement": improvement,
"interpretation": (
f"V2 improves correlation by {improvement:.3f}"
if improvement > 0.01 else
"V2 shows minimal improvement over V1"
),
"n_samples": len(human_scores),
}
|