2nd/10_Wiki/Topics/AI_and_ML/Explainable-AI-XAI.md

---
id: wiki-2026-0508-explainable-ai-xai
title: Explainable AI (XAI)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [XAI, interpretability, SHAP, LIME, attention, mechanistic interpretability, saliency]
duplicate_of: none
source_trust_level: A
confidence_score: 0.95
verification_status: applied
tags: [ai, interpretability, xai, shap, lime, attention, mechanistic, transparency]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: SHAP / LIME / Captum / TransformerLens
---

# Explainable AI (XAI)

## 매 한 줄
> **"매 black-box 의 decision 의 understand"**. 매 SHAP / LIME (post-hoc), 매 attention / saliency, 매 mechanistic interpretability (modern). 매 EU AI Act 의 high-risk 의 require. 매 trade-off: 매 accuracy ↔ interpretability (often false dichotomy).

## 매 핵심

### 매 dimension
- **Local vs global**: 매 single prediction vs overall model.
- **Model-specific vs agnostic**: 매 internals vs black-box.
- **Post-hoc vs intrinsic**: 매 after training vs by design.

### 매 method
- **Feature importance**: SHAP, LIME, permutation.
- **Saliency**: Grad-CAM, Integrated Gradients.
- **Attention**: 매 transformer.
- **Counterfactual**: 매 minimal change.
- **Mechanistic**: 매 circuit, SAE, attribution.
- **Concept-based**: TCAV.

### 매 modern (mechanistic)
- **TransformerLens** (Anthropic).
- **Sparse Autoencoders** (SAE).
- **Activation Patching**.
- **Probing**.
- **Anthropic Circuits**: 매 Towards Monosemanticity.

### 매 응용
1. **Healthcare**: 매 diagnostic explain.
2. **Finance**: 매 credit decision.
3. **Justice**: 매 risk score.
4. **Debugging**: 매 model failure.
5. **Compliance**: 매 EU AI Act.
6. **AI safety**: 매 alignment audit.

## 💻 패턴

### SHAP (TreeExplainer)
```python
import shap
import xgboost as xgb

model = xgb.XGBClassifier().fit(X, y)
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_test)
shap.summary_plot(shap_values, X_test)
shap.force_plot(explainer.expected_value, shap_values[0], X_test.iloc[0])
```

### LIME
```python
from lime.lime_tabular import LimeTabularExplainer
explainer = LimeTabularExplainer(X_train, feature_names=cols, class_names=['neg', 'pos'])
exp = explainer.explain_instance(X_test[0], model.predict_proba, num_features=10)
exp.show_in_notebook()
```

### Integrated Gradients (Captum)
```python
import captum
from captum.attr import IntegratedGradients

ig = IntegratedGradients(model)
attributions = ig.attribute(input_tensor, target=label_idx, n_steps=50)
```

### Grad-CAM (vision)
```python
from captum.attr import LayerGradCam
import torchvision.models as models

model = models.resnet50(pretrained=True).eval()
gc = LayerGradCam(model, model.layer4[-1])
attribution = gc.attribute(input_image, target=class_idx)
```

### Attention visualization (transformer)
```python
import torch
def attention_rollout(attentions, layer):
    """매 average heads, multiply layers."""
    A = attentions[layer]  # 매 [B, H, T, T]
    A = A.mean(dim=1)  # 매 average heads
    return A
```

### Counterfactual explanation
```python
def counterfactual(model, x, target_class, max_iter=100):
    """매 minimum change to flip prediction."""
    x_cf = x.clone().detach().requires_grad_(True)
    optim = torch.optim.Adam([x_cf], lr=0.01)
    for _ in range(max_iter):
        pred = model(x_cf)
        if pred.argmax() == target_class: return x_cf
        loss = F.cross_entropy(pred, torch.tensor([target_class])) + 0.1 * (x_cf - x).norm()
        optim.zero_grad(); loss.backward(); optim.step()
    return x_cf
```

### Permutation importance
```python
from sklearn.inspection import permutation_importance
result = permutation_importance(model, X_val, y_val, n_repeats=10, random_state=0)
sorted_idx = result.importances_mean.argsort()[::-1]
for i in sorted_idx[:10]:
    print(f'{cols[i]}: {result.importances_mean[i]:.4f}')
```

### TCAV (concept-based)
```python
def tcav_score(model, layer, concept_examples, random_examples, target_class):
    cav = train_cav(layer.activations(concept_examples), layer.activations(random_examples))
    sensitivities = []
    for x in target_class_examples:
        grad = compute_gradient_at_layer(model, x, target_class, layer)
        sensitivities.append((grad @ cav) > 0)
    return np.mean(sensitivities)
```

### Mechanistic — activation patching
```python
import transformer_lens as tl

model = tl.HookedTransformer.from_pretrained('gpt2')
def patched_forward(prompt_clean, prompt_corrupt, layer):
    _, clean_cache = model.run_with_cache(prompt_clean)
    
    def patch_hook(activation, hook):
        return clean_cache[hook.name]
    
    return model.run_with_hooks(prompt_corrupt, fwd_hooks=[(f'blocks.{layer}.attn.hook_z', patch_hook)])
```

### Sparse Autoencoder (SAE)
```python
class SAE(nn.Module):
    def __init__(self, d_model, d_sae=32768, l1_coef=1e-3):
        super().__init__()
        self.W_enc = nn.Linear(d_model, d_sae)
        self.W_dec = nn.Linear(d_sae, d_model, bias=False)
        self.l1_coef = l1_coef
    
    def forward(self, x):
        z = F.relu(self.W_enc(x))
        x_recon = self.W_dec(z)
        recon_loss = F.mse_loss(x_recon, x)
        sparsity_loss = self.l1_coef * z.abs().sum(-1).mean()
        return x_recon, recon_loss + sparsity_loss
```

### Probing classifier
```python
def probe(activations, labels, layer):
    X = activations[layer].detach().numpy()
    clf = LogisticRegression(max_iter=1000).fit(X, labels)
    return clf.score(X_val, y_val)
```

### Model card explainability
```yaml
explainability:
  method: SHAP TreeExplainer
  global_top_features:
    - credit_score: 0.34
    - debt_to_income: 0.18
    - employment_years: 0.12
  per_decision_explanation: available_in_ui
  counterfactual_offered: yes
```

### LLM explanation prompt
```python
def llm_explain_decision(prediction, features):
    prompt = f"""You are explaining an ML decision. Use ONLY the features given.
Prediction: {prediction}
Features: {features}

Output:
1. Top 3 reasons
2. What change would flip the decision
3. Limitations of this explanation"""
    return llm.generate(prompt)
```

## 매 결정 기준
| 상황 | Method |
|---|---|
| Tabular ML | SHAP TreeExplainer |
| Black-box agnostic | LIME / SHAP KernelExplainer |
| Vision | Grad-CAM / Integrated Gradients |
| NLP | Attention + token attribution |
| LLM internals | TransformerLens / SAE |
| User-facing | Counterfactual + plain language |
| Compliance | Model card + global + local |

**기본값**: 매 SHAP global + local + 매 counterfactual + 매 model card. 매 LLM internals = SAE + activation patching.

## 🔗 Graph
- 부모: [[AI]] · [[Interpretability]]
- 변형: [[SHAP]] · [[LIME]] · [[Mechanistic-Interpretability]]
- 응용: [[Model-Card]] · [[EU-AI-Act]] · [[AI Safety]]
- Adjacent: [[Sparse-Autoencoder]] · [[Ethics & AI]]

## 🤖 LLM 활용
**언제**: 매 high-risk EU. 매 healthcare / finance. 매 model debug.
**언제 X**: 매 low-stakes. 매 explanation 의 misleading 의 risk.

## ❌ 안티패턴
- **Single number trust**: 매 SHAP 의 misuse.
- **Saliency as causal**: 매 correlation only.
- **Attention = explanation**: 매 not always (Jain & Wallace 2019).
- **Post-hoc only**: 매 design 의 ignore.
- **Explanation 의 misleading**: 매 false confidence.

## 🧪 검증 / 중복
- Verified (Lundberg SHAP 2017, Ribeiro LIME 2016, Olah Anthropic 2024).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — XAI methods + 매 SHAP / LIME / IG / counterfactual / SAE / probing code |