---
id: wiki-2026-0508-catastrophic-forgetting
title: Catastrophic Forgetting & Continual Learning
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [파괴적 망각, continual learning, lifelong learning, EWC, replay buffer, LoRA, mixture of experts]
duplicate_of: none
source_trust_level: A
confidence_score: 0.93
verification_status: applied
tags: [continual-learning, catastrophic-forgetting, ewc, replay, transfer-learning, lifelong-learning, lora, llm-finetune]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: PyTorch / Avalanche / continual-learning libs
---

# Catastrophic Forgetting

## 📌 한 줄 통찰
> **"매 new task 의 학습 의 매 old 의 destroy"**. 매 NN 의 weight 의 overwrite. 매 lifelong learning 의 fundamental challenge. 매 modern LLM era 의 highly relevant — 매 fine-tune 의 base capability 의 lose.

## 📖 핵심

### 매 mechanism
- 매 SGD 의 모든 weight 의 update.
- 매 same weight 가 매 multiple task 의 store.
- 매 new task 의 gradient 의 old 의 wipe.

### 매 3 approach (Continual Learning)

#### 1. Regularization-based
- **EWC** (Elastic Weight Consolidation): 매 past 의 important weight 의 protect.
- **SI** (Synaptic Intelligence).
- **LwF** (Learning without Forgetting): 매 distillation.

#### 2. Replay-based
- **Experience replay**: 매 old data 의 sample.
- **Generative replay** (DGR): 매 generative model 의 old 의 reconstruct.
- **Reservoir sampling**.

#### 3. Architecture-based
- **Progressive Networks**: 매 column 의 add.
- **PackNet**: 매 weight 의 mask.
- **Dynamic expansion**.

#### 4. Modern (LLM)
- **LoRA**: 매 base 의 frozen + 매 adapter 의 train.
- **Adapter modules**.
- **Mixture of Experts** (MoE).
- **Soft prompt tuning**.

### 매 evaluation metric
- **Average accuracy**: 매 모든 past task.
- **Backward transfer (BWT)**: 매 old task 의 degradation.
- **Forward transfer (FWT)**: 매 new task 의 boost.
- **Forgetting rate**.

### 매 setting
- **Class-incremental**: 매 new class.
- **Task-incremental**: 매 distinct task.
- **Domain-incremental**: 매 same task, 매 new domain.
- **Online**: 매 stream.

### 매 modern LLM 의 응용
1. **Fine-tune drift**: 매 helpful 의 acquire 가, 매 reasoning 의 lose.
2. **Domain adapt**: 매 medical fine-tune 가, 매 general 의 weak.
3. **RLHF**: 매 alignment tax.
4. **Continual pretraining**: 매 new knowledge.

→ 매 LoRA 의 popular reason: 매 base 의 keep.

### 매 lib
- **Avalanche** (PyTorch): 매 best.
- **Continual-AI**: 매 community.
- **Mammoth**.

### 매 biological 의 inspiration
- 매 brain 의 hippocampus 의 fast learning + 매 neocortex 의 consolidation.
- 매 sleep 의 replay 의 role (Bayesian brain).
- 매 modular brain.

## 💻 패턴

### EWC (Elastic Weight Consolidation)
```python
import torch
import torch.nn as nn

class EWC:
    def __init__(self, model, dataset, lambda_=1000):
        self.model = model
        self.lambda_ = lambda_
        self.params = {n: p for n, p in model.named_parameters() if p.requires_grad}
        self.fisher = self._compute_fisher(dataset)
        self.opt_params = {n: p.data.clone() for n, p in self.params.items()}
    
    def _compute_fisher(self, dataset):
        fisher = {n: torch.zeros_like(p) for n, p in self.params.items()}
        self.model.eval()
        for x, y in dataset:
            self.model.zero_grad()
            output = self.model(x)
            loss = F.cross_entropy(output, y)
            loss.backward()
            for n, p in self.params.items():
                fisher[n] += p.grad.data.pow(2) / len(dataset)
        return fisher
    
    def penalty(self):
        loss = 0
        for n, p in self.params.items():
            loss += (self.fisher[n] * (p - self.opt_params[n]).pow(2)).sum()
        return self.lambda_ * loss

# 매 train new task with EWC
ewc = EWC(model, old_task_loader)
for x, y in new_task_loader:
    loss = F.cross_entropy(model(x), y) + ewc.penalty()
    loss.backward()
    optimizer.step()
```

### Experience Replay
```python
class ReplayBuffer:
    def __init__(self, capacity=10000):
        self.buffer = []
        self.capacity = capacity
    
    def add(self, data):
        if len(self.buffer) >= self.capacity:
            # 매 reservoir sampling
            idx = random.randint(0, len(self.buffer))
            if idx < self.capacity:
                self.buffer[idx] = data
        else:
            self.buffer.append(data)
    
    def sample(self, n):
        return random.sample(self.buffer, min(n, len(self.buffer)))

replay = ReplayBuffer()

# 매 train new task + 매 mix replay
for x, y in new_task_loader:
    new_loss = F.cross_entropy(model(x), y)
    
    if replay.buffer:
        replay_batch = replay.sample(BATCH_SIZE // 2)
        rx, ry = collate(replay_batch)
        replay_loss = F.cross_entropy(model(rx), ry)
        loss = new_loss + replay_loss
    else:
        loss = new_loss
    
    loss.backward()
    optimizer.step()
    
    # 매 store new for future
    for x_i, y_i in zip(x, y):
        replay.add((x_i, y_i))
```

### LoRA (modern, LLM-friendly)
```python
from peft import LoraConfig, get_peft_model
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained('meta-llama/Llama-3-8B')

# 매 task 1 의 LoRA
lora_t1 = LoraConfig(r=16, lora_alpha=32, target_modules=['q_proj', 'v_proj'])
model_t1 = get_peft_model(base, lora_t1)
train(model_t1, task1_data)
model_t1.save_pretrained('./lora-task1')

# 매 task 2 — 매 base 의 fresh + 매 다른 LoRA
model_t2 = get_peft_model(base, LoraConfig(r=16, lora_alpha=32, target_modules=['q_proj', 'v_proj']))
train(model_t2, task2_data)
model_t2.save_pretrained('./lora-task2')

# 매 inference 시 의 swap
def serve(prompt, task):
    base.load_adapter(f'./lora-{task}')
    return base.generate(prompt)
```

→ 매 base 의 untouched — 매 forgetting X.

### Generative Replay
```python
def generative_replay(generator, classifier, new_task_loader):
    for x, y in new_task_loader:
        # 매 new task loss
        new_loss = F.cross_entropy(classifier(x), y)
        
        # 매 old replay (generated)
        n_replay = x.size(0)
        replay_x = generator.sample(n_replay)
        replay_y = classifier_old(replay_x).argmax(-1)  # 매 old 의 prediction 의 supervise
        replay_loss = F.cross_entropy(classifier(replay_x), replay_y)
        
        loss = new_loss + 0.5 * replay_loss
        loss.backward()
        optimizer.step()
```

### PackNet (architecture-based)
```python
class PackNet:
    """매 weight 의 task 별 mask."""
    def __init__(self, model, prune_ratio=0.5):
        self.model = model
        self.task_masks = {}  # 매 task → 매 mask
    
    def train_task(self, task_id, loader):
        # 매 train normally
        train(self.model, loader)
        
        # 매 prune low-magnitude weights
        for name, p in self.model.named_parameters():
            threshold = p.abs().quantile(self.prune_ratio)
            mask = p.abs() > threshold
            self.task_masks[(task_id, name)] = mask
            p.data *= mask  # 매 freeze unmasked
    
    def forward_task(self, task_id, x):
        # 매 use only task's mask
        with torch.no_grad():
            for name, p in self.model.named_parameters():
                p.data *= self.task_masks[(task_id, name)]
        return self.model(x)
```

### Continual eval (Avalanche)
```python
from avalanche.benchmarks.classic import SplitMNIST
from avalanche.training import EWC
from avalanche.evaluation.metrics import accuracy_metrics, forgetting_metrics

scenario = SplitMNIST(n_experiences=5)
model = MyModel()
strategy = EWC(model, optimizer, criterion=F.cross_entropy, ewc_lambda=400)

for experience in scenario.train_stream:
    strategy.train(experience)
    results = strategy.eval(scenario.test_stream)
    print(f'Avg accuracy: {results["Top1_Acc_Stream/eval_phase/test_stream/Task000"]}')
```

### LLM fine-tune drift detection
```python
def detect_drift(base_model, finetuned_model, eval_set):
    """매 base capability 의 forgetting 의 measure."""
    base_scores = []
    ft_scores = []
    for example in eval_set:
        base_scores.append(score(base_model, example))
        ft_scores.append(score(finetuned_model, example))
    
    drift = np.mean(base_scores) - np.mean(ft_scores)
    if drift > 0.05:
        log(f'Significant capability loss: {drift:.3f}')
    return drift
```

## 🤔 결정 기준
| 상황 | Approach |
|---|---|
| LLM fine-tune | LoRA / Adapter |
| Class-incremental | EWC + Replay |
| Streaming | Reservoir + Online EWC |
| Few-shot | Prompt tuning |
| Domain shift | Domain-adversarial |
| Strong constraint | Architecture-based (PackNet) |
| General | Replay (best) |

**기본값**: LoRA / Adapter for LLM. Replay + EWC for vision.

## 🔗 Graph
- 부모: [[Continual-Learning]]
- 변형: [[EWC]] · [[Replay-Buffer]]
- 응용: [[LoRA]] · [[Adapter]] · [[Mixture-of-Experts]]
- Adjacent: [[Bayesian-Brain-Hypothesis]] · [[Biological-Intelligence]] · [[Bias vs Variance Trade-off]] · [[Auto-Encoding]]

## 🤖 LLM 활용
**언제**: 매 sequential task. 매 LLM domain adapt. 매 streaming data. 매 lifelong agent.
**언제 X**: 매 single static dataset. 매 IID assumption.

## ❌ 안티패턴
- **Naive fine-tune**: 매 catastrophic forgetting.
- **No EWC / replay**: 매 old task 의 lose.
- **Replay buffer 의 unbounded**: 매 storage 폭발.
- **No drift measurement**: 매 silent capability loss.
- **Same LR for all task**: 매 some 의 dominate.

## 🧪 검증 / 중복
- Verified (Kirkpatrick EWC 2017, Lopez-Paz GEM, Rebuffi iCaRL).
- 신뢰도 A.
- Related: [[LoRA]] · [[Mixture-of-Experts]] · [[Continual-Learning]] · [[Bayesian-Brain-Hypothesis]] · [[Biological-Intelligence]].

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — 3 approach + 매 EWC / replay / LoRA / PackNet code + LLM drift |