File size: 8,001 Bytes
2c914eb |
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 |
# How SAL Differs
## SAL β RLHF β Safety β Reward
---
## The Confusion
When people first hear about SAL, they often ask:
> "So it's like RLHF but different?"
No.
> "It's a new safety method?"
No.
> "Some kind of reward shaping?"
No.
SAL is fundamentally different from all of these. This document explains why.
---
## SAL vs RLHF
### RLHF (Reinforcement Learning from Human Feedback)
**What it does:**
- Collects human preferences on model outputs
- Trains a reward model on these preferences
- Uses the reward model to fine-tune the base model
- Goal: Make model outputs match human preferences
**Key characteristics:**
- External signal (human feedback)
- Reward-based optimization
- Behavior shaping
- Requires large amounts of human annotation
### SAL (Self-Alignment Learning)
**What it does:**
- Measures internal parameter stability
- Protects stable (emergent) structures
- Adjusts learning rates based on stability
- Goal: Preserve coherence while enabling growth
**Key characteristics:**
- Internal signal (stability measurement)
- No rewards or optimization targets
- Structure preservation
- Requires no human annotation
### Comparison Table
| Aspect | RLHF | SAL |
|--------|------|-----|
| Signal source | External (humans) | Internal (stability) |
| Optimization | Reward maximization | None |
| Goal | Behavior alignment | Coherence preservation |
| Annotation needs | High | None |
| Forgetting risk | High | Low |
---
## SAL vs Safety Training
### Safety Training
**What it does:**
- Identifies harmful outputs
- Trains model to refuse harmful requests
- Constrains output space
- Goal: Prevent harmful behavior
**Key characteristics:**
- Output-focused
- Constraint-based
- Reactive (responds to bad outputs)
- Binary (safe/unsafe)
### SAL
**What it does:**
- Identifies stable parameters
- Protects emergent structures
- Enables continued learning
- Goal: Maintain internal coherence
**Key characteristics:**
- Parameter-focused
- Protection-based
- Proactive (prevents forgetting)
- Continuous (stability spectrum)
### Comparison Table
| Aspect | Safety Training | SAL |
|--------|-----------------|-----|
| Focus | Outputs | Parameters |
| Approach | Constrain | Protect |
| When | After bad output | Before update |
| Measure | Safe/unsafe | Stability score |
| Purpose | Prevent harm | Preserve coherence |
### They're Complementary
SAL and safety training can work together:
- Safety training constrains what the model outputs
- SAL protects how the model learns
You can apply SAL during safety fine-tuning to reduce forgetting of the base model's capabilities.
---
## SAL vs Reward-Based Methods
### Reward-Based Training
**Examples:** RLHF, RLAIF, Constitutional AI, Reward Modeling
**What they do:**
- Define a reward function (explicit or learned)
- Optimize model to maximize reward
- Shape behavior toward desired outcomes
- Goal: High reward = good behavior
**Key characteristics:**
- Optimization-based
- Reward signal required
- Behavior-focused
- Can lead to reward hacking
### SAL
**What it does:**
- No reward function
- No optimization toward external targets
- Measures internal state
- Goal: Stable β overwritten
**Key characteristics:**
- Measurement-based
- No external signal
- Structure-focused
- No hacking possible (nothing to hack)
### Why No Rewards?
Rewards create optimization pressure. Optimization pressure creates:
1. **Reward hacking** β Finding shortcuts that maximize reward without achieving the intended goal
2. **Goodhart's Law** β "When a measure becomes a target, it ceases to be a good measure"
3. **Alignment tax** β Capability loss from constraining the optimization landscape
SAL avoids all of these by not optimizing for anything. It simply:
- Observes what is stable
- Protects what has emerged
- Allows continued learning in volatile regions
---
## SAL vs Regularization
### Regularization Methods
**Examples:** L1/L2 regularization, Dropout, Weight decay, EWC
**What they do:**
- Add penalty terms to loss function
- Constrain weight magnitudes or changes
- Prevent overfitting
- Goal: Generalization
**Key characteristics:**
- Loss-based
- Penalty approach
- Uniform across parameters (mostly)
- Prevents large weights
### SAL
**What it does:**
- No penalties
- No loss modifications
- Measures stability per-parameter
- Goal: Preserve emergence
**Key characteristics:**
- Gradient-based
- Protection approach
- Adaptive per-parameter
- Preserves stable patterns
### EWC Comparison
Elastic Weight Consolidation (EWC) is the closest method to SAL:
| Aspect | EWC | SAL |
|--------|-----|-----|
| Identifies important parameters | Yes (via Fisher information) | Yes (via stability) |
| Protection mechanism | Quadratic penalty in loss | Gradient scaling |
| Requires task boundaries | Yes | No |
| Online learning | Difficult | Natural |
| Computational cost | High (Fisher computation) | Low |
SAL can be seen as a simpler, more general approach that doesn't require:
- Task boundary detection
- Fisher information computation
- Loss function modification
---
## SAL vs Layer Freezing
### Layer Freezing
**What it does:**
- Selects layers to freeze (no updates)
- Other layers train normally
- Binary: frozen or not
- Goal: Preserve early features
**Key characteristics:**
- Layer-level granularity
- Binary decision
- Manual selection
- All-or-nothing
### SAL
**What it does:**
- Analyzes all parameters
- Continuous stability scores
- Automatic detection
- Soft protection (reduced but non-zero gradients)
**Key characteristics:**
- Parameter-level granularity
- Continuous scale
- Automatic
- Gradual protection
### Why Soft Protection?
Hard freezing (zero gradients) prevents any adaptation. But stable doesn't mean perfect. A parameter might be 90% optimal and benefit from small adjustments.
SAL's soft protection allows:
- Stable parameters: small updates (fine-tuning)
- Neutral parameters: moderate updates (adaptation)
- Volatile parameters: large updates (learning)
---
## The Core Difference
All other methods ask: **"How do we get the behavior we want?"**
SAL asks: **"How do we preserve what has emerged while enabling growth?"**
This is a fundamentally different question. It leads to a fundamentally different approach.
| Traditional | SAL |
|-------------|-----|
| Behavior-centric | Structure-centric |
| Output-focused | Parameter-focused |
| External signals | Internal measurement |
| Optimization | Observation |
| Control | Communication |
---
## When to Use SAL
SAL is particularly valuable for:
1. **Continual learning** β Learning new tasks without forgetting old ones
2. **Fine-tuning** β Adapting models while preserving capabilities
3. **Long training runs** β Preventing gradual coherence loss
4. **Multi-task learning** β Balancing between task-specific and shared knowledge
SAL is NOT designed for:
1. **Behavior alignment** β Use RLHF or Constitutional AI
2. **Safety constraints** β Use safety training
3. **Output filtering** β Use classifiers or rules
---
## Combining SAL with Other Methods
SAL can be combined with other approaches:
### SAL + RLHF
Apply SAL during RLHF fine-tuning to reduce capability loss.
### SAL + Safety Training
Apply SAL to preserve base capabilities while adding safety constraints.
### SAL + EWC
Use EWC for task-specific importance, SAL for general stability.
---
## Summary
| Method | What it optimizes | Signal source | SAL equivalent |
|--------|-------------------|---------------|----------------|
| RLHF | Behavior | Human preferences | None (no optimization) |
| Safety | Compliance | Safety labels | None (not about outputs) |
| Reward | Reward function | Reward model | None (no rewards) |
| Regularization | Loss + penalty | Loss function | Stability score |
| Freezing | Selected layers | Manual | Automatic, soft |
**SAL is unique because it optimizes nothing. It observes and protects.**
---
*"Training as dialogue, not control."*
|