File size: 17,431 Bytes
518db7a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 | import torch
import torch.nn as nn
from typing import Optional, Any, cast
from ..utils.sleep import EpisodicBuffer, SleepReplayScheduler
class NeuroscienceEnhancer(nn.Module):
"""
Neuroscience-inspired enhancement layer for SEM V6.
Integrates five biological learning mechanisms:
1. **STDP (Spike-Timing Dependent Plasticity)**: Temporal causality learning
where synaptic strength changes based on relative timing of pre- and
post-synaptic spikes (Bi & Poo, 1998).
2. **Sleep Consolidation**: Offline memory replay during designated "sleep"
phases to strengthen important patterns (Wilson & McNaughton, 1994).
3. **Neuromodulator Dynamics**: Three neuromodulators (ACh, NE, 5-HT)
dynamically adjust learning rates based on context:
- ACh (Acetylcholine): Enhances plasticity during learning
- NE (Norepinephrine): Increases exploration during arousal
- 5-HT (Serotonin): Stabilizes weights during consolidation
(Sara, 2009; Hasselmo, 2006)
4. **Lateral Inhibition**: Winner-take-all competition where highly active
neurons suppress neighbors to enforce sparsity.
5. **Predictive Coding**: Error-driven learning where prediction errors
propagate backward to update representations (Rao & Ballard, 1999).
These mechanisms enhance Module C (ChebyKAN propagator) with biological
realism while maintaining the frozen architecture constraint.
References:
- Bi, G., & Poo, M. (1998). Synaptic modifications in cultured
hippocampal neurons: dependence on spike timing, synaptic strength,
and postsynaptic cell type. Journal of Neuroscience, 18(24), 10464-10472.
- Wilson, M. A., & McNaughton, B. L. (1994). Reactivation of hippocampal
ensemble memories during sleep. Science, 265(5172), 676-679.
- Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual
cortex: a functional interpretation of some extra-classical
receptive-field effects. Nature Neuroscience, 2(1), 79-87.
- Sara, S. J. (2009). The locus coeruleus and noradrenergic modulation
of cognition. Nature Reviews Neuroscience, 10(3), 211-223.
- Hasselmo, M. E. (2006). The role of acetylcholine in learning and
memory. Current Opinion in Neurobiology, 16(6), 710-715.
"""
def __init__(
self,
manifold_dim: int = 16384,
sparsity: float = 0.05,
device: str = "cuda",
enable_stdp: bool = True,
enable_sleep: bool = True,
enable_neuromodulation: bool = True,
enable_lateral_inhibition: bool = True,
enable_predictive_coding: bool = True,
awake_steps: int = 1000,
sleep_steps: int = 100,
sleep_replay_batch: int = 32,
buffer_max_size: int = 1000,
):
"""
Initialize NeuroscienceEnhancer.
Args:
manifold_dim: Dimensionality of the hypergraph manifold (default: 16384)
sparsity: Target sparsity for lateral inhibition (default: 0.05, i.e., 5%)
device: Device for tensor operations ('cuda' or 'cpu')
enable_stdp: Enable STDP learning mechanism
enable_sleep: Enable sleep consolidation system
enable_neuromodulation: Enable neuromodulator dynamics (ACh, NE, 5-HT)
enable_lateral_inhibition: Enable lateral inhibition (k-WTA)
enable_predictive_coding: Enable predictive coding error computation
awake_steps: Number of training steps per awake phase (default: 1000)
sleep_steps: Number of replay steps per sleep phase (default: 100)
sleep_replay_batch: Batch size for sleep replay (default: 32)
buffer_max_size: Maximum size of episodic replay buffer (default: 1000)
"""
super().__init__()
# GPU requirement check (per CLAUDE.md)
assert torch.cuda.is_available(), "GPU required for Module E (per CLAUDE.md)"
self.manifold_dim = manifold_dim
self.sparsity = sparsity
self.k = int(manifold_dim * sparsity) # Number of winners for k-WTA
self.device = torch.device(device)
# Feature flags
self.enable_stdp = enable_stdp
self.enable_sleep = enable_sleep
self.enable_neuromodulation = enable_neuromodulation
self.enable_lateral_inhibition = enable_lateral_inhibition
self.enable_predictive_coding = enable_predictive_coding
# Sleep/wake cycle scheduler (subtask-2-3)
self.sleep_scheduler: Optional[SleepReplayScheduler]
self.episodic_buffer: Optional[EpisodicBuffer]
if self.enable_sleep:
self.sleep_scheduler = SleepReplayScheduler(
awake_steps=awake_steps,
sleep_steps=sleep_steps,
sleep_replay_batch=sleep_replay_batch
)
# Episodic replay buffer (subtask-2-1)
self.episodic_buffer = EpisodicBuffer(
max_size=buffer_max_size,
device=str(self.device)
)
else:
self.sleep_scheduler = None
self.episodic_buffer = None
# Neuromodulator levels (learnable parameters)
# Initialized to biologically plausible baseline values
if self.enable_neuromodulation:
self.ach = nn.Parameter(torch.tensor(1.0, device=self.device)) # Acetylcholine
self.ne = nn.Parameter(torch.tensor(0.5, device=self.device)) # Norepinephrine
self.serotonin = nn.Parameter(torch.tensor(0.3, device=self.device)) # Serotonin
# Placeholder for future components (to be implemented in subsequent subtasks)
# - STDP learner (subtask-1-2) - to be integrated
# - Predictive coding error module (subtask-3-3) - to be implemented
def set_sleep_mode(self, is_sleeping: bool) -> None:
"""
Set sleep/wake mode for the enhancer.
Args:
is_sleeping: True for sleep mode (offline consolidation),
False for awake mode (online learning)
Note:
When using the sleep scheduler, prefer using step() for automatic
sleep/wake transitions instead of manually setting sleep mode.
"""
# Manual override (bypasses scheduler if enabled)
if self.enable_sleep and self.sleep_scheduler is not None:
# Synchronize scheduler state with manual override
self.sleep_scheduler.awake = not is_sleeping
def step(self) -> None:
"""
Advance one step in sleep/wake cycle.
Automatically transitions between awake and sleep modes based on
the configured scheduler. Should be called once per training step.
Example:
>>> enhancer = NeuroscienceEnhancer(enable_sleep=True)
>>> for step in range(10000):
... if enhancer.is_awake():
... # Online training
... loss = train_step(data)
... enhancer.add_episode(episode)
... else:
... # Sleep consolidation
... replay_batch = enhancer.sample_episodes(batch_size=32)
... consolidate(replay_batch)
... enhancer.step()
"""
if self.enable_sleep and self.sleep_scheduler is not None:
self.sleep_scheduler.step()
def is_awake(self) -> bool:
"""
Check if currently in awake (online learning) mode.
Returns:
True if awake, False if sleeping
"""
if self.enable_sleep and self.sleep_scheduler is not None:
return self.sleep_scheduler.is_awake()
return True # Default to awake if sleep disabled
def is_sleeping(self) -> bool:
"""
Check if currently in sleep (offline consolidation) mode.
Returns:
True if sleeping, False if awake
"""
if self.enable_sleep and self.sleep_scheduler is not None:
return self.sleep_scheduler.is_sleeping()
return False # Default to not sleeping if sleep disabled
def add_episode(self, episode: dict[str, Any]) -> None:
"""
Add episode to replay buffer during awake phase.
Args:
episode: Episode dictionary containing 'sdr', 'reward', 'timestamp'
Example:
>>> episode = {
... 'sdr': torch.randn(16384, device='cuda'),
... 'reward': 1.5,
... 'timestamp': 100.0
... }
>>> enhancer.add_episode(episode)
"""
if self.enable_sleep and self.episodic_buffer is not None:
self.episodic_buffer.add(episode)
def sample_episodes(self, batch_size: int, **kwargs: Any) -> list[dict[str, Any]]:
"""
Sample episodes from replay buffer for sleep consolidation.
Args:
batch_size: Number of episodes to sample
**kwargs: Additional arguments passed to buffer.sample()
(e.g., prioritize=True, reverse_temporal=True)
Returns:
List of episode dictionaries
Example:
>>> # Sample for reverse temporal replay during sleep
>>> batch = enhancer.sample_episodes(
... batch_size=32,
... reverse_temporal=True
... )
"""
if self.enable_sleep and self.episodic_buffer is not None:
return self.episodic_buffer.sample(batch_size, **kwargs)
return []
def get_phase_progress(self) -> float:
"""
Get progress through current sleep/wake phase.
Returns:
Progress as fraction [0, 1] (0.0 = phase start, 1.0 = phase end)
"""
if self.enable_sleep and self.sleep_scheduler is not None:
return self.sleep_scheduler.get_phase_progress()
return 0.0
def get_neuromodulator_states(self) -> tuple[float, float, float]:
"""
Get current neuromodulator levels.
Returns:
Tuple of (ACh, NE, Serotonin) levels
"""
if self.enable_neuromodulation:
return (
self.ach.item(),
self.ne.item(),
self.serotonin.item()
)
else:
return (1.0, 0.0, 0.0) # Defaults when neuromodulation disabled
def compute_effective_learning_rate(self, base_lr: float) -> float:
"""
Compute effective learning rate modulated by neuromodulators.
Formula (per spec):
lr_effective = base_lr * ach * (1 + ne) * (1 - 0.5 * serotonin)
Args:
base_lr: Base learning rate from optimizer
Returns:
Effective learning rate after neuromodulator modulation
"""
if not self.enable_neuromodulation:
return base_lr
# Clamp neuromodulators to safe range [0, 2] to prevent instability
ach_clamped = torch.clamp(self.ach, 0.0, 2.0)
ne_clamped = torch.clamp(self.ne, 0.0, 2.0)
serotonin_clamped = torch.clamp(self.serotonin, 0.0, 2.0)
lr_effective = base_lr * ach_clamped * (1 + ne_clamped) * (1 - 0.5 * serotonin_clamped)
return cast(float, lr_effective.item())
def compute_predictive_coding_error(
self,
prediction: torch.Tensor,
target: torch.Tensor,
return_magnitude: bool = False,
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
"""
Compute predictive coding error for error-driven learning.
Implements Rao & Ballard (1999) predictive coding framework where
prediction errors drive learning through hierarchical error propagation.
Error neurons compute the difference between top-down predictions and
bottom-up sensory input, and these errors are used to update
representations at each level of the hierarchy.
Mathematical formulation:
error = target - prediction
error_magnitude = ||error||_2 (L2 norm)
In predictive coding, weights are updated proportional to error magnitude:
Δw ∝ error_magnitude * gradient
This creates a Bayesian inference framework where:
- Higher-level predictions influence lower-level representations
- Prediction errors are minimized through gradient descent
- Hierarchical structure emerges naturally
Args:
prediction: Model's prediction (batch, manifold_dim)
This represents the top-down prediction from higher
hierarchical levels
target: Ground truth target (batch, manifold_dim)
This represents the bottom-up sensory input or
desired output from lower hierarchical levels
return_magnitude: If True, also return L2 norm of error
(useful for monitoring convergence)
Returns:
Tuple of:
- error: Prediction error tensor (batch, manifold_dim)
Sign indicates direction of error (target > pred: +, target < pred: -)
- error_magnitude: L2 norm of error per sample (batch,)
Only returned if return_magnitude=True, else None
Example:
>>> enhancer = NeuroscienceEnhancer(manifold_dim=16384, device='cuda')
>>> prediction = torch.randn(32, 16384, device='cuda')
>>> target = torch.randn(32, 16384, device='cuda')
>>> error, magnitude = enhancer.compute_predictive_coding_error(
... prediction, target, return_magnitude=True
... )
>>> # Use error for weight updates: Δw ∝ error
>>> # Monitor magnitude to verify error reduction over training
Note:
In the hierarchical predictive coding framework:
- Level N+1 predicts activity at Level N
- Error at Level N = actual(N) - predicted(N)
- This error is used to:
1. Update Level N+1's predictions (top-down)
2. Update Level N's representations (bottom-up)
- Iterative minimization of prediction error across hierarchy
implements Bayesian inference
Reference:
Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual
cortex: a functional interpretation of some extra-classical
receptive-field effects. Nature Neuroscience, 2(1), 79-87.
"""
# Compute raw prediction error (target - prediction)
# This represents the surprise signal that drives learning
error = target - prediction
# Optionally compute error magnitude for monitoring convergence
error_magnitude = None
if return_magnitude:
# L2 norm per sample: ||error||_2
# Used to verify that error decreases over training iterations
# (acceptance criterion from spec)
error_magnitude = torch.norm(error, p=2, dim=1)
return error, error_magnitude
def apply_lateral_inhibition(self, activations: torch.Tensor) -> torch.Tensor:
"""
Apply k-Winners-Take-All lateral inhibition (vectorized).
Keeps only the top-k activations, zeros out the rest to enforce sparsity.
Args:
activations: Input activations (batch, manifold_dim)
Returns:
Sparse activations with exactly k active neurons per sample
"""
if not self.enable_lateral_inhibition:
return activations
# Find top-k indices for each sample in batch
_, top_k_indices = torch.topk(activations, self.k, dim=-1)
# Vectorized: set top-k positions to 1 (no Python loop)
sparse_activations = torch.zeros_like(activations)
sparse_activations.scatter_(-1, top_k_indices, 1.0)
return sparse_activations
def forward(
self,
u: torch.Tensor,
target: Optional[torch.Tensor] = None
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
"""
Forward pass through neuroscience enhancement layer.
Args:
u: Input state from Module C propagator (batch, manifold_dim)
target: Optional target for predictive coding error computation
Returns:
Tuple of:
- Enhanced state after neuroscience mechanisms
- Prediction error (if target provided and predictive coding enabled)
"""
enhanced_u = u
prediction_error = None
# Apply lateral inhibition (if enabled and awake)
if self.enable_lateral_inhibition and not self.is_sleeping():
enhanced_u = self.apply_lateral_inhibition(enhanced_u)
# Compute predictive coding error (if enabled and target provided)
if self.enable_predictive_coding and target is not None:
prediction_error, _ = self.compute_predictive_coding_error(
prediction=enhanced_u,
target=target,
return_magnitude=False
)
# Note: STDP updates and sleep consolidation are handled externally
# via callbacks and training loop orchestration (to be implemented)
return enhanced_u, prediction_error
|