File size: 22,350 Bytes
ebaf2ce | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 | """
MLE Inference Module: Reasoning Engine
========================================
The complete reasoning pipeline that integrates all modules:
Query β Encoding β Routing β Binding β Energy Minimization β Response
The "chain of thought" is an internal trajectory through state space:
1. ENCODE: Transform input query into a binary vector
2. ROUTE: Activate relevant memory regions (top-500 beam)
3. BIND: Construct composite representations from activated memories
4. RELAX: Minimize energy to find coherent state
5. DECODE: Extract response from final state
Multiple reasoning iterations are possible (iterative deepening):
each iteration uses the previous result as context for the next.
"""
import numpy as np
from typing import Optional, List, Dict, Any, Tuple
import logging
import time
import hashlib
from ..memory.sparse_address_table import SparseAddressTable
from ..routing.recursive_jit_router import RecursiveJITRouter, RoutingResult
from ..binding.semantic_binding import BindingEngine, BinaryBinding
from ..energy.energy_model import EnergyModel, EnergyFunction, RelaxationDynamics, HopfieldDynamics
from ..utils.simd_ops import (
N_BITS, N_WORDS,
random_binary_vector, random_binary_vectors,
hamming_distance, hamming_similarity,
xor_vectors, majority_vote
)
logger = logging.getLogger(__name__)
class QueryEncoder:
"""
Encode text/symbolic queries into binary vectors.
Encoding strategies:
1. Named concept lookup (exact match in codebook)
2. Hash-based encoding (deterministic binary vector from string)
3. Composite encoding (bind multiple concepts together)
"""
def __init__(self, binding_engine: BindingEngine):
self.binding = binding_engine
def encode(self, query: Any) -> np.ndarray:
"""Encode a query into a binary vector.
Supports:
- str: hash-based or codebook lookup
- np.ndarray: pass-through (already a vector)
- list of str: composite binding of concepts
- dict: role-filler structure encoding
"""
if isinstance(query, np.ndarray):
return query.copy()
elif isinstance(query, str):
return self._encode_string(query)
elif isinstance(query, (list, tuple)):
return self._encode_composite(query)
elif isinstance(query, dict):
return self._encode_structure(query)
else:
raise TypeError(f"Cannot encode query of type {type(query)}")
def _encode_string(self, text: str) -> np.ndarray:
"""Encode a string to binary vector.
First tries codebook lookup, then falls back to hash.
"""
# Try codebook lookup
vec = self.binding.get_concept(text)
if vec is not None:
return vec.copy()
# Hash-based encoding: deterministic binary vector from string
return self._hash_to_vector(text)
def _hash_to_vector(self, text: str) -> np.ndarray:
"""Deterministic hash-based encoding.
Uses SHA-512 repeatedly to fill 4096 bits.
"""
n_hashes = N_BITS // 512 + 1 # SHA-512 = 512 bits
bits = []
for i in range(n_hashes):
h = hashlib.sha512(f"{text}:{i}".encode()).digest()
bits.append(np.frombuffer(h, dtype=np.uint8))
all_bytes = np.concatenate(bits)[:N_BITS // 8]
return all_bytes.view(np.uint64).copy()
def _encode_composite(self, concepts: List[str]) -> np.ndarray:
"""Encode a composite query by bundling encoded concepts."""
vectors = [self._encode_string(c) for c in concepts]
if len(vectors) == 1:
return vectors[0]
return BinaryBinding.bundle(*vectors)
def _encode_structure(self, structure: Dict[str, str]) -> np.ndarray:
"""Encode a structured query (role:filler pairs).
Example: {"subject": "cat", "relation": "is_a"} β
bind(encode("subject"), encode("cat")) β bind(encode("relation"), encode("is_a"))
"""
pairs = []
for role, filler in structure.items():
role_vec = self._encode_string(role)
filler_vec = self._encode_string(filler)
pairs.append(BinaryBinding.bind(role_vec, filler_vec))
return BinaryBinding.bundle(*pairs) if len(pairs) > 1 else pairs[0]
class ResponseDecoder:
"""
Decode final state vector into interpretable response.
Strategies:
1. Nearest neighbor: find closest named concept(s) in codebook
2. Unbinding: recover specific fillers for given roles
3. Activation readout: return most-activated memory entries
"""
def __init__(self, memory: SparseAddressTable, binding_engine: BindingEngine):
self.memory = memory
self.binding = binding_engine
def decode(self, state: np.ndarray, top_k: int = 5,
roles: Optional[List[str]] = None) -> Dict[str, Any]:
"""Decode state vector into response.
Returns dict with:
- nearest_concepts: closest named concepts with similarities
- nearest_memories: closest memory entries
- role_fillers: decoded role-filler pairs (if roles provided)
- confidence: overall decoding confidence
"""
result = {}
# 1. Nearest concepts in codebook
concept_scores = []
for name, vec in self.binding._codebook.items():
sim = hamming_similarity(state, vec)
concept_scores.append((name, float(sim)))
concept_scores.sort(key=lambda x: x[1], reverse=True)
result['nearest_concepts'] = concept_scores[:top_k]
# 2. Nearest memory entries
memory_results = self.memory.query_nearest(state, k=top_k, use_lsh=True)
nearest_memories = []
for idx, dist in memory_results:
meta = self.memory._metadata[idx]
name = meta.get('name', f'entry_{idx}') if meta else f'entry_{idx}'
sim = 1.0 - dist / N_BITS
nearest_memories.append({
'index': idx,
'name': name,
'distance': dist,
'similarity': sim,
'metadata': meta,
})
result['nearest_memories'] = nearest_memories
# 3. Role-filler decoding
if roles:
role_fillers = {}
for role_name in roles:
role_vec = self.binding.get_concept(role_name)
if role_vec is None:
role_vec = QueryEncoder(self.binding)._hash_to_vector(role_name)
filler = BinaryBinding.unbind(state, role_vec)
# Find nearest concept to recovered filler
filler_scores = []
for name, vec in self.binding._codebook.items():
sim = hamming_similarity(filler, vec)
filler_scores.append((name, float(sim)))
filler_scores.sort(key=lambda x: x[1], reverse=True)
role_fillers[role_name] = filler_scores[:3]
result['role_fillers'] = role_fillers
# 4. Confidence (based on how close the best match is)
if concept_scores:
result['confidence'] = concept_scores[0][1]
else:
result['confidence'] = 0.0
return result
class ReasoningStep:
"""Record of a single reasoning step in the inference chain."""
__slots__ = ['step_num', 'state', 'energy', 'energy_components',
'routing_result', 'decoded', 'timestamp']
def __init__(self, step_num: int):
self.step_num = step_num
self.state: Optional[np.ndarray] = None
self.energy: float = float('inf')
self.energy_components: Dict[str, float] = {}
self.routing_result: Optional[RoutingResult] = None
self.decoded: Optional[Dict] = None
self.timestamp: float = 0.0
class ReasoningEngine:
"""
The MLE Reasoning Engine β core inference loop.
Architecture:
ββββββββββββ ββββββββββββ ββββββββββββ ββββββββββββ
β Query ββββββΆβ Router ββββββΆβ Binder ββββββΆβ Energy β
β Encoder β β (Beam) β β (Compose)β β (Relax) β
ββββββββββββ ββββββββββββ ββββββββββββ ββββββββββββ
β β
β ββββββββββββ ββββββββββββ β
βββββββββββββ Response βββββββ Decode βββββββββββββ
β β β β
ββββββββββββ ββββββββββββ
Reasoning loop:
1. Encode query β binary vector
2. Route β find relevant memories (beam search)
3. Activate memories, construct bindings
4. Minimize energy (Hopfield + binary relaxation)
5. Decode final state β response
6. [Optional] Use result as new query β iterate
"""
def __init__(
self,
memory: Optional[SparseAddressTable] = None,
beam_width: int = 500,
max_routing_depth: int = 3,
max_reasoning_steps: int = 5,
energy_mode: str = 'hybrid',
hopfield_beta: float = 8.0,
relaxation_iterations: int = 50,
):
# Memory
self.memory = memory or SparseAddressTable()
# Binding engine
self.binding = BindingEngine(use_binary=True)
# Router
self.router = RecursiveJITRouter(
memory=self.memory,
beam_width=beam_width,
max_depth=max_routing_depth,
)
# Energy model
energy_fn = EnergyFunction()
self.energy_model = EnergyModel(
mode=energy_mode,
energy_fn=energy_fn,
binary_dynamics=RelaxationDynamics(
energy_fn,
max_iterations=relaxation_iterations,
n_candidates=32,
),
hopfield_dynamics=HopfieldDynamics(
beta=hopfield_beta,
max_iterations=20,
),
)
# Encoder/Decoder
self.encoder = QueryEncoder(self.binding)
self.decoder = ResponseDecoder(self.memory, self.binding)
# Config
self.max_reasoning_steps = max_reasoning_steps
# ββ Knowledge Management ββββββββββββββββββββββββββββββββββββββββββββββ
def add_concept(self, name: str, metadata: Optional[Dict] = None) -> np.ndarray:
"""Add a named concept to both memory and binding codebook."""
vec = self.binding.register_concept(name)
self.memory.store_concept(name, content=vec, metadata=metadata)
return vec
def add_relation(self, subject: str, relation: str, obj: str,
metadata: Optional[Dict] = None) -> int:
"""Store a semantic relation as a bound vector in memory."""
# Ensure all concepts exist
for name in [subject, relation, obj]:
if self.binding.get_concept(name) is None:
self.add_concept(name)
# Create bound representation
s = self.binding.get_concept(subject)
r = self.binding.get_concept(relation)
o = self.binding.get_concept(obj)
bound = BinaryBinding.encode_triple(s, r, o)
# Store in memory
meta = metadata or {}
meta.update({
'type': 'relation',
'subject': subject,
'relation': relation,
'object': obj,
'name': f"{subject}_{relation}_{obj}",
})
# Address = blend of subject and relation (for routing)
address = BinaryBinding.bundle(s, r, o)
return self.memory.store(address, bound, metadata=meta,
name=f"{subject}_{relation}_{obj}")
# ββ Core Reasoning ββββββββββββββββββββββββββββββββββββββββββββββββββββ
def reason(self, query: Any, max_steps: Optional[int] = None,
roles: Optional[List[str]] = None,
verbose: bool = False) -> Dict[str, Any]:
"""
Execute full reasoning pipeline.
Args:
query: input query (string, vector, list of strings, or dict)
max_steps: override max reasoning iterations
roles: roles to decode in response
verbose: print step-by-step info
Returns:
Dict with:
- response: decoded response dict
- reasoning_chain: list of ReasoningStep objects
- total_energy_reduction: cumulative energy decrease
- latency_ms: total time
"""
t0 = time.perf_counter()
steps = max_steps or self.max_reasoning_steps
chain: List[ReasoningStep] = []
# ββ Step 0: Encode query ββββββββββββββββββββββββββββββββββββββ
query_vec = self.encoder.encode(query)
current_state = query_vec.copy()
if verbose:
logger.info(f"Query encoded. Starting reasoning with max {steps} steps.")
for step_num in range(steps):
step = ReasoningStep(step_num)
step.timestamp = time.perf_counter() - t0
# ββ Step 1: Route β find relevant memories ββββββββββββββββ
routing = self.router.route_and_activate(current_state)
step.routing_result = routing
if verbose:
logger.info(
f"Step {step_num}: Routed to {len(routing.indices)} candidates "
f"(depth={routing.depth}, latency={routing.latency_ms:.1f}ms)"
)
if len(routing.indices) == 0:
step.state = current_state
chain.append(step)
break
# ββ Step 2: Gather context from activated memories ββββββββ
active_indices = routing.indices[:min(100, len(routing.indices))]
context_vectors = np.ascontiguousarray(
self.memory._contents[active_indices]
)
act_strengths = routing.activations[:len(active_indices)]
# ββ Step 3: Construct bindings for coherence checking βββββ
bindings = self._extract_bindings(active_indices)
# ββ Step 4: Energy minimization βββββββββββββββββββββββββββ
energy_result = self.energy_model.minimize(
current_state,
context_vectors,
act_strengths,
bindings=bindings,
)
new_state = energy_result['final_state']
step.state = new_state
step.energy = energy_result['final_energy']
if verbose:
logger.info(
f"Step {step_num}: Energy {step.energy:.4f} "
f"(iterations={energy_result.get('total_iterations', '?')})"
)
chain.append(step)
# ββ Check convergence βββββββββββββββββββββββββββββββββββββ
if step_num > 0:
state_change = hamming_distance(current_state, new_state) / N_BITS
energy_change = abs(chain[-2].energy - step.energy)
if state_change < 0.01 and energy_change < 1e-4:
if verbose:
logger.info(f"Reasoning converged at step {step_num}")
break
current_state = new_state
# ββ Decode final state ββββββββββββββββββββββββββββββββββββββββ
final_state = chain[-1].state if chain else query_vec
response = self.decoder.decode(final_state, top_k=10, roles=roles)
# Add decoding to last step
if chain:
chain[-1].decoded = response
total_time = (time.perf_counter() - t0) * 1000
# Energy trajectory
energies = [s.energy for s in chain if s.energy != float('inf')]
energy_reduction = (energies[0] - energies[-1]) if len(energies) >= 2 else 0.0
return {
'response': response,
'reasoning_chain': chain,
'total_energy_reduction': energy_reduction,
'latency_ms': total_time,
'num_steps': len(chain),
'final_state': final_state,
}
def _extract_bindings(self, indices: np.ndarray) -> List[Tuple]:
"""Extract binding triples from activated memory entries for coherence checking."""
bindings = []
for idx in indices:
meta = self.memory._metadata[idx]
if meta and meta.get('type') == 'relation':
subj = meta.get('subject')
rel = meta.get('relation')
obj = meta.get('object')
if subj and rel and obj:
s_vec = self.binding.get_concept(subj)
r_vec = self.binding.get_concept(rel)
o_vec = self.binding.get_concept(obj)
if s_vec is not None and r_vec is not None and o_vec is not None:
bound = self.memory._contents[idx]
# Check: unbind(bound, bind(s,r)) should β o
role_composed = BinaryBinding.bind(s_vec, r_vec)
bindings.append((bound, role_composed, o_vec))
return bindings[:20] # Limit for performance
# ββ Analogy Solving βββββββββββββββββββββββββββββββββββββββββββββββ
def solve_analogy(self, a: str, b: str, c: str,
candidates: Optional[List[str]] = None) -> Dict[str, Any]:
"""
Solve analogy: a is to b as c is to ?
Uses binding-based approach:
1. Compute relation R = bind(a, b) [XOR extracts difference]
2. Apply: query = bind(c, R)
3. Route query through memory
4. Decode nearest match
"""
t0 = time.perf_counter()
# Get or create concept vectors
for name in [a, b, c]:
if self.binding.get_concept(name) is None:
self.add_concept(name)
va = self.binding.get_concept(a)
vb = self.binding.get_concept(b)
vc = self.binding.get_concept(c)
# Compute analogy query
relation = BinaryBinding.bind(va, vb)
query = BinaryBinding.bind(vc, relation)
# Route and find matches
result = self.reason(query, max_steps=2)
# Also do direct codebook comparison
codebook_results = self.binding.solve_analogy(a, b, c, candidates)
return {
'query': f"{a} : {b} :: {c} : ?",
'codebook_ranking': codebook_results[:10],
'memory_response': result['response'],
'latency_ms': (time.perf_counter() - t0) * 1000,
}
# ββ Association βββββββββββββββββββββββββββββββββββββββββββββββββββ
def associate(self, concept: str, top_k: int = 5) -> List[Tuple[str, float]]:
"""Find concepts most associated with the given concept."""
vec = self.binding.get_concept(concept)
if vec is None:
vec = self.encoder.encode(concept)
# Route through memory
routing = self.router.route(vec, beam_width=100)
if len(routing.indices) == 0:
return []
# Collect associated concepts
associations = []
seen = set()
for idx in routing.indices:
meta = self.memory._metadata[idx]
if meta:
name = meta.get('name', '')
if name and name != concept and name not in seen:
sim = 1.0 - routing.distances[len(associations)] / N_BITS \
if len(associations) < len(routing.distances) else 0.0
associations.append((name, float(sim)))
seen.add(name)
if len(associations) >= top_k:
break
return associations
# ββ Composition βββββββββββββββββββββββββββββββββββββββββββββββββββ
def compose(self, *concepts: str) -> Dict[str, Any]:
"""Compose multiple concepts and find what the composition represents."""
vectors = []
for c in concepts:
vec = self.binding.get_concept(c)
if vec is None:
self.add_concept(c)
vec = self.binding.get_concept(c)
vectors.append(vec)
# Bundle all concepts
if len(vectors) == 1:
composed = vectors[0]
else:
composed = BinaryBinding.bundle(*vectors)
# Decode what this composition means
result = self.reason(composed, max_steps=2)
return {
'input_concepts': concepts,
'response': result['response'],
'latency_ms': result['latency_ms'],
}
# ββ Stats βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def stats(self) -> Dict[str, Any]:
"""Return engine statistics."""
return {
'memory': self.memory.stats(),
'codebook_size': len(self.binding._codebook),
'beam_width': self.router.beam_width,
'energy_mode': self.energy_model.mode,
}
|