drift_neuron_discovery.py

"""
Drift Neuron Discovery via L1-Regularized Probing
===================================================
Inspired by:
- Semantic Entropy Neurons (NeurIPS 2024): L1 probes find sparse neuron sets
- Neuron Circuits (Arora & Wu, 2026): MLP activations are a sparse basis
- Calibration paper (Radharapu et al.): Brier score loss for calibration

Key idea: Train L1-regularized linear probes on MLP ACTIVATIONS (not outputs!)
to find the minimal set of neurons that predict whether knowledge has drifted.

If a small set of "drift neurons" exists, this validates that:
1. The model encodes temporal validity internally
2. Drift detection can be done with minimal compute
3. These neurons can potentially be steered (connecting to YaPO)

Usage:
    python drift_neuron_discovery.py \
        --model Qwen/Qwen2.5-7B-Instruct \
        --dataset data/knowledge_drift_dataset.json \
        --output data/drift_neurons/ \
        --max_samples 500
"""

import argparse
import json
import os
import logging
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from tqdm import tqdm
from collections import defaultdict
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import roc_auc_score, average_precision_score

logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)


# ============================================================
# MODEL LOADING WITH MLP ACTIVATION HOOKS
# ============================================================

class MLPActivationExtractor:
    """
    Extract MLP activations (pre-down-projection) from each layer.
    
    Per Arora & Wu (2026): MLP activations are a PRIVILEGED BASIS
    due to element-wise nonlinearity (SiLU/GeLU). Much sparser than
    MLP outputs or residual stream.
    """
    def __init__(self, model):
        self.model = model
        self.activations = {}
        self.hooks = []
        self._register_hooks()
    
    def _register_hooks(self):
        """Register forward hooks on MLP intermediate activations."""
        for name, module in self.model.named_modules():
            # For Qwen2.5: the gate/up projection output before down projection
            # Architecture: gate_proj -> silu -> * up_proj -> down_proj
            if 'mlp.gate_proj' in name or 'mlp.up_proj' in name:
                layer_idx = name.split('.')[2] if 'layers' in name else name.split('.')[1]
                hook = module.register_forward_hook(
                    lambda mod, inp, out, n=name, l=layer_idx: 
                    self._save_activation(n, l, out)
                )
                self.hooks.append(hook)
    
    def _save_activation(self, name, layer_idx, output):
        key = f"layer_{layer_idx}"
        if key not in self.activations:
            self.activations[key] = {}
        if 'gate' in name:
            self.activations[key]['gate'] = output.detach()
        elif 'up' in name:
            self.activations[key]['up'] = output.detach()
    
    def get_mlp_activations(self):
        """
        Get the actual MLP hidden activations: SiLU(gate) * up
        This is the privileged basis from Arora & Wu (2026).
        """
        result = {}
        for key, acts in self.activations.items():
            if 'gate' in acts and 'up' in acts:
                # SiLU activation applied to gate, then elementwise multiply with up
                hidden = F.silu(acts['gate']) * acts['up']
                result[key] = hidden
        return result
    
    def clear(self):
        self.activations = {}
    
    def remove_hooks(self):
        for h in self.hooks:
            h.remove()


def extract_features(model, tokenizer, query, extractor, device="cuda"):
    """
    Extract MLP activations and hidden states for a single query.
    Returns per-layer features at the last token position.
    """
    prompt = f"<|im_start|>system\nAnswer concisely.<|im_end|>\n<|im_start|>user\n{query}<|im_end|>\n<|im_start|>assistant\n"
    
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    
    extractor.clear()
    with torch.no_grad():
        outputs = model(**inputs, output_hidden_states=True)
    
    # Get MLP activations (privileged basis)
    mlp_acts = extractor.get_mlp_activations()
    
    # Get hidden states (residual stream)
    hidden_states = outputs.hidden_states
    
    features = {}
    
    # MLP activations at last token, per layer
    for key, act in mlp_acts.items():
        layer_idx = int(key.split('_')[1])
        features[f"mlp_act_layer_{layer_idx}"] = act[0, -1, :].cpu()  # [ffn_dim]
    
    # Hidden states at last token, per layer
    for i, h in enumerate(hidden_states):
        features[f"hidden_layer_{i}"] = h[0, -1, :].cpu()  # [hidden_dim]
    
    return features


# ============================================================
# DRIFT PROBE WITH L1 REGULARIZATION
# ============================================================

class DriftProbeL1(nn.Module):
    """
    Linear probe with L1 regularization for drift detection.
    
    Trained with Brier score loss (proper scoring rule) for calibration.
    L1 regularization forces sparsity, revealing "drift neurons."
    """
    def __init__(self, input_dim):
        super().__init__()
        self.linear = nn.Linear(input_dim, 1)
    
    def forward(self, x):
        return torch.sigmoid(self.linear(x))
    
    def get_nonzero_weights(self, threshold=1e-4):
        """Get indices of neurons with non-negligible weights."""
        weights = self.linear.weight.data.abs().squeeze()
        nonzero = (weights > threshold).nonzero(as_tuple=True)[0]
        return nonzero, weights[nonzero]


def brier_score_loss(pred, target):
    """Brier score: proper scoring rule for calibration."""
    return ((pred.squeeze() - target.float()) ** 2).mean()


def train_probe(features, labels, layer_key, l1_lambda=0.01, epochs=200, lr=0.001):
    """
    Train an L1-regularized probe on a specific layer's features.
    
    Returns: trained probe, metrics, discovered neurons
    """
    X = torch.stack(features).float()
    y = torch.tensor(labels).float()
    
    input_dim = X.shape[1]
    
    # 5-fold cross-validation
    skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
    fold_metrics = []
    best_probe = None
    best_auroc = 0
    
    for fold, (train_idx, val_idx) in enumerate(skf.split(X, y)):
        X_train, X_val = X[train_idx], X[val_idx]
        y_train, y_val = y[train_idx], y[val_idx]
        
        probe = DriftProbeL1(input_dim)
        optimizer = torch.optim.Adam(probe.parameters(), lr=lr, weight_decay=0)
        
        for epoch in range(epochs):
            probe.train()
            pred = probe(X_train)
            
            # Brier score + L1 regularization
            loss = brier_score_loss(pred, y_train)
            l1_reg = l1_lambda * probe.linear.weight.abs().sum()
            total_loss = loss + l1_reg
            
            optimizer.zero_grad()
            total_loss.backward()
            optimizer.step()
        
        # Evaluate
        probe.eval()
        with torch.no_grad():
            val_pred = probe(X_val).squeeze().numpy()
            val_labels = y_val.numpy()
        
        try:
            auroc = roc_auc_score(val_labels, val_pred)
            auprc = average_precision_score(val_labels, val_pred)
        except ValueError:
            auroc, auprc = 0.5, 0.5
        
        # Brier score (lower is better)
        brier = np.mean((val_pred - val_labels) ** 2)
        
        # Count non-zero neurons
        nonzero_idx, nonzero_weights = probe.get_nonzero_weights()
        n_neurons = len(nonzero_idx)
        
        fold_metrics.append({
            "auroc": auroc, "auprc": auprc, "brier": brier, "n_neurons": n_neurons
        })
        
        if auroc > best_auroc:
            best_auroc = auroc
            best_probe = probe
    
    avg_metrics = {
        "auroc": np.mean([m["auroc"] for m in fold_metrics]),
        "auroc_std": np.std([m["auroc"] for m in fold_metrics]),
        "auprc": np.mean([m["auprc"] for m in fold_metrics]),
        "brier": np.mean([m["brier"] for m in fold_metrics]),
        "n_neurons": np.mean([m["n_neurons"] for m in fold_metrics]),
    }
    
    # Get drift neurons from best probe
    nonzero_idx, nonzero_weights = best_probe.get_nonzero_weights()
    sorted_order = nonzero_weights.argsort(descending=True)
    drift_neurons = [
        {"neuron_idx": nonzero_idx[i].item(), "weight": nonzero_weights[i].item()}
        for i in sorted_order[:50]  # Top 50 by weight magnitude
    ]
    
    return best_probe, avg_metrics, drift_neurons


# ============================================================
# MAIN ANALYSIS
# ============================================================

def run_full_analysis(model, tokenizer, samples, output_dir, device="cuda", max_samples=None):
    os.makedirs(output_dir, exist_ok=True)
    
    if max_samples:
        drifted = [s for s in samples if s.get("is_drifted_query")]
        not_drifted = [s for s in samples if not s.get("is_drifted_query") 
                       and s.get("temporal_zone") == "post_cutoff"]
        import random; random.seed(42)
        n_d = min(len(drifted), max_samples // 2)
        n_nd = min(len(not_drifted), max_samples - n_d)
        samples = random.sample(drifted, n_d) + random.sample(not_drifted, n_nd)
        logger.info(f"Sampled {n_d} drifted + {n_nd} non-drifted = {len(samples)} total")
    
    extractor = MLPActivationExtractor(model)
    
    # === STEP 1: Extract features ===
    logger.info("Extracting MLP activations and hidden states...")
    all_features = defaultdict(list)
    all_labels = []
    
    for sample in tqdm(samples, desc="Extracting features"):
        try:
            features = extract_features(model, tokenizer, sample["query"], extractor, device)
            for key, feat in features.items():
                all_features[key].append(feat)
            all_labels.append(1 if sample.get("is_drifted_query") else 0)
        except Exception as e:
            logger.error(f"Error extracting: {e}")
    
    n_drifted = sum(all_labels)
    n_total = len(all_labels)
    logger.info(f"Extracted features: {n_total} samples ({n_drifted} drifted, {n_total-n_drifted} non-drifted)")
    
    # === STEP 2: Train probes per layer ===
    logger.info("Training L1-regularized drift probes...")
    
    layer_results = {}
    
    # Try different L1 strengths
    l1_values = [0.001, 0.005, 0.01, 0.05, 0.1]
    
    for layer_key in sorted(all_features.keys()):
        if not all_features[layer_key]:
            continue
        
        best_l1_result = None
        best_auroc = 0
        
        for l1 in l1_values:
            probe, metrics, neurons = train_probe(
                all_features[layer_key], all_labels, layer_key, l1_lambda=l1
            )
            
            if metrics["auroc"] > best_auroc:
                best_auroc = metrics["auroc"]
                best_l1_result = {
                    "layer": layer_key, "l1_lambda": l1,
                    "metrics": metrics, "drift_neurons": neurons,
                }
        
        if best_l1_result:
            layer_results[layer_key] = best_l1_result
    
    # === STEP 3: Find best layers and neurons ===
    print("\n" + "=" * 90)
    print("  DRIFT NEURON DISCOVERY RESULTS")
    print("=" * 90)
    
    # Sort by AUROC
    sorted_layers = sorted(layer_results.items(), key=lambda x: x[1]["metrics"]["auroc"], reverse=True)
    
    print(f"\n{'Layer':<30} {'AUROC':>8} {'AUPRC':>8} {'Brier':>8} {'#Neurons':>10} {'L1':>8}")
    print("-" * 80)
    
    for layer_key, result in sorted_layers[:20]:
        m = result["metrics"]
        print(f"{layer_key:<30} {m['auroc']:>8.4f} {m['auprc']:>8.4f} {m['brier']:>8.4f} "
              f"{m['n_neurons']:>10.1f} {result['l1_lambda']:>8.3f}")
    
    # Best layer
    if sorted_layers:
        best_layer, best_result = sorted_layers[0]
        print(f"\n  🏆 BEST LAYER: {best_layer}")
        print(f"     AUROC: {best_result['metrics']['auroc']:.4f} ± {best_result['metrics']['auroc_std']:.4f}")
        print(f"     Active neurons: {best_result['metrics']['n_neurons']:.0f}")
        print(f"     Top drift neurons:")
        for n in best_result["drift_neurons"][:10]:
            print(f"       Neuron {n['neuron_idx']:>6d}: weight = {n['weight']:.4f}")
    
    # === STEP 4: MLP activations vs hidden states comparison ===
    print("\n\n  === MLP ACTIVATIONS vs HIDDEN STATES ===")
    mlp_layers = [(k, v) for k, v in sorted_layers if "mlp_act" in k]
    hidden_layers = [(k, v) for k, v in sorted_layers if "hidden" in k]
    
    if mlp_layers and hidden_layers:
        best_mlp = mlp_layers[0][1]["metrics"]["auroc"]
        best_hidden = hidden_layers[0][1]["metrics"]["auroc"]
        print(f"  Best MLP activation probe:  AUROC = {best_mlp:.4f} ({mlp_layers[0][0]})")
        print(f"  Best hidden state probe:    AUROC = {best_hidden:.4f} ({hidden_layers[0][0]})")
        
        if best_mlp > best_hidden:
            print(f"  ✅ MLP activations are MORE informative (consistent with Arora & Wu 2026)")
        else:
            print(f"  ℹ️  Hidden states are more informative for this task")
    
    # === SAVE RESULTS ===
    save_results = {}
    for k, v in layer_results.items():
        save_results[k] = {
            "l1_lambda": v["l1_lambda"],
            "metrics": v["metrics"],
            "drift_neurons": v["drift_neurons"],
        }
    
    with open(os.path.join(output_dir, "drift_neuron_results.json"), 'w') as f:
        json.dump(save_results, f, indent=2)
    
    # Save the layer ranking
    ranking = [{"layer": k, **v["metrics"], "l1": v["l1_lambda"]} for k, v in sorted_layers]
    with open(os.path.join(output_dir, "layer_ranking.json"), 'w') as f:
        json.dump(ranking, f, indent=2)
    
    logger.info(f"Results saved to {output_dir}")
    return layer_results


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", default="Qwen/Qwen2.5-7B-Instruct")
    parser.add_argument("--dataset", default="data/knowledge_drift_dataset.json")
    parser.add_argument("--output", default="data/drift_neurons/")
    parser.add_argument("--max_samples", type=int, default=None)
    parser.add_argument("--device", default="auto")
    parser.add_argument("--post_cutoff_only", action="store_true")
    args = parser.parse_args()
    
    from transformers import AutoModelForCausalLM, AutoTokenizer
    
    logger.info(f"Loading model: {args.model}")
    tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    model = AutoModelForCausalLM.from_pretrained(
        args.model, torch_dtype=torch.float16, device_map=args.device, trust_remote_code=True,
    )
    model.eval()
    
    with open(args.dataset, 'r') as f:
        dataset = json.load(f)
    samples = dataset["samples"]
    if args.post_cutoff_only:
        samples = [s for s in samples if s.get("temporal_zone") == "post_cutoff"]
    
    device = "cuda" if torch.cuda.is_available() else "cpu"
    run_full_analysis(model, tokenizer, samples, args.output, device, args.max_samples)


if __name__ == "__main__":
    main()