sdil: normalize causal vectorizer regression

1 files changed, 31 insertions, 11 deletions

diff --git a/sdil/core.py b/sdil/core.py
index 20e94eb..d816549 100644
--- a/sdil/core.py
+++ b/sdil/core.py
@@ -452,7 +452,8 @@ class SDILConfig:
                  pert_sigma=1e-2, pert_every=5, pert_ndirs=1, momentum=0.0, wd=0.0,
                  settle_steps=0, kappa=0.0, feedback="error", p_update_on_neutral=True,
                  normalize_delta=False, pert_mode="layerwise",
-                 raw_scale_control="none", direct_node_pert=False):
+                 raw_scale_control="none", direct_node_pert=False,
+                 vectorizer_optimizer="sgd", vectorizer_eps=1e-6):
         self.eta = eta
         self.eta_A = eta_A
         self.eta_P = eta_P
@@ -480,23 +481,42 @@ class SDILConfig:
         # eta) from that noise -- like normalized SGD.
         self.normalize_delta = normalize_delta
         self.direct_node_pert = direct_node_pert
+        if vectorizer_optimizer not in ("sgd", "nlms"):
+            raise ValueError(f"unknown vectorizer optimizer: {vectorizer_optimizer}")
+        self.vectorizer_optimizer = vectorizer_optimizer
+        self.vectorizer_eps = vectorizer_eps
 
 
-def _update_apical_vectorizer(net, h, c, r_list, qs, eta_A):
+def _update_apical_vectorizer(net, h, c, r_list, qs, cfg):
     """Apply the local causal-regression update to all apical pathways."""
     B = c.shape[0]
     context = h[-2]
+    context_features = None
+    if net.vectorizer_mode == "context_gated":
+        context_features = (c.unsqueeze(2) * torch.tanh(context).unsqueeze(1)).flatten(1)
     for l in range(net.L - 1):
         calibration_error = qs[l] - r_list[l]
-        dA = calibration_error.t() @ c / B
-        net.A[l] += eta_A * dA
+        scaled_error = calibration_error
+        if cfg.vectorizer_optimizer == "nlms":
+            if net.vectorizer_mode == "linear":
+                feature_power = c.square().sum(1, keepdim=True)
+            elif net.vectorizer_mode == "soma_gated":
+                # Each cell i sees [c, tanh(h_i)c], so its effective feature
+                # power is ||c||^2 (1+tanh(h_i)^2).
+                feature_power = (c.square().sum(1, keepdim=True)
+                                 * (1.0 + torch.tanh(h[l + 1]).square()))
+            else:
+                feature_power = (c.square().sum(1, keepdim=True)
+                                 + context_features.square().sum(1, keepdim=True))
+            scaled_error = calibration_error / feature_power.clamp_min(cfg.vectorizer_eps)
+        dA = scaled_error.t() @ c / B
+        net.A[l] += cfg.eta_A * dA
         if net.vectorizer_mode == "soma_gated":
-            dgate = (calibration_error * torch.tanh(h[l + 1])).t() @ c / B
-            net.A_gate[l] += eta_A * dgate
+            dgate = (scaled_error * torch.tanh(h[l + 1])).t() @ c / B
+            net.A_gate[l] += cfg.eta_A * dgate
         elif net.vectorizer_mode == "context_gated":
-            features = (c.unsqueeze(2) * torch.tanh(context).unsqueeze(1)).flatten(1)
-            dgate = calibration_error.t() @ features / B
-            net.A_gate[l] += eta_A * dgate
+            dgate = scaled_error.t() @ context_features / B
+            net.A_gate[l] += cfg.eta_A * dgate
 
 
 def apical_calibration_step(net, x, y, y_onehot, cfg,
@@ -531,7 +551,7 @@ def apical_calibration_step(net, x, y, y_onehot, cfg,
         estimator = (simultaneous_node_perturbation_targets
                      if mode == "simultaneous" else node_perturbation_targets)
         qs = estimator(net, x, y, sigma=cfg.pert_sigma, n_dirs=directions)
-        _update_apical_vectorizer(net, h, c, r_list, qs, cfg.eta_A)
+        _update_apical_vectorizer(net, h, c, r_list, qs, cfg)
     return loss, {"error": c, "did_pert": True}
 
 
@@ -605,7 +625,7 @@ def sdil_step(net, x, y, y_onehot, cfg, step, prev_error=None):
 
         # ================= apical vectorizer A via node perturbation =======
         if did_pert and cfg.learn_A:
-            _update_apical_vectorizer(net, h, c, r_list, qs, cfg.eta_A)
+            _update_apical_vectorizer(net, h, c, r_list, qs, cfg)
 
         # ================= predictor P (neutral) ==========================
         # KEY identification condition. P must learn the soma->apical coupling