figures: lock claim-critical training protocols

1 files changed, 38 insertions, 7 deletions

diff --git a/experiments/plot_main_figures.py b/experiments/plot_main_figures.py
index 7b17b2c..9005caa 100644
--- a/experiments/plot_main_figures.py
+++ b/experiments/plot_main_figures.py
@@ -99,15 +99,31 @@ def require_scaling_protocol(rows):
         for key, expected in (("dataset", "cifar10"), ("width", 64),
                               ("epochs", 5), ("batch_size", 128),
                               ("act", "tanh"), ("residual", 1),
-                              ("device", "cuda")):
+                              ("momentum", 0.9), ("device", "cuda")):
             require_value(errors, row, key, args.get(key), expected)
         name = method(row)
         require_value(errors, row, "depth", args.get("depth"),
                       args.get("depth") if args.get("depth") in DEPTHS else "one of " + repr(DEPTHS))
         require_value(errors, row, "eta", args.get("eta"), 0.02 if name == "SDIL" else 0.05)
+        if name == "FA":
+            for key, expected in (("train_examples", 0), ("max_batches", 0),
+                                  ("feedback_scale", 1),
+                                  ("canonical_preprocess", 1)):
+                require_value(errors, row, key, args.get(key), expected)
+        else:
+            for key, expected in (("max_steps", 0), ("w_scale", 1),
+                                  ("n_in", 3072)):
+                require_value(errors, row, key, args.get(key), expected)
         if name == "SDIL":
-            require_value(errors, row, "pert_mode", args.get("pert_mode"), "simultaneous")
-            require_value(errors, row, "pert_ndirs", args.get("pert_ndirs"), 16)
+            for key, expected in (
+                    ("eta_A", 0.02), ("eta_P", 0.002), ("pert_sigma", 0.01),
+                    ("pert_every", 4), ("pert_mode", "simultaneous"),
+                    ("pert_ndirs", 16), ("use_residual", 1), ("learn_A", 1),
+                    ("learn_P", 1), ("p_neutral", 1), ("p_warmup_steps", 200),
+                    ("p_warmup_eta", 0.05), ("nuis_rho", 0),
+                    ("predictor_mode", "diagonal"), ("normalize_delta", 0),
+                    ("settle_steps", 0), ("kappa", 0), ("feedback", "error")):
+                require_value(errors, row, key, args.get(key), expected)
     if errors:
         raise RuntimeError("mixed CIFAR scaling protocols:\n" + "\n".join(errors))
 
@@ -131,6 +147,9 @@ def require_ep_protocol(groups):
             if name == "EP":
                 require_value(errors, row, "batch_size", args.get("batch_size"), 20)
                 require_value(errors, row, "ep_persistent", args.get("ep_persistent"), 1)
+                require_value(errors, row, "ep_dt", args.get("ep_dt"), 0.5)
+                require_value(errors, row, "canonical_preprocess",
+                              args.get("canonical_preprocess"), 1)
                 resolved = row.get("resolved_protocol", {})
                 for key in ("beta", "free_steps", "nudge_steps", "learning_rates"):
                     require_value(errors, row, f"resolved_protocol.{key}",
@@ -142,9 +161,20 @@ def require_ep_protocol(groups):
                 require_value(errors, row, "act", args.get("act"), "tanh")
                 require_value(errors, row, "residual", args.get("residual"), 0)
                 require_value(errors, row, "eta", args.get("eta"), 0.05)
+                require_value(errors, row, "momentum", args.get("momentum"), 0.9)
+                require_value(errors, row, "max_steps", args.get("max_steps"), 0)
                 if name == "SDIL":
-                    require_value(errors, row, "pert_mode", args.get("pert_mode"), "simultaneous")
-                    require_value(errors, row, "pert_ndirs", args.get("pert_ndirs"), 16)
+                    for key, expected in (
+                            ("eta_A", 0.02), ("eta_P", 0.002),
+                            ("pert_sigma", 0.01), ("pert_every", 4),
+                            ("pert_mode", "simultaneous"), ("pert_ndirs", 16),
+                            ("use_residual", 1), ("learn_A", 1), ("learn_P", 1),
+                            ("p_neutral", 1), ("p_warmup_steps", 200),
+                            ("p_warmup_eta", 0.05), ("nuis_rho", 0),
+                            ("predictor_mode", "diagonal"), ("normalize_delta", 0),
+                            ("settle_steps", 0), ("kappa", 0),
+                            ("feedback", "error")):
+                        require_value(errors, row, key, args.get(key), expected)
     if errors:
         raise RuntimeError("mixed exact-EP protocols:\n" + "\n".join(errors))
 
@@ -481,8 +511,9 @@ GTX 1080. **a,** CIFAR-10 test accuracy versus wall time for width-64 residual M
 five epochs. The black line is the mean-based nondominated frontier among local-learning methods.
 BP is a nonlocal reference and is excluded from frontier construction. Labels give hidden depth.
 **b,** SDIL versus canonical equilibrium propagation (EP) on exact 784–500–10 (d1) and
-784–500–500–10 (d2) forward architectures, the same first 50,000 MNIST training examples, and the
-same epoch counts. EP retains its published raw-pixel, hard-sigmoid energy dynamics and relaxation
+784–500–500–10 (d2) forward architectures and the same first 50,000 MNIST training examples.
+Within each matched architecture, every method uses the same budget: 25 epochs for d1 and 60 epochs
+for d2. EP retains its published raw-pixel, hard-sigmoid energy dynamics and relaxation
 schedule; SDIL retains its z-scored, tanh feedforward dynamics. Thus architecture, examples, and
 epochs are matched, but preprocessing and state dynamics are intentionally method-native.