summaryrefslogtreecommitdiff
path: root/KP_RAW_TRAFFIC_SCALING.md
diff options
context:
space:
mode:
authorYurenHao0426 <Blackhao0426@gmail.com>2026-08-27 10:35:06 -0500
committerYurenHao0426 <Blackhao0426@gmail.com>2026-08-27 10:35:06 -0500
commit2b9caf80d9a1054ec585a03718fc37106f8c7aae (patch)
tree230c9d478498fbce5249edbe74b1abbccdabbfe4 /KP_RAW_TRAFFIC_SCALING.md
parentdb21295addcfeb830ae72b330a8e7784c75c5232 (diff)
exp: freeze raw KP traffic scaling control
Diffstat (limited to 'KP_RAW_TRAFFIC_SCALING.md')
-rw-r--r--KP_RAW_TRAFFIC_SCALING.md51
1 files changed, 51 insertions, 0 deletions
diff --git a/KP_RAW_TRAFFIC_SCALING.md b/KP_RAW_TRAFFIC_SCALING.md
new file mode 100644
index 0000000..9682500
--- /dev/null
+++ b/KP_RAW_TRAFFIC_SCALING.md
@@ -0,0 +1,51 @@
+# Raw KP under four-RMS mixed traffic
+
+## Question
+
+The A6000 crossover shows that SDIL remains close to clean reciprocal
+Kolen--Pollack (KP) learning as ResNet depth increases. This control asks
+whether the same KP substrate remains trainable when it receives SDIL's
+four-RMS soma-predictable apical traffic without subtracting the predictable
+component.
+
+The compared teaching signals are
+
+```text
+clean KP: u = s
+raw KP: u = s + t
+SDIL: u = s + t - prediction(t)
+```
+
+Here `s` is the reciprocal KP instruction and `t` is the frozen traffic field.
+Raw KP executes the same 64-example closed-form predictor fit and the same
+instruction-off neutral projection as SDIL, but uses the mixed apical vector
+unchanged. The predictor and projection are therefore charged equally while
+the subtractive direction is the only effective difference.
+
+## Frozen validation panel
+
+Run raw KP at ResNet-20, ResNet-32, and ResNet-56 with the settings used by the
+A6000 crossover:
+
+- CIFAR-10 45,000/5,000 train/validation split; test remains untouched;
+- standard CIFAR `6n+2` ResNet, base width 16, option-A shortcuts, BatchNorm;
+- model, loader, split, and traffic seeds `0`, `0`, `2027`, and `5000`;
+- batch size 128, 200 epochs, standard crop/flip augmentation;
+- SGD momentum 0.9, weight decay `1e-4`, learning rate 0.1, drops at epochs
+ 100 and 150;
+- reciprocal KP feedback, four-RMS traffic, one 64-example closed-form neutral
+ fit, and one paired neutral projection per ordinary minibatch;
+- validation evaluation after every epoch and at the final endpoint.
+
+The learning rate and traffic strength stay fixed across depth. Every depth
+is retained whether it completes, becomes nonfinite, or reaches the 48-hour
+timeout. The primary metric is final validation accuracy; stability and the
+first nonfinite event are reported alongside it.
+
+## Interpretation
+
+The panel is paired with the clean-KP and SDIL cells from the A6000 crossover.
+A growing raw-KP deficit with depth, together with stable SDIL endpoints,
+supports the claim that subtracting soma-predictable traffic preserves KP's
+scaling behavior under mixed feedback. Clean KP remains the clean-feedback
+reference.