fix(training): correct loop bounds and indexing in weights update

The change fixes incorrect loop bounds and indexing when updating weights from the hidden layer to output layer. The outer loop now iterates over network.inputCount instead of network.hiddenCount, and the inner loop iterates over network.hiddenCount instead of network.outputCount. This ensures that each input node properly updates its respective connections to the hidden nodes, correcting potential miscalculations in the weight adjustments during training.

Author: dlidstrom

Cpp/neural.cpp

@@ -129,9 +129,10 @@ void Trainer::Train(const Vector& input, const Vector& y, double lr) {
         }
     }
 
-    for (size_t r = 0; r < network.hiddenCount; r++) {
-        for (size_t c = 0; c < network.outputCount; c++) {
-            network.weightsHidden[r * network.hiddenCount + c] -= lr * gradHidden[c] * input[r];
+    for (size_t r = 0; r < network.inputCount; r++) {
+        for (size_t c = 0; c < network.hiddenCount; c++) {
+            network.weightsHidden[r * network.hiddenCount + c] -=
+                lr * gradHidden[c] * input[r];
         }
     }