Clarify some terminology.HEADmain

1 files changed, 6 insertions, 7 deletions

diff --git a/src/lib/src/train.c b/src/lib/src/train.c
index fe9f598..7559ece 100644
--- a/src/lib/src/train.c
+++ b/src/lib/src/train.c
@@ -239,17 +239,16 @@ void nnTrain(
 
         // Compute this layer's gradient.
         //
-        // By "gradient" we mean the expression common to the weights and bias
-        // gradients. This is the part of the expression that does not contain
-        // this layer's input.
+        // By 'gradient' we mean the subexpression common to all the gradients
+        // for this layer.
+        // For linear layers, this is the subexpression common to both the
+        // weights and bias gradients.
         //
         // Linear:   G = id
         // Relu:     G = (output_k > 0 ? 1 : 0)
         // Sigmoid:  G = output_k * (1 - output_k)
         switch (layer->type) {
         case nnLinear: {
-          // TODO: Just copy the pointer?
-          *gradient = nnMatrixBorrow(&errors[l]);
           break;
         }
         case nnRelu:
@@ -294,7 +293,7 @@ void nnTrain(
           nnMatrix*     layer_biases  = &linear->biases;
 
           // Outer product to compute the weight deltas.
-          nnMatrixMulOuter(layer_input, gradient, &weight_deltas[l]);
+          nnMatrixMulOuter(layer_input, &errors[l], &weight_deltas[l]);
 
           // Update weights.
           nnMatrixScale(&weight_deltas[l], params->learning_rate);
@@ -304,7 +303,7 @@ void nnTrain(
           // This is the same formula as for weights, except that the o_j term
           // is just 1.
           nnMatrixMulSub(
-              layer_biases, gradient, params->learning_rate, layer_biases);
+              layer_biases, &errors[l], params->learning_rate, layer_biases);
         }
       }