trainNetwork: loss output vs. manual calculation

Dear Matlab community,

I have recently become a bit puzzled when it comes to the trainNetwork function, specifically the diagnostis printed.

I get the following values during the last epoch:

|======================================================================================================================|

| Epoch | Iteration | Time Elapsed | Mini-batch | Validation | Mini-batch | Validation | Base Learning |

| | | (hh:mm:ss) | RMSE | RMSE | Loss | Loss | Rate |

|======================================================================================================================|

| 1000 | 252750 | 12:58:24 | 0.04 | 0.17 | 0.0007 | 0.0136 | 6.2500e-05 |

| 1000 | 252800 | 12:58:34 | 0.03 | 0.16 | 0.0006 | 0.0120 | 6.2500e-05 |

| 1000 | 252850 | 12:58:43 | 0.04 | 0.16 | 0.0007 | 0.0133 | 6.2500e-05 |

| 1000 | 252900 | 12:58:52 | 0.03 | 0.17 | 0.0004 | 0.0143 | 6.2500e-05 |

| 1000 | 252950 | 12:59:01 | 0.03 | 0.16 | 0.0005 | 0.0121 | 6.2500e-05 |

| 1000 | 253000 | 12:59:11 | 0.03 | 0.17 | 0.0004 | 0.0150 | 6.2500e-05 |

One observes that the training loss is much lower than the validation loss, sign of overtraining but not the issue here.

If I now use the trained network, predict the responses and calculate the loss manually I receive:

training: 0.137 / validation: 0.149

This is systematic and leads me to wonder if the "Mini-batch Loss" is not the MSE of the training data.