So I trained CNN model for people detection on caltech-pedestrian dataset: Then I was curious and evaluated the model in every 1000th iteration on Evaluation toolbox(I guarantee, there is no bug in evaluation).

However, plot of the performance does not look so good. The miss rate spikes between 20K(20,000) and 30K iterations.

I am confused what does that mean. I mean usually we would expect the miss rate to decrease as we train the model more.

I am using yolo object detection network https://github.com/AlexeyAB/darknet/blob/master/cfg/tiny-yolo-voc.cfg

So learning options are the same as https://github.com/AlexeyAB/darknet/blob/master/src

enter image description here

I also tried 10x smaller learning rate. However, no luck: enter image description here

  • 1
    $\begingroup$ It's weird to have such similar learning curves given the different learning and miss rates. What's your optimizer? How big is your network? Are you minibatching? If so, what size? $\endgroup$
    – Emre
    May 23, 2017 at 7:47
  • $\begingroup$ I am using darknet framework. It uses momentum for update. My network is really big, I put link for configuration (it is modified version of inceptionV3). Batch size is 64. YOLO waits to fill the batch by performing multiple subdivisions. github.com/AlexeyAB/darknet/blob/master/src/… $\endgroup$ May 23, 2017 at 8:35
  • 1
    $\begingroup$ It could just be an artifact of minibatching, which introduces noise that has a desirable regularization effect. See what happens when you increase the minibatch size. $\endgroup$
    – Emre
    May 23, 2017 at 18:05

1 Answer 1


This is a relatively common phenomena called double descent.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.