I am using conv1d to classify EEG signals, but my val_accuracy stuck at 0.65671. No matter what changes i do, it never go beyond 0.65671. Here is the architecture



Shape of training data is (5073,3072,7) and for test data it is (1908,3072,7).

I have tried reducing the number of neurons in each layer, changing activation function, and add more layers. But this upper limit has not changed mostly.

I have tried one hot encoding of binary class, using keras.utils.to_categorical(y_train,num_classes=2) but this issue does not resolve.

I have tried learning rate of 0.0001, but it does not work. I have tried some kernel_initializer and optimizers but nothing help


   Train on 5073 samples, validate on 1908 samples
Epoch 1/8
 - 23s - loss: 0.6865 - acc: 0.5757 - val_loss: 0.6709 - val_acc: 0.6564

Epoch 00001: val_acc improved from -inf to 0.65645, saving model to weights.hdf5
Epoch 2/8
 - 22s - loss: 0.6760 - acc: 0.5837 - val_loss: 0.6569 - val_acc: 0.6567

Epoch 00002: val_acc improved from 0.65645 to 0.65671, saving model to weights.hdf5
Epoch 3/8
 - 21s - loss: 0.6661 - acc: 0.5843 - val_loss: 0.6669 - val_acc: 0.6111

Epoch 00003: val_acc did not improve from 0.65671
Epoch 4/8
 - 21s - loss: 0.6622 - acc: 0.5915 - val_loss: 0.6579 - val_acc: 0.6253

Epoch 00004: val_acc did not improve from 0.65671
Epoch 5/8
 - 22s - loss: 0.6575 - acc: 0.5939 - val_loss: 0.6540 - val_acc: 0.6255

Epoch 00005: val_acc did not improve from 0.65671
Epoch 6/8
 - 21s - loss: 0.6554 - acc: 0.5940 - val_loss: 0.6448 - val_acc: 0.6399

Epoch 00006: val_acc did not improve from 0.65671
Epoch 7/8
 - 21s - loss: 0.6511 - acc: 0.6042 - val_loss: 0.6584 - val_acc: 0.6195

Epoch 00007: val_acc did not improve from 0.65671
Epoch 8/8
 - 21s - loss: 0.6487 - acc: 0.6059 - val_loss: 0.6647 - val_acc: 0.6030

Epoch 00008: val_acc did not impr

ove from 0.65671

  • $\begingroup$ Did you find a solution for this? I'm using EEG and my models end up getting stuck at around 65% as well $\endgroup$ – Levi Moreira Jul 14 at 13:22
  • $\begingroup$ no, i did not. . there are papers, claiming good accuracy. I implemented some of them and reproduce their results. Issue is that their models are very data specific. Changing the data resulted in very low accuracy $\endgroup$ – Talha Anwar Jul 14 at 13:35
  • $\begingroup$ Yeah, I tried reproducing some of the papers but not even that :/ They claim good accuracy but when I try to replicate the experiments it always gets stuck at that 65% area. I've tried multiple architectures and it's always the same, so frustrating. $\endgroup$ – Levi Moreira Jul 14 at 13:39
  • $\begingroup$ are you using the same data as mentioned in paper for reproducing the paper? $\endgroup$ – Talha Anwar Jul 14 at 13:47
  • $\begingroup$ yep, I'm trying to replicate some experiments around seizure detection. I'm using the CHB MIT dataset and trying to do cross patient classification. So I train with some patients and test on the others. The results are always around 65% no matter the architecture I pick :/ $\endgroup$ – Levi Moreira Jul 14 at 13:50

I am using 1D CNNs for EEG/EMG classification as well. One thing that seems to help for me is playing around with the number of filters, and yours seem quite low. I have used up to 80 filters on a layer, at times with good results. Also you may want to reverse how you are doing things and add more filters at the beginning and reduce with each successive layer.

| improve this answer | |
  • $\begingroup$ I have tried this, but this does not help me $\endgroup$ – Talha Anwar Sep 11 '19 at 11:25
  • $\begingroup$ Which kind of eeg signal are you trying to classify? $\endgroup$ – stefanLopez Sep 11 '19 at 13:56
  • $\begingroup$ Scizhophrenia vs normal $\endgroup$ – Talha Anwar Sep 12 '19 at 7:12
  • $\begingroup$ Hmm, so this is not a standard EEG waveform classification then. You may want to try something other than a 1D CNN. 1D CNNs are great if you have a time series waveform that you are looking for, but I imagine it takes a much more holistic approach to classify schizophrenia. Specifically, you may want to try models that focus more on the relationship between sites in the brain. $\endgroup$ – stefanLopez Sep 12 '19 at 16:36
  • $\begingroup$ any idea from where should i start, or what should i look for, Btw i try machine learning too, but that does not help $\endgroup$ – Talha Anwar Sep 12 '19 at 17:07

I would like to see you data set :) I am also doing some signal classification.

Unless there is some simple bug in data preprocessing stage: (check what you didn't show here first!)

  • As correctly pointed to you by @stefanLopez your number of filters is way too low.
  • Next, filter length is too short to capture anything serious.
  • Remove batchnorm while testing.
  • Reduce dropout while testing.
  • Test with ELU (Exponential Linear Unit) activation.
  • Last, use more FC layers with more neurons.
  • Try using glorot (commonly known as Xavier) initializer.

Example model:






Tell if it helps.

| improve this answer | |

I hit the same issue, with a different network/task.

I'm using a fully-connected network to regress a vector from an image. Pretty quickly, after 1-2 epochs, both training and validation seem to be stuck in some values. Curiously, they also vary around second decimal, despite being an order of magnitude larger than in your case (my: loss ~7.2, error ~7.9).

The reason was a bug in the batch generator function, which could come to a state where it always returns the same batch for validation. I've found the bug by creating a debug data set, which had only 10 samples (images).

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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