Reducing bias in cnn model

#1

I have tried using dropouts, normalised the values still the validation accuracy is fluctuating a lot sometime avove 90 sometime below 90.

PLEASE TELL ME THE WAYS TO REDUCE THIS BIAS IN ACCURACY

#2

please someone reply

#3

Dropout is technique to reduce overfitting and thus increasing the bias.

In case, you model is not giving great results and has high bias (i.e. giving bad accuracy on both validation as well as training dataset), you should try to look at the following:

  1. Try increasing number of neurons in each layer
  2. Try increasing number of layers
  3. If your data is image or any other spatial data where the nearby pixels matter, try using CNN (Conv. Neural Network) or any complex variants of CNN such as inception v3.
  4. In case, you data has a sequence (e.g. words in a sentence, price in stock, chat conversation etc). Try using the RNN (LSTM)
  5. Also, observe is the data is imbalance
  6. Try coming up with new set of features

In case, during training, the error is not at all dropping in epochs. Try the following:

  1. If the error is increasing quickly instead of reducing, chances are that the learning rate is too high. Try reducing the learning rate.
  2. In case, error is not changing at all. There might be vanishing gradients problem. Try initializing the neural network and use RELU instead of sigmoid activation function.

If the data is not sufficient:

  1. Use pretrained neural network from various model zoos:
    – https://github.com/onnx/models
    – https://modelzoo.co/ ,
    – https://caffe.berkeleyvision.org/model_zoo.html
  2. Augmenting the data
  3. Gathering more data