13
votes
Accepted
Effect of NOT changing filter weights of CNN during backprop
By not changing the weights of the convolutional layers of a CNN, you are essentially feeding your classifier (the fully connected layer) random features (i.e. not the optimal features for the ...
- 8,038
13
votes
Artificially expanding the datasets through rotation of images in MNIST
The problem is that numbers are not invariant to rotations.
For example, see what happens when you rotate a 4 in steps of 90 degrees:
So unless your task includes recognizing numbers which are ...
- 5,490
10
votes
Accepted
Too low accuracy on MNIST dataset using a neural network
What am I missing?
Incorrect architecture for the classification task. You have a single binary output, trained using binary_crossentropy, so the NN can only ...
- 29.2k
9
votes
Accepted
Are CNNs insensitive to rotations and shifts in images?
If you use max-pooling layers, they may be insensetive to small shifts but not that much. If you want your network to be able to be invariant to transformations, ...
- 14.2k
7
votes
Accepted
0.1 accuracy on MNIST fashion dataset following official Tensorflow/Keras tutorial
You haven't normalized your image dataset such as setting the pixel values between 0-1 which could help classifier converge faster.
Please do it by doing the operation below.
...
5
votes
What is the state-of-the art ANN architecture for MNIST?
Actually, what I'm going to discuss is not an architecture and is like a module used in networks. It performs really well on character based data-sets like MNIST ...
- 14.2k
5
votes
Accepted
GAN discriminator converging to one output
What you are experiencing is called mode collapse, which can occur in GAN Training and is one of its "training instability problems". If you are using Vanilla GAN one effective way is to implement ...
- 346
5
votes
Artificially expanding the datasets through rotation of images in MNIST
In somewhat of a response to your comment to Sammy’s post, the problem doesn’t restrict to five-degree rotations. The problem allows for all rotations.
An $8$ rotated 90 degrees is $\infty$ and no ...
- 4,244
4
votes
Accepted
DC GAN with Batch Normalization not working
Golden Rule: In Keras, if using Batch Normalization layer, train the discriminator on real and fake images separately. Don't combine them.
I was able to solve it by changing the discriminator ...
- 724
4
votes
MNIST Digit dataset requires login
You can download it from here :
Just click on it, and the download will automatically started
https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz
- 141
3
votes
Need a little help Understanding how to build model's in Keras
I have implemented your model to the astonishment, there is a very minute error that is hard to notice.
The way, I was able to get better accuracy is by changing the optimizer to "SGD" or "ADAM".
As ...
- 441
3
votes
Accepted
Why my model can't recognise my own hand written digit?
You have to remember that machine learning model do not understand any concepts as we do, humans. It cannot generalize something it hasn't seen. And yours hasn't seen black digits on white background, ...
- 1,102
3
votes
Are CNNs insensitive to rotations and shifts in images?
Convolution is shift-equivariant except for border effects. Fully connected layers aren't.
Pooling (without subsampling/stride) can be seen as a kind of smoothing, and its output is often the same ...
- 131
3
votes
Difference between grayscaled and binary mnist dataset
MNIST is not an interesting data set. You use MNIST to learn how to do machine learning that you will you on interesting data sets. Thus...
If you just want to figure out how to do the Keras code for ...
- 4,244
3
votes
Accepted
Which neural network is better?
When you train a neural network, you usually use 3 sets: one for training, one for development, one for testing.
Your training set is here for (obviously) training your model: the performance of your ...
- 156
2
votes
Accepted
Why use the .idx data format?
Generally, you will find datasets being distributed in CSV format for their simplicity and human readable format that you could ingest in any programming language with just the packages that the ...
- 38
2
votes
Why not use more than 3 hidden layers for MNIST classification?
Empirically, the network performance does not increase much for a fully-connected network on MNIST when you add layers, but you can probably find ways to improve it on networks with 3+ hidden layers, ...
- 29.2k
2
votes
Softmax vs Sigmoid in RBM/Auto Encoder final layer
This choice mainly depends on what your output represents. Given a vector $\mathbf{x}$, the sigmoid function is given by
$$ \sigma(x_i) = \frac{\exp(x_i)}{1 + \exp(x_i)} $$
while the softmax is given ...
- 1,114
2
votes
Accepted
padding on mnist for LeNet Architecture
Basically, it does exactly what you specify. The used numpy function appends values in each dimension. The amount of "pads" on each axis is specified by ...
- 458
2
votes
Pre-trained CNN for one-shot learning
Because the layers in the CNNs must be already able to extract features from images.
Normally this procedure takes thousands of iterations to be completed, so if they were not trained we could't ...
- 338
2
votes
Accepted
What is the reason that CNN classify some images horribly wrong
Always remember classification is computational process with group of numbers in matrices and when you train a model for particular set it goes through the data and build a matrices that is called as ...
- 157
2
votes
Too low accuracy on MNIST dataset using a neural network
There are a thousand tricks you can use to improve accuracy on MNIST. I am indebted to the Yassine Ghouzam Kaggle Kernel for most of these ideas:
Normalize the data. This allows the optimization to ...
- 422
2
votes
Are CNNs insensitive to rotations and shifts in images?
Try MS COCO dataset: it is VERY diverse, and try training the network for detection/segmentation. The best-performing networks like Mask R-CNN produce about 44% mAP on test data or 68% at 0.5 IoU. It ...
- 777
2
votes
Accepted
How mean and deviation come out with MNIST dataset?
mean : It is the mean of all pixel values in the dataset ( 60000 × 28 × 28 ). This mean is calculated over the whole dataset.
...
- 2,220
2
votes
Accepted
Finding correlation between MNIST digits
You can’t. Correlation is a measure of a variable changes as another variable changes. One goes up by a certain amount, the other usually goes up too: positive correlation. And so on.
What you can ...
- 923
2
votes
Artificially expanding the datasets through rotation of images in MNIST
Sammy and Dave have accurately answered the question that Nielsen intended, but based on your comment on Sammy's answer I think you are wondering whether having arbitrarily many slightly-rotated ...
- 1,530
2
votes
Accepted
Separating styles of numbers for simple digit classification
Interesting question. You are right in assuming that some 1s may be confused for 7s, same with 8s and 3s for instance.
Generally creating different classes as you suggest doesn't happen, simply ...
- 6,106
2
votes
How to make an MNIST classifier work with blank images?
Try thresholding on the image. I believe, you will get ~95% of what is required.
Then try other classical image processing techniques depending on the issue.
...
- 5,759
2
votes
Accepted
Training and validation loss are almost the same (perfect fit?)
Informal explanation:
We usually expect the validation loss to be higher than the training loss due to overfitting. This occurs because the model fits random noise in the training samples, it "...
- 192
2
votes
Using python AI mnist to recognize my picture, trained accuracy is 97.99%, but accuracy to my img is less than 20%
It means that your model is overfitting- when you see that the model performs well on the training data but does not perform well on the test data.
There are various steps to deal with this type of ...
- 171
Only top scored, non community-wiki answers of a minimum length are eligible