Deep learning network for classification

2
votes

I know this is a generic question but I would like some feedback. So, I have an image classification task, which I am trying to solve with neural networks. I already have my feature vectors. There are 3 salient feature vectors (point coordinates in 2d space), each of which can belong to three classes. These classes are not the same among these 3 features. (for instance, one could be "up down straight" the other "round square triangular"). These are combined in an overall classification task in 6 output classes (not 27 as the number of combinations would maybe imply). So, what's the question: Are deep learning networks / convolutional networks a good solution to this? Or would you prefer something else? (I have just began learning about these and cant make up my mind yet, so I would appreciate some feedback from more experienced users)

1
matlabimage-processingclassificationdeep-learning
Probably not, since you have low dimensional existing features. Techniques like CNNs, or (stacked) autoencoders shine when it comes to learning features from unlabeled data. Stanford has some good tutorials. – Maurits
thanks a lot for your response :D I ll look into them. – Es.T
There is a Matlab toolbox for Deep Learning in github: github.com/rasmusbergpalm/DeepLearnToolbox – tashuhka

1 Answers

0
votes

As I understood, your problem is not so complicated to implement deep learning. Because you have low dimensional features and few class outputs. Deep learning is mostly for highly nonlinear and large-size classification problems. I suggest you you to use AdaBoost or SVM classifier for this kind of simple classification tasks. Matlab has great tools for above techniques.