Hi, I am looking for deep learning approaches to pattern recognition on very small data sets.
Previous attempts using convolutional neural networks have been far less powerful than common machine learning classifiers like support vector machines, random forests or multi layer perceptrons.
Are there new alternatives or approaches that suggest promising and satisfying multi-class (3 - 10 classes) predictions on very small data sets (size approx. 50x500 - 100x10,000)?
Johannes Burdack
https://www.researchgate.net/post/How_to_classify_a_small_number_of_data
https://www.researchgate.net/post/Machine_Learning_Method_for_Smaller_Data_set
Dear Johannes Burdack,
despite the power of deep learning,I think that deep learning cannot be a good selection for classifying very small datasets. I suggest using some simpler classifiers such as neural networks for very small dataset classification.
I hope this comment will be usefull for you.
If you happen to have a small dateset with labels. but have many unlabeled examples from the same distribution, Generative adversarial networks can be a interesting choice to try with an semi-supervised appraoch
Article Semi-Supervised Learning with Generative Adversarial Networks
reports training of a high accuracy classifier with less generalization error with a very small set of labeled examples.
Generative Adversarial Networks (GANs) can be used for data augmentation to enhance the quality of Deep Learning trainings since DL models are data-hungry models.
- Badges
- Science topic