Is Deep Learning better that the rest of supervised machine learning methods for both classification and regression problems?
Is there any scientific article to support this statement?
Under what conditions the performance of Deep Learning is better ?
The answer is: it depends.
Convolutional neural networks have been consistently doing better than other methods for image classification (ImageNet) type problems, and computer vision applications have become the poster child for deep learning.
If there are enough data points for training, deep networks work as well or better than other techniques. Of course, this statement is about empirical results,
There is also work on theoritical justifications of deep learning.
One of the prime advantages is more flexible prior models.
See this link below for a detailed explanation.
http://rinuboney.github.io/2015/10/18/theoretical-motivations-deep-learning.html
However, there are several domains where the data is not enough yet for training deep nets, and often the classes are highly imbalanced.
In such cases, until we have more data, expert opinion will matter as priors.
Also, there are cases of multimodal data (which helps make a human decide) and temporal (history) data, which may be missing in the dataset used for machine learning.
In such cases, human intervention is still useful at present.
The answer is: it depends.
Convolutional neural networks have been consistently doing better than other methods for image classification (ImageNet) type problems, and computer vision applications have become the poster child for deep learning.
If there are enough data points for training, deep networks work as well or better than other techniques. Of course, this statement is about empirical results,
There is also work on theoritical justifications of deep learning.
One of the prime advantages is more flexible prior models.
See this link below for a detailed explanation.
http://rinuboney.github.io/2015/10/18/theoretical-motivations-deep-learning.html
However, there are several domains where the data is not enough yet for training deep nets, and often the classes are highly imbalanced.
In such cases, until we have more data, expert opinion will matter as priors.
Also, there are cases of multimodal data (which helps make a human decide) and temporal (history) data, which may be missing in the dataset used for machine learning.
In such cases, human intervention is still useful at present.
- Badges
- Science topic