I am not an expert on the topic, but I do recall of some problems a friend of mine had when working with these: class imbalance and overfitting. She also had some sort of problem since she had a lot of classes (+/- 50 classes), but I mean, that's a problem for almost all types of learners. Something that a lot of people are concerned about is that no theory gives bounds on the amounts of layers to be used, therefore, it is usually a trial and error thing.

Cheers,

In my opinion, the weakness of CNNs lays in the amount of data you provide to them. If you provide them with less, expect the CNNs to perform poorly. CNNs have millions of parameters and with small dataset, would run into an over-fitting problem because they needs massive amount of data to quench the thirst. So, you give much data, CNNs is more strong and willing to give you better performance, you give less data CNNs is very weak...

Cheers

For example try this: https://www.quora.com/What-are-Convolutional-Neural-Networks-and-Deep-Belief-Network-weakness

Yes it's me

I was looking for some papers or refs iclm nips cvpr so as to have a better understanding  of which actually are those weakness 

Thanks

Overfitting is the most problems of CNNs. As reported in the following paper when we change camera and illumination, most CNNs cannot preform well.    

https://www.researchgate.net/publication/312463511_How_Popular_CNNs_Perform_in_Real_Applications_of_Face_Recognition 

Conference Paper How Popular CNNs Perform in Real Applications of Face Recognition

One of the weaknesses was pointed out by Severyn et al. (link below).

It was the dependence of CNNs on the initial parameter tuning (for a good point) to avoid local optima. Thus, a weakness of CNNs is the considerable amount of work they require to initialize according to the problem at hand. This would require some expert knowledge in the domain.

Conference Paper Twitter Sentiment Analysis with Deep Convolutional Neural Networks

CNN is still very reliant on human intervention.

Mostly, CNN has problem of overfitting and its mostly computationally expensive because it has to take a large databaset for training

1. Data requirements leading to overfitting & underfitting

2. Parameter-to-memory relationship

3. Non-expressive learning

4. Non-expressive logic

5. Parameter tuning requirements

6. Computationally expensive

Adding more layers in plain nets (stack layers) leads to higher training error and results in saturation in accuracy

Recommend you to read the review paper which published in 2021

Article Review of deep learning: concepts, CNN architectures, challe...