If multiple deep learning (DL) algorithms are merged together to create models then the system will be complex.
To analyze this, how to calculate complexity?
Is there any formal way or mathematical proof to analyze this kind of complexity of DL models?
Thanks in advance.
Dear Anik Islam Abhi
Check these papers, they can help
Article Advanced Meta-Heuristics, Convolutional Neural Networks, and...
Article Novel Feature Selection and Voting Classifier Algorithms for...
Dear Anik Islam Abhi ,
You can find out about your question , here :
https://ieeexplore.ieee.org/document/9195618
Dear Anik Islam Abhi ,
I hope you may find useful this recent survey.
Article Evaluating the Quality of Machine Learning Explanations: A S...
I think the complexity of a DL model depends on so many factors. Some of them are:
- no. parameters
- Time and FLOPs (no. floating-point operations) for inferring an instance
- The memory that is required to save the model
- The memory that is required during the inference
- How parallelized the model is (in case that the device supports the parallel computing)
...
There's a paper I would like to recommend: Preprint RepVGG: Making VGG-style ConvNets Great Again
Have a look at this too : https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.641.5324&rep=rep1&type=pdf
Article On the Complexity of Neural Network Classifiers: A Compariso...
Anik Islam Abhi you can use such parameters:
- learning
- processing time
- number of weights (training parameters)
- number of units (neurons)
complexity of a Deep learning model depends on
learning , processing time , number of weights , number of layer and number of neurons in layer
https://towardsdatascience.com/model-complexity-accuracy-and-interpretability-59888e69ab3d
This is a good question that is generally unanswered (or heavily dependent on your application). There are different factors that can give you an idea of the complexity of a neural network such as the number of layers, types of layers used in your NN (this is because basically Conv2D and Conv2DTranspose layers with a high number of filters have more parameters that need to be trained than a simple Dense layer), and so on and so forth. However, depending on your architecture, sometimes you are training more than a single NN (for example GANs where you are training a generator and discriminator which takes much more training and pre-processing time than a very deep dense neural network). Also it depends on the width (less layers but more perceptrons or filters in your layers) and depth (less perceptrons or filters in your layers but more layers) effects. One good reference for this is "Wide Residual Networks", where they express the idea that "it is more computationally effective to widen the layers than have thousands of small kernels as GPU is much more efficient in parallel computations on large tensors". Hope it helps!
- Badges
- Science topic