As is well known the main concern of the support vector machines (SVM) is to provide the maximum margin between different patterns; and when coupled with the kernels trick or modeling it will provide avery flexible framework for pattern recognition. Now the question is:
Is it safe to say that the Support Vector Machines and the Kernels will replace the Neural Networks in supervised learning?
In my opinion, no. For two main reasons:
- Theoretical perspective: due to no-free-lunch theorems, it is hard to make such predictions.
- Practical use-cases: SVM may have a stronger mathematical foundation (I can't say this is 100% accurate, but it is widely believed). However, neural nets use a more flexible representation, which makes it "easier" in practical cases to extract features. In addition, neural net representation, fast learning algorithms, and GPUs make it possible to use very deep architectures, which is a significant advantage over SVM especially in computer vision (
Article Benefits of depth in neural networks
).
Yes . Support vector machine is safe in medical applications . It has strong decssion for detection and classification
Hi Gamal,
The term "replace", that you used, is not accurate. This is because you are addressing two different algorithms, where each has its own manner to solve a problem. Moreover, without training/testing those two methods, we don't know which one of them produces best classification result (i.e., they could produce a different result). Consequently, it's better to test support vector machines and neural network first before making the final decision about the one that has the best performance.
HTH.
Samer Sarsam, PhD.
Dear Samer,
First of all, thank you very much for your feedback. from a conceptual point of view, SVM can be seen as an extension of neural network since the main difference is the maximization of the distances between classes. Yes, the kernels add another capability to SVM but SVM can be seen as an extended class of neural network. This does not deny the complexity of its mathematics for non-linear modeling.
Regards,
Gamal
Dear Gamal,
You're welcome.
To make things clearer, in your recent reply, you've said that " from a conceptual point of view, SVM can be seen as an extension of neural network", while in your main question, you asked whether "Support Vector Machines and the Kernels will replace the Neural Networks in supervised learning". As you can see, in the two sentences you're using two terms: extension and replace. Those terms have different meaning in our domain. I might agree that SVM can be seen as an extension of neural network, but it can't replace it.
Cheers,
Dr. Samer Sarsam
Hello Gamal,
The thing with SVM is that it is a linear classifier. Most machine learning algorithms function well as long as the dataset has linear properties (either directly or after the application of some filters). That's where neural networks come in. Though it is associated with machine learning, it is a completely new domain (new from the previous ML models). It's specialty lies in the fact that it can learn from non-linear data with significant accuracy. It has many models (like ANN, CNN, RNN, SOM, Boltzman Machine, etc) which are suited for different types of datasets.
Hence, SVM can never replace neural networks as the later is more advanced and holds much potential in the future.
I hope this helps.
Have a good day,
Vaibhav.
SVMs and Neural networks are oceans apart, as perceived by me. A supervised network will learn approximately from the data presented and map input space to output space. They have tolerance to approximation and soft computing kind. While SVM draws linear and non-linear discriminators to partition the solution space. They are hard-limiting kind.
A key point is how SVMs cannot easily scale well to large datasets (due to both RAM requirements and training time that scales superlinearly with training samples). Additionally, SVMs are more difficult to paralellise than NNs and fewer such frameworks are available as of now compared to NNs.
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling