This is problem dependent. You can run simulations using both ANN and SVM. And If you run a statistical test to compare and see if there's any significant difference between the performance of the two classifiers, you might get positive or negative results. From my experience, for some data sets you will get comparable / very similar results. But for some other data sets, SVM might perform better significantly.

As the two previous answers point out, this depends on the problem and dataset. I'd only like to add a few more points to think about that might help with the choice (but in case of doubt, Kee Huong Lai's recommendation to evaluate both approaches is the best way).

Neural networks (especially deep ones) typically perform better on very (very, very) large datasets. On smaller application domains, you either have to choose a very small network (in which case SVMs will probably perform better) or you need very good techniques to keep the network from overfitting. Also, neural networks profit a lot if the datapoints are structured in a way that can be exploited by the architecture (e.g., in the case of image or audio data, using convolutions, or in the case of time series using RNNs or LSTMs, as Bahadorreza Ofoghi's answer pointed out). If that is the case with your dataset, neural networks may be a better choice.

Another thing to think about is that SVMs are (in my opinion at least) conceptually simpler and therefore lend themselves better to post-hoc analysis. In many engineering applications, where one has to fulfill quality and safety criteria, it can be a significant problem to analyze what the trained network is actually doing.

Accuracy of classifiers or learning algorithms I believe is dependent on the problem at hand in relation to say complexity or quality/integrity of the data. So, I don't think there is a straight answer like saying one is better than the other globally without making reference to the problem at hand. Therefore, the only litmus test is try the classifiers or learning algorithms on your specific data set and you can come out with the best for your data set. Then you would be able to set a reference point for researchers that may be interested in doing further works in relation to data set similar to yours. I wish you all the best.

Hi Faiq,

I completely agree with the answers provided by previous colleagues. I will add some more points.

As stated previously, SVMs are typically used for classification problems. In some of these problems they perform really well, in some cases better than NN, due to the fact that for constructing the decision boundary, SVMs only rely on the Support Vectors, which are the training samples that lay exactly on the hyperplanes used to define the margin. In addition, in the case of not linearly separable data, SVMs have the powerful "kernel trick", which allows to map the data to a very high dimension space in which the data can be separable by a hyperplane, almost at cost 0.

Having said that, I will also add that in problems like multivariate regression, NN are typically used in which you use a linear neuron (in the final layer) per variable you want to predict.

I also agree with previous answers, that all these concepts depend so much in each particular problem, so it is always better to design proper experiments in order to make a reliable comparison of both models.

Regards,

Carlos.

Text classification: neural networks vs support vector machines

By Waleed Zaghloul, Sang M. Lee andSilvana Trimi

 Abstract

Purpose

– The purpose of this paper is to compare the performance of neural networks (NNs) and support vector machines (SVMs) as text classifiers. SVMs are considered one of the best classifiers. NNs could be adopted as text classifiers if their performance is comparable to that of SVMs.

Design/methodology/approach

– Several NNs are trained to classify the same set of text documents with SVMs and their effectiveness is measured. The performance of the two tools is then statistically compared.

Findings

– For text classification (TC), the performance of NNs is statistically comparable to that of the SVMs even when a significantly reduced document size is used.

Practical implications

– This research finds not only that NNs are very viable TC tools with comparable performance to SVMs, but also that it does so using a much reduced size of document. The successful use of NNs in classifying reduced text documents would be its great advantage as a classification tool, compared to others, as it can bring great savings in terms of computation time and costs.

Originality/value

– This paper is of value by showing statistically that NNs could be adopted as text classifiers with effectiveness comparable to SVMs, one of the best text classifiers currently used. This research is the first step towards utilizing NNs in text mining and its sub‐areas.

.....

http://www.emeraldinsight.com/doi/abs/10.1108/02635570910957669