"Developing AI" is a very broad concept. There are many problems in AI (planning, learning, forescasting, AI applied in robotics, etc.). But lately Support Vector Machines are displacing ANNs in classification and forecasting problems .

Artificial intelligence: an oxymoron.

Depends on what you want to do, how fast you want it to be, and how much time you're willing to spend on training.

As demonstrated by Schmidhuber and colleagues, for high-dimensional nonlinear supervised classification tasks, ANNs often do very well and it's not uncommon to see an ANN outperform specially-developed algorithms. In addition, the strength of ANNs dramatically increases as more and more layers are added. However, the problem is training time. For vision applications, for example, it is not uncommon for ANNs to require days or weeks of training (somewhat less on a dedicated GPU board). The required training time increases dramatically as more layers are added, which is why people usually restrict themselves to a small number of layers.

It is not true that ANNs are outperformed by SVM/kernel methods. Often, in studies showing SVMs outperforming ANNs, training was not carried out properly. This is another sticking point: training ANNs can be very tricky, and certain problems may require their own special training procedures. For example, on handwriting recognition, distortion of input is necessary. On other problems, noise injection (either on the inputs or on the weights) is required to acheive good convergence. Yet on other problems, initialization by, say, deep belief nets must be done otherwise the net will not find a good minimum. In choosing to use ANNs you have to take all these considerations into account.

How good are ANN for developing AI ?

I think that this question is ill-posed. If I reformulate:

What kind of learning framework is appropriate to simulate behavior of live animals ?

Then the answer is reinforcement learning: a learner tries to maximize his rewards in an unkown environnement.

There are a lot of learning techniques which are well suited for this task: Monte-Carlo, Temporal Differences, Q-Learning...

One of the simplest is the multi-armed bandit, that allows to find a trade-off between exploration (find a good policy) and exploitation (use the good policy).

To my knowledge ANN are not used for applications of reinforcement learning. However, one of the oldest rules to simulate neural network behavior is the Hebb rules. It consists in reinforce weights of neurons which have a simultaneous activation. There are also some recent papers in which the Perceptrons are used for reinforcement learning tasks (see Banditron).

Hi Raphael, I would say that you got it wrong here. The question is generalized so that no concept in AI is missed. If i targeted a specific region then the diversity of idea would be limited .Hope you understand my point of view. However , I do appreciate your point of view

Hi Heman,

I understand your point of view.

I thought your question was too broad.

By reformulating it, I could answer. I try again.

We know that biological neural networks are well suited for Natural Intelligence. We can expect that a good simulation of biological neural networks should be well suited for Artificial Intelligence.

However, we observe that if we would like to solve a practical problem, neural networks are just one of the learning machines, which we can use. Moreover, the neural networks, which is the most used, the MLP, is not plausible from a biological point of view.

My point of view is that there are two possible ways:

- find the best simulation of biological neural networks to achieve Artificial Intelligence (the Graal),

- solve some tasks of Artificial Intelligence, using learning machines.

Thanks Raphael, Indeed your answer provide a good insight on ANN..:)

Current artificial neural network (ANN) techniques focus on patern recognition and pattern completion.

Whilst quite effective at this important component of intelligent behaviour, real brains do far more than this. It may be hundreds or thousands of years before we discover and replicate the brain sucessfully, although most AI researchers for the last 40 years have been saying it is only 20 years away; although perhaps it always will be 20 years away!

We shouldn't expect a pattern recognition ANN to become conscious as it is only part of what the brain does with the sensory data. Or, as I like to put it, half a cow won't moo!

One thing that is clear to me is that is that any successful AI will be the result of allowing the brain simulation to interact with the world and learn from it be virtue of being 'situated' in a robotic body. An isolated electronic "brain in a jar", will never become intelligent. The fundamental purpose of a brain is to plan and control movement.

Perhaps the best (and the least appreciated) approach to developing true AI was done by Sir Steve Grand when he created the robot Lucy. His book is well worth reading : http://www.amazon.co.uk/Growing-Lucy-Build-Android-Twenty/dp/0753818051

Depends on the technique which you want to follow. It is a vast topic. Currently it focused on pattern recognition and comparison. You can work on brain neural networks as a future research.

i guess every one almost have the same response,

ANN compare to AI (all techniques in AI) ... it depends on what you are working on, i believe each technique in AI have their own unique to which field or area that they are suit for.

probably you can make clear about it.

ANN can't "do" AI at all! Not if you mean supervised single or multilayer Neural Networks with common transfer functions or kernels (which definition includes SVM).

The issue is one of structure, representation and organization, not just training.

We haven't even begun to scratch the surface with intelligent systems.

Loebner Prize entries, and even Watson, may have good memories.

But they don't have the intelligence of a preschooler.

BNN can "do" BI I assume! Where the B is Biological in place of A for Artificial.

So there is the promise that some future form of ANN, with a lot of help from AI/ANN researchers, and most likely a broad array of Cognitive Scientists as well, may reproduce a convincing level of intelligent behaviour.

Most importantly, the systems that get labeled ANNs today are supervised.

But BNN are unsupervised - they learn about the world/culture/language themselves.

The first step is to understand the sensory/motor aspects, the inputs and outputs.

The ANN needs to slot into a very complex dynamic real time real world interaction.