Lots of sensors - 2006/09/20 11:04Thinking about a typical problem that would require some kind of machine consciousness an automatic surveillance system come to my mind. Given a system with a lot of sensors (let's say hundreds or more) where the available computational power is not enough to process all information coming from them at a given time, some attention mechanism is needed to filter incoming sensing data.
The same way our (conscious) brain ignores most of the time the information coming from skin sensors, a surveillance system with many sensors would have to focus only in interesting areas. For instance, when a suspicious movement has been detected in a particular area, especial attention should be put in other sensors located close to that area. Other distant sensors reporting no relevant information might be ignored. This could be achieved by a basic attention mechanism.
Everyone is welcome to comment on this! Cheers, Raúl.
The administrator has disabled public write access. Please, register to participate in the
forum.
JoSeK
Visitor
Re:Lots of sensors - 2006/10/03 06:21Reading this, two things come to my head:
1.- Use a big cluster of computers (SSI cluster) to control all the sensors.
2.- Use datamining to study what sensor are really relevant.
But I don't know much about robots...
The administrator has disabled public write access. Please, register to participate in the
forum.
Raúl
Moderator Moderator
Posts: 591
Karma: 10
Re:Lots of sensors - 2006/10/13 08:29Hi JoSeK, thanks for your comments. As I see it, ideally you would have plenty of computational power to process all information coming from the sensors. That would be the case of a SSI cluster as you suggest. However, what I would like to discuss here is a possible strategy to deal with sensors data when not enough processing power is available (for cost reasons for instance).
This leads us to the problem of filtering or selecting data for processing. Some information has to be neglected, and some other has to be processed with the available CPU power.
The use of data mining techniques to analyse what sensor information is more relevant at any given time I think would require additional (huge? ) pre-processing power. The thing is that in such a system (automatic surveillance) the interesting data source locations are very dynamic. Taking this problem to the field of machine consciousness I’d propose an attention mechanism to drive system’s focus. The problem then is how to build such an attention subsystem. Does it make sense? Cheers, Raúl.Raúl Arrabales Moreno. conscious-robots.com/raul
The administrator has disabled public write access. Please, register to participate in the
forum.
JoSeK
Visitor
Re:Lots of sensors - 2006/10/14 07:54Raul, the problem you propose is really interesting but also abstract and difficult to solve. The human brain can pay more attention to certain "sensor" in certain moments as it has learnt to do so. I figure baby's attention mechanism is a primitive one and it grows and learns to pay attention to the most important stimulus.
How to make it artificial? The first ideas that comes to my head are related to my previous response. Datamining and data filtering. Trainning a neural network to detect what stimulus are most important at each situation seems to be a realistic solution but it generates the problem of how to generate the trainning data (or capture and label it). As you say data source locations are very dynamic but the neural network is supposed to learn this dynamism.
I can't think about other realistic solutions as a more abstract attention subsystem could be something like the "real" artificial intelligence...a holy grail search.
The administrator has disabled public write access. Please, register to participate in the
forum.
Raúl
Moderator Moderator
Posts: 591
Karma: 10
Re:Lots of sensors - 2006/10/19 07:34What sounds really interesting to me is how you have ended up relating attention and learning. It is clear that an artificial system needs to learn what is valuable to pay attention to. Trying to simplify things, learning could be involved *only* in attention focus selection rules, but not in the attention mechanism itself.
The instance JoseK mentioned about human babies, I think could be considered the ontogenesis of the attention mechanism, i.e. the basic innate mechanism is there but the rules to have it working adaptively have to be learned. Taking this idea to a simplified problem in which an autonomous mobile robot has to use the attention mechanism, the rules could be learn using reinforcement learning based on robot’s goals.
I think it would be interesting as well to find out the way to integrate the mentioned data mining techniques into the attention mechanism rule set. Data mining learning?Raúl Arrabales Moreno. conscious-robots.com/raul
The administrator has disabled public write access. Please, register to participate in the
forum.
JoSeK
User Junior Boarder
Posts: 7
Karma: 1
Re:Lots of sensors - 2006/10/19 10:26Hmmmm...Now I understand better what you want to do. I was 'obsessed' by the way of learning what we must pay attention to but as you say, the attention mechanism can be a rule set (for instance, if you need to implement a rule engine, there exist Drools, now integrated in JBoss, which is a very good open source rules engine).
This rule set can be static or dynamic (if you pretend to do the learning stage). Learning rules can be very simple as there are some Machine Learning algorithms that produce set of rules (some variations of decision trees, association rules, etc.) and you could easily integrate the new rules into the existent dataset.
I would use reinforcement learning only to modify the weight or priority of the rules.
The administrator has disabled public write access. Please, register to participate in the
forum.
Raúl
Moderator Moderator
Posts: 591
Karma: 10
Re:Lots of sensors - 2006/11/10 08:37Thanks JoSeK, I had a look to Drools rule engine. The thing is that I am now focused on rule learning for autonomous robots (using the simulation engine that comes with Microsoft Robotics Studio). I think I have to stick to C# coding for the time being if I want to take advantage of the services architecture implemented in Robotics Studio…Raúl Arrabales Moreno. conscious-robots.com/raul
The administrator has disabled public write access. Please, register to participate in the
forum.
miel
continental
Forums... 
) pre-processing power. The thing is that in such a system (automatic surveillance) the interesting data source locations are very dynamic. Taking this problem to the field of machine consciousness I’d propose an attention mechanism to drive system’s focus. The problem then is how to build such an attention subsystem. Does it make sense?
Conscious Robots RSS Feed
Follow us
