|
Post by helgipall on Nov 1, 2012 4:05:46 GMT -5
Please post two questions each from the reading material listed for attention on the course wiki website (Oct Tue 30 + Nov Fri 2nd). I have a recent paper of my own to the reading material, "On Attention Mechanisms for AGI Architectures", available here: www.perseptio.com/publications/Helgason-AGI-2012.pdf
|
|
paolo
New Member
Posts: 13
|
Post by paolo on Nov 1, 2012 16:51:21 GMT -5
Your paper "On Attention Mechanisms for AGI Architectures: A Design Proposal" is really interesting and the topics are well explained.
1) In the section 3 "Prior work" you state that NARS implements the attention using a computational control strategy called controlled concurrency where task execution is controlled by two prioritizatioin parameters: urgency and durability. Can you explain to us how the durability paramenter works? In the same way, you state that in LIDA attention is a core process of each operating cycle, consisting of three phases: sensing, attending and action selection. Can you explain deeper the sensing and the action selection phases?
2) In an AGI systems it could happen that, while the system is performing some tasks, a new unexpected event occurs. Now, suppose that the new unexpected event (with its information) makes one of the previous tasks that the system was performing (and it is still performing) pointless and wasteful. How does the Knudsen attention framework react to that?
|
|
|
Post by fabrizio on Nov 1, 2012 17:04:21 GMT -5
After have read your paper, seems logic that an AGI requires the implementation of the "Attention".
1) How may an AGI establish what is useful and what is not? What is so important to require attention (for an AGI)?
2) From the abstract: "In the domain of intelligent systems the management of system resources is typically called “attention”." Then, is the attention (in AGI) the only process to focus resources on a task that has become "relevant"? Or is it something more?
3) How attention is linked with the Stroop effect? Considering an AGI, should there be a priority for processing data coming from a sensor (e.g. visual sensor) before data coming from other sensor (e.g. audio sensor)?
|
|
palli
New Member
Posts: 10
|
Post by palli on Nov 1, 2012 17:13:31 GMT -5
"Managing the balance between top-down and bottom-up attention, in terms of resource allocation, is part of the role of attention. [..] Balance between the two is difficult to specify in advance, as it depends on the environment and context of the system." Do you think that this will be an emergent property of the system? Is there something you have to do to influence the balance? Do you have any data on this or something to tell us?
As I see this, you are working on your goal (top-down/planning) until you are interupted by something interesting from the outside, e.g. someone calling your name. Where does task switching fit into your framework? That is where do new ideas that pop into your head come from? Or distractions, some other goals, from "within". Do the always come from the bottom-up mechanism, sub-consious(?), (but not outside) or do they fit under consious thought (top-down)?
|
|
|
Post by krummi on Nov 2, 2012 4:06:29 GMT -5
(1) "Novelty and unexpectedness are evaluated based on the operating experience of the system, data or patterns of data that have occurred before receive lower bias than previously unseen data. To accomplish this task under tight temporal (and likely also memory) constraints, it is necessary to compress prior experience of the system in some way, preferably in data structures that allow for efficient look-up and comparison."
This seems like a difficult problem to attack. How should the prior experience be compressed? What kind of data structures are we talking about? Won't the memory be exhausted pretty quickly?
(2) "Predictive capabilities: Capacity to generate predictions with regards to future expectations must be supported. Predictions are necessary control data for (topdown) attention in addition to goals."
I'm guessing this is something that has already been implemented. If so, how does this work?
|
|