主頁
- 書單
--- 50 Key Thinkers
- 測驗
- 分類
- 我的研究
|
I'm working on a connectionist model of visual attention.
The model is implemented in C/C++, and was initially tested for basic pattern recognition and then visual search tasks.
What I'm now trying to do at this stage is to upgrade the model to achieve scale-invariance
(i.e. it can recognise a pattern even if it has different sizes.) After achieving that, I would like to test the model
with Navon figures and see if it shows any interesting behaviours.
But, before that, let me explain what connectionist approach and Navon figures are.
|
Connectionists Approach
Connectionists approach is quite popular for the past decade or so.
It is an alternative to the traditional cognitive approach of input--processing--output, the so-called
"information processing approach".
In the connectionists approach, there are less distinction between input, processing and output,
but a network of units that has connections between them, and so pass information and activation between them.
It was argued that this structure is more like the structure of brain cells, with cells spreading activities
between them by firing action potentials and so on. I personally do not think that a structure similar to brain cells
automatically gaurantees that it is better, but connectionists network nonetheless is a very powerful approach.
One major strength of the connectionists approach is its ability to learn. And its way ot learning is to
change the connections between them -- some think that it is also the way brain cells works.
In a non-psychology example, a connectionist network inside a military submarine has to decide whether
the object in front is a rock or a mine. Different frequencies reflected from the object are fed into the network,
the network would produce an output (rock or mine), and an external program or person would tell the network
if it is the correct answer or not.
If it is the correct answer, the network would reinforce the current connections, and if the answer is wrong,
the network would degrade the connections. After a certain number of trials, the network succeed in distinguishing
between rock and mine -- all without being told how to do it.
When looking into the network and see how they connect, it was found that, by learning,
the network obviously learned to associate different patterns of frequencies to rock and mine;
but, more interestingly, is that it do it by detecting whether there are metal in the object or not.
This property of connectionist network is particularly interesting for psychologists
as many behaviours involved learning of some sort.
|
Navon Figures
This is another classic. The follow figures are used by Navon in his study in 1977. Have a look and see what you think.
So what do not see? What do you notice first? Navon found that his subjects noticed the global,
the overall shape of the figure first -- that is "S" for the first two, "H" for the last two.
It was also found that one can ignore the local elements if asked just to focus on the global figure,
but not the other way round.
Of course, there are many further questions to be answered, like whether the processing of
the global figure starts first, and the local processing only starts when the global one finishes;
or do they start at the same time but the global one is faster? And are there any other factors
affecting the order and time of the process?
|
Best viewed with 1024x768 screen
Last updated: 1 Jul 2004
|
|
|
