Wednesday, April 08, 2009

Action-selection and Attention-allocation: a common problem and a common solution?

I have recently blogged a bit about action-selection and operant learning, emphasizing that the action one chooses, out of many possible, is driven by maximizing the utility function associated with the set of possible actions, so perhaps a quick read of last few posts would help appreciate where I come from .

To recap, whenever an organism makes a decision to indulge in an act (an operant behavior), there are many possible actions from which it has to choose the most appropriate one. Each action leads to a possibly different Outcome and the organism may value the outcomes differentially. this valuation may be both objective (how the organism actually 'likes' the outcome once it happens, or it may be subjective and based on how keenly the organism 'wants' the outcome to happen independent on whether the outcome is pleasurable or not. Also, it is never guaranteed that the action would produce the desired/expected outcome. There is always some probability associated that the act may or may not result in the expected outcome. Also, on a macro level the organism may lack sufficient energy required to indulge in the act or to carry it out successfully to completion. Mathematically, with each action one can associate a utility U= E x V (where U is utility of act; E is expectancy as to whether one would be able to carry the act and if so whether the act would result in desired outcome; and V is the Value (both subjective and objective0 that one has assigned to the outcome. The problem of action-selection then is simply to maximize the utility given different acts n and to choose the action with maximum utility.

Today I had an epiphany; doesn't the same logic apply to allocating attention to the various stimuli that bombard us. Assuming a spotlight view of attention, and assuming that there are limited attentional resources, one is constantly faced with the problem of finding which stimuli in the world are salient and need to be attended to. Now, the leap I am making is that attention-allocation just like choosing to act volitionally is an operant and not a reactive, but pro-active process. It may be unconscious, but still it involves volition and 'choosing'. Remember, that even acts can be reactive and thus there is room for reactive attention; but what I am proposing is that the majority of attention is pro-active- actively choosing between stimuli and focusing on one to try and better predict the world. We are basically prediction machines that want to predict beforehand the state of the world that is most relevant to us and this we do by classical or pavlovian conditioning. We try to associate stimuli (CS) with stimuli(UCS) or response (UCR) and thus try to ascertain what state of world at time T would be given that stimulus (CS) has happened. Apart from prediction machines we are also Agents that try to maximize rewards and minimize punishments by acting on this knowledge and acting and interacting with the world. There are thousands of actions we can indulge in- but we choose wisely; there are thousands of stimuli in the external world, but we attend to salient features wisely.

Let me elaborate on the analogy. While selecting an action we maximize reward and minimize punishment, basically we choose the maximal utility function; while choosing which stimuli to attend to we maximize our foreknowledge of the world and minimize surprises, basically we choose the maximal predictability function; we can even write an equivalent mathematical formula: Predictability P = E x R where P is the increase in predictability due to attending to stimulus 1 ; E is probability that stimulus 1 correctly leads to prediction of stimulus 2; and R is the Relevance of stimulus 2(information) to us. Thus the stimulus one would attend, is the one that leads to maximum gain in predictability. Also, similar to the general energy level of organism that would bias as to whether, and how much, the organism acts or not; there is a general arousal level of the organism that biases whether and how much it would attend to stimuli.

So, what new insights do we gain from this formulation? First insight we may gain is by elaborating the analogy further. We know that basal ganglia in particular and dopamine in general is involved in action-selection. Dopamine is also heavily involved in operant learning. We can predict that dopamine systems , and the same underlying mechanisms, may also be used for attention-allocation. Dopamine may also be heavily involved in classical learning as well. Moreover, the basic computations and circuitry involved in allocating attention should be similar to the one involved in action-selection. Both disciplines can learn from each other and utilize methods developed in one field for understanding and elaborating phenomenon in the other filed. For eg; we know that dopamine while coding for reward-error/ incentive salience also codes for novelty and is heavily involved in novelty detection. Is the novelty detection driven by the need to avoid surprises, especially while allocating attention to a novel stimulus.

What are some of the prediction we can make form this model: just like the abundant literature on U= E x V in decision making and action selection literature, we should be able to show the independent and interacting effects of Expectancy and Relevance on attention-grabbing properties of stimulus. The relevance of different stimuli can be manipulated by pairing them with UCR/UCS that has different degrees of relevance. The expectancy can be differentially manipulated by the strength of conditioning; more trials would mean that the association between the CS and UCS is strong; also the level of arousal may bias the ability to attend to stimuli. I am sure that there is much to learn in attention research from the research on decision-making and action-selection and the reverse would also be true. It may even be that attention-allocation is actually conceptualized in the above terms; if so I plead ignorance of knowledge of this sub-field and would love to get a few pointers so that I can refine my thinking and framework.

Also consider the fact that there is already some literature implicating dopamine in attention and the fact that dopamine dysfunction in schizophrenia, ADHD etc has cognitive and attentional implications is an indication in itself. Also, the contextual salience of drug-related cues may be a powerful effect of dapomine based classical conditioning  and attention allocation hijacking the normal dopamine pathways in addicted individuals. 

Lastly, I got set on this direction while reading an article on chaining of actions to get desired outcomes and how two different brain systems ( a cognitive (Prefrontal) high road one based on model-based reinforcement learning and a unconscious low road one (dorsolateral striatal) based on model-free reinforcement learning)may be involved in deciding which action to choose and select. I believe that the same conundrum would present itself when one turns attention to the attention allocation problem, where stimuli are chained together and predict each other in succession); I would predict that there would be two roads involved here too! but that is matter for a future post. for now, would love some honest feedback on what value, if any, this new conceptualization adds to what we already know about attention allocation.

Sphere: Related Content

6 comments:

Ward said...

Another fine post. I appreciate your big picture thinking of trying to draw different research avenues together.

thanks

Mariana Soffer said...

Very interesting.
I was thinking if this could also apply to where we direct our mental awarnes focus.
Meaning: what goes on in our head can have the same mechanism of selection. In what are we focusing our internal attention when we are thinking (for example about our past).

kim said...

How dare you! I'm browsing my regular RSS feeds in an attempt to avoid some work on attentional optimization in a category learning task, and here you blog about the very thing I'm trying to avoid!

Our study was conducted with an eye-tracker as participants learned to classify images into four categories. Although all three features were relevant to do the entire task, only two were relevant to any given category. Quick eye-movements don't seem all that costly, though additional (irrelevant) info is uninformative, so do people bother restricting their attention for relevant information only? We found, essentially, that people don't bother looking at features of a stimulus that are not important for making a category decision after they have successfully learned a category and stopped making performance errors.

My colleagues are interested in how attentional optimization (characterized by changes in allocation) can be incorporated into models of category learning. Categorization is a great way to study attentional learning in a lab setting, I think, because it's a very easy way to force people to consider emphasizing information that is important, and ignoring information that is unimportant, irrelevant, distracting, or even misleading. I'm interested in how one's definition of an "optimal" pattern may change, based on the participant's goals -- I think the results we found may imply, at least under our conditions, that an effort (or cost) goal is being satisfied only after an accuracy (or benefit) goal is reached. But would this be different if we emphasize speed? (Though undergraduate students who are participating in research for credit are probably emphasizing speed goals quite a lot, given they probably just want to leave the experiment and go home as soon as possible.) And how would varying the predictability of the features affect optimal patterns?

I'm more familiar with the category learning & (visual) attention literature than I am about the things you've mentioned.. so feel free to blog more about it! ;) In categorization research, the closest thing to what you have described -- predicting where the brain will allocate attention -- is going to be in the modeling literature I believe. Rehder & Hoffman, John Kruschke, and Brad Love are some names I can think of off the top of my head if you are interested in computational models.

But outside the realm of category learning, and speaking more directly to your ideas of prediction and anticipation, I recently heard about some eyetracking experiments in Matthew Crocker's lab with linguistic stimuli. Participants heard sentences (in German) while viewing a scene, and they made anticipatory eye movements to areas of the screen based on the sentences (they were essentially predicting what words would come next). I believe they created a model based on a simple recurrent network to predict this kind of attentional allocation.

Anyway, enough with my procrastinating ramblings -- thanks very much for the links, and any more posts on attention allocation are very welcomed!

Sandy G said...

Ward, Mariana thanks a lot.
Mariana,
There is merit in what you say. While action-selection can be roughly equated to the motor side of things and attention-allocation to sensory side of things, we can have a cognitive side of things whereby awareness of internal thoughts, motives, beliefs, desires (meta cognition)etc is constricted and the problem of what thoughts/ beliefs etc to bring into awareness and ruminate on becomes a similar problem. However, to start with, I would just focus on attention as related to external sensory cues to make things less complicated and more intuitive.

Sandy G said...

Kim, thanks for that elaborate response. Conversations and feedback like these make the whole effort of blogging worthwhile.
The studies you are doing seem very promising.

Initially when I was writing this post I did consider the problem of feature binding and what features one would attend to in order to identify (or in your case categorize) the object given the fact that there are multiple relevant and irrelevant features available to attend to; I think your experimental setup provides an excellent setup to test some of the ideas I propose. Suppose there were three categories, and two categories were close (required 2 features each to categorize) and differed only on the basis of one of the feature value (say having a dark boundary). Also suppose that finding an object of one of these category entails a monetary reward and thus makes importance of that feature that distinguishes that category from the others more Relevant/ important to the subjects. Now after learning the task, when a novel object is presented would attention be shifted to stimulus/feature that differentiates between the two categories, one of which is a desired category. Maybe I am sounding dense , but the idea is to test whether relevance of the distinguishing feature affects attention allocation to that feature while categorizing. Maybe you can drop me an email at sandygautam[AT]yahoo[DOT]com and we can discuss in detail.

A similar experimental setup can be devised to gauge the effects of feature-category binding (expectancy of stimulus 1 predicting stimulus 2 in my model) on attention allocation. Will attention be shifted to features relevant to the category that is more deterministic and not fuzzy and has been learned well.

I'm very excited and look forward to a collaboration/ brainstorming on the same. do let me know your email id/ start a conversation.

Meanwhile I will look up the modeling literature you have pointed me to. Many thanks for the comment!!

Mariana Soffer said...

T hank's for your encouragement Sandy, you might want to check my website that talks about this kind of subjects, I think you might enjoy it.

Take care
M