What happened to associative learning?

<< Where am i? | Home | Missing Docs - again! >>

What happened to associative learning?

With so much research being done on Artificial Intelligence and learning, it is striking that all research seems to concentrate on reinforcement learning, while research on associative learning almost doesn't exist.
When you do find research on associative learning, it seems to be reinforcement learning in disguise.
What's up with this?

The difference between associative learning and reinforcement learning

Reinforcement learning means that an agent is rewarded for doing something right and punished for doing something wrong. In Artificial Intelligence, an evaluation function is used to determine how good the result of an action was. Usually as a result a neural network is adjusted.
Associative learning is based on making associations between things happening. When the sky is dark, we know that it's going to rain. Why? Because we have seen dark sky and rain together many times, and have associated dark skies with rain.

Pavlov's dog

Everyone knows the experiment of Pavlov on dogs. When he fed his dog, he sounded a bell. The dog made an association between the bell and the food. when ever he heard the bell, saliva started dripping from his mouth in anticipation of the food.

Associative learning in computers

Why is there almost no research to associative computer learning? This is the way we train our own brains. It is very close to the essence of how we think.
It seems to me that it can not be very hard to implement this: Just monitor the parts of the computer system, and when two parts are activated at the same time, create a link between the two, and make this link stronger every time this occurs.

This would be especially practical in swarm intelligence systems like White et al. describe. You could build a special swarm that does just this: Look at the energy levels and create links between places with high activity by leaving trails. Then in the future, when one gets activated, the other gets activated too. Furthermore, since the trails go away gradually, the association will be broken once the two don't get activated at the same time anymore.

Have AI researchers always just overlooked this simple possibility, or is there something fundamentally wrong with my reasoning?
Would my proposed system work?

Print posted @ Sunday, February 08, 2004 8:13 PM

Feedback

# re: What happened to associative learning?

Associative learning sounds a lot like Bayesian techniques used in AI. 2/9/2004 12:12 AM | Wesner Moise

# re: What happened to associative learning?

As far as i understand it i have never used baysian networks, a bayesian network is based on reinforcement learning with a feedback loop. The evaluation function is based on the feedback loop: Is the predicted result the same as the actual result? So we try to predict something, then see if the prediction is corect, and learn purely based on that.
Maybe one could argue that associative learning is a pure consequence of reinforcement learning, because reinforcement learning can create associations. But i don't agree: I think there is another mechanism at work in associative learning. 2/9/2004 6:53 AM | Henk Devos

# re: What happened to associative learning?

I'm glad I found someone mentioning this....

I've been researching in associative learning and was wondering why nobody tried to implement it yet....so I decided to do it myself....

I currently have a somewhat silly version of a Pavlovian Dog which I just use to demonstrate the idea....but it's not really a sound implementation....

I'm trying to figure out a way that would be both reliable and efficient....especially when the number of states can be infinite (like in most real-world systems).

Maybe we can work together on this....I'm still reading around....any ideas would be great.

mak@aucegypt.edu 1/31/2005 2:22 PM | mak

# re: What happened to associative learning?

check out the above link 1/31/2005 2:25 PM | mak

# re: What happened to associative learning?

I don't see things the same way. Reinforcement learning has to do with attaining goals, sort of optimization problems, while Associative Learning has to do with correlating things which usually implies no goal whatsoever. Actually both can work together, for example having AL for correlating actions and the respective state transitions they bring about, while at the same time, having some RL for the agent to pursue a global goal. What worths having a dog associate the bell with the food if the dog doesn't seek being fed. 2/2/2005 1:24 PM | Rodrigo da Silva Guerra

# re: What happened to associative learning?

would love for someone to share insights on associative learning .Am trying to do some work based on that. 8/24/2005 10:29 PM | neena george

# re: What happened to associative learning?

Curiously, we had been working on such things much in the last 4 years. Take a look at the work of B. Porr on our "Publications" site, where there is quite some stuff on How to Achieve a Goal with associative learning. In Neural Comp. Feb 2005 we wrote a big review on such topics (Called "Temporal Sequence Learning....") 9/12/2005 12:18 PM | F. Woergoetter

# re: What happened to associative learning?

This is a topical area of great interest to me and my
colleagues. We are in the process of developing a rather large educational resource that relies on an associative informational architecture. Basically, we began by asking two very fundamental questions "how do you know what you know?" and "how do we learn best?"

The answer to the first question is deceptively simple. We know what we know because we are able to create mental links or associations between concepts, sensory inputs, data, etc.with existing reference points. Simple enough.

What we might call traditional learning paradigms rely on a linear learning process or construct. Think in terms of a university text book. Chapter one lays out fundamental concepts, theories or basic building blocks. Chapter two builds on those building blocks and following chapters go into more detail about specific topics. For example, if you wanted to learn about a subject you might check out a library book and skim through chapters until you had a feel for the topic (developed enough associations)and then dig into a particular topic.

Today we live in a search and learn world. In short, we learn through sifting through a large volume of sources, taking in discreet or small amounts of information until we find something that meets our needs. As we go through this process we are able to construct larger meaning out of these small bites of information. There has been some meaningful research into this sort of search and learn process. Some have borrowed a term from manufacturing (just in time production) and applied a new twist by calling this process just-in-time learning. Just in time learning can be thought of as the acquisition of knowledge or skills as they are needed.

To support this sort of learning, information can be better organized by anticipating the sorts of questions or associations that a user/learner may have. Most internet content is still organized in a fairly linear fashion. News sites often include a few hyperlinks to related stories (associations), and while this is a start, it still fairly primitive. The primary reason that better information support does not take place is simply that it is very, very time consuming.

Now, the 2nd question "how do we learn best?" The short answer is through the use of newly acquired information and through exploration. The more senses that can be involved the better (think neuropathway stimulation). A prime example is learning a foreign language. Language learning requires memorization of vocabulary words, sentence structure/patterns,listening, repeating and practice. But in addition, exploration and experimentation is a critical part of the learning process.

At www.3DAuto.com we've incorporated the sort of associative learning process mentioned above through the extensive use of hyperlinking related terms, phrases and concepts. In addition, we have developed an extensive catalog of animations,3d models, interactive models, photos and video.

How does all of this apply and why is it important?
Well, let's assume that you want to learn about an automobile brake, or perhaps you think you may have a problem with the brake system. But let's assume you don't know much about a drum brake. Traditionally you might read a book, see a photo or an exploded diagram and from those references begin to understand how a drum brake works. From the descriptions and diagrams you could create a mental picture of how the drum brake assembly might work. But what if you could watch an animation of a drum brake in action? Perhaps it would be helpful if you could hold one in your hand or disassemble the brake assembly? Maybe you'd like to be able to look at a large photo gallery of various drum brake assemblies and individual parts. Perhaps it might also be helpful to read about the sort of problems that are most common with drum brakes or even see symptoms of problems....I think you get my point.

This is heavy lifting from a development perspective. It is extremely time consuming, exhausting and financially taxing. However, it is our shared belief that in the near future there will be a substantial divide emerge between the current generation of online information structure/learning processes and a much more user-centric model.

In closing, I think associative learning is going to have a big impact on how we all learn in the future. Associative learning creates a big umbrella under which a great deal of exploration, experimentation and development will take place. Appreciate feedback, thoughts or simply drop by and see what we're up to.

Cheer and all the best to all. 3/27/2006 5:29 PM | cjones

# re: What happened to associative learning?

You mention that associative and reinforcement learning are distinct, but this does not seem so. Everything has a consequence, whether this be immediately or delayed. In the Pavlovian dog situation, the reward and association are presented at the same time. However, in your example of the thunder strom and darkness the darkness precedes the consequence of being rained upon. Therefore, if the darkness is sufficient to make to notice it, it will have consequences on your behaviour to endeavour to not get rained upon. Essentially both situations provide a consequence of the associated stimulus, thereby changing behaviour,

So in your view, does the difference lye in the temporal dynamics of the consequence and association?

10/5/2006 7:04 PM | Rich

# re: What happened to associative learning?

very good 5/26/2008 9:26 PM | acer battery

# re: What happened to associative learning?

I am just a mother doing my own research and I'm wondering with my 8 yr old ADHD son if I can combine the AL and the RL to condition his cognitive behaviors and achieve a better "therapy" for him... any ideas?? 11/4/2008 4:57 PM | Tiffany Volore

# re: What happened to associative learning?

Associative learning (AL) has far more implications then reinforced learning(RL). AL is a superset of RL. I see AL as providing intellignce to link data sets in multi-directions while RL is a one-way relation of data sets. A==>B or A|B==>C. reinforcement is simply applying a direction to the association of data sets. I see RL as a model created to help us realize the benefits of predictive intelligence but at the heart of it is still associative learning.

If we can solve associative learning, RL should follow easily behind. 2/5/2009 5:27 PM | Rajeev Nischal

Henk Devos