I'm looking at neural nets from a biological perspective, so if no one knows an answer, please direct me to another forum that would have an answer. I'm wondering how are our brains able to infer something about one dataset given another entirely different dataset. For example (this is going to sound really weird, but oh well): I went through the drive through at Taco Bell at lunch, and I ordered my food. Then I drove forward to the first window, paid, and I drove to the next window and picked up my food. Both people at the windows were guys. I started wondering which of the two I had talked to via the microphone when I ordered my food - the one at the first window or the second. Then today I went through again. This time, I picked up my food at the second window and before I started driving away I heard the guy say something like, "Hi, can I take your order?" and I knew he was talking to someone preparing to order, not me. So the question is: how was my brain able to infer that the person at the second window was the person that had taken my order both times? I'm thinking that because of things like this, the neural network in the human brain must be composed of a series of subnetworks that "talk" to each other. Each subnetwork has a set of neurons. I think the subnetworks function similarly to either Kohonen networks or ART networks in that they are able to self-organize and classify information based on content (associative memories). Then I think the different subnetworks are linked together via their individual nodes. More specifically, I think each subnetwork applies to a different type of knowledge. For instance, one might store the characteristics of every human face I remember, while another subnetwork might store the sounds of their voices. What do you guys think of this idea? Have you heard of this before, and do you know of any links that might have something about this? [Edited by - chadjohnson on November 3, 2005 3:04:09 PM]

Quote: Original post by chadjohnson
So the question is: how was my brain able to infer that the person at the second window was the person that had taken my order both times?

There's actually a multi-stage process going on here involving both deduction and induction. On the first time through the drive-through, you obtain certain evidence about the entities in the environment and the relationships between them. However, you don't have enough information to answer the question "which entity was I talking to". On the second time through you obtained evidence about which entity takes orders. You overheard one of them talking and could reasonably deduce that they were an order taker. However, this doesn't mean that the person in that window took your order the day before. However, many of the entities and relationships were similar between the two experiences (or at least the ones you believe relevant to the problem at hand... more on this in a moment) so you can make an induction that the second window always takes orders and then, from this infer that the day before, your order was taken by the second window.

Now, here's a few things to consider. If on any given day you receive evidence that the second window doesn't take the order (so you overhear the first window guy doing it) then your confidence in your inference about that first experience must be lessened. In fact, if you are a logical agent, your confidence in that inference about the first experience should be directly proportional to the support lent that inference by the observed evidence.

Now, your induction might also be flawed. Let's assume for a moment that in fact only one person working at the store is trained to take orders. However, they may be able to do so from either of the two windows. Subsequently, any latter observation of the location of the person when they took the order does not tell you about the location of the person during that first experience of them taking your order. The best you can do is observe many such instances and estimate the likelihood (probability), based on frequencies of events, that they were in the first or second window during that first event.

Now, does the brain do this? Well, yes and no. Does it do it with associative memory? Yes and no. It is fairly evidenct that certain regions of the brain, such as the hippocampus (in each temporal lobe) are made up of associative memory (AM). There are some very good AM models out there that predict limited hippocampal functionality well. However, the inference tasks described above require reasoning skills usually produced by the frontal lobe, which is not simply associative memory (indeed, there's probably little 'memory' in the frontal lobe... the temporal lobes seem to house working and long term memory). There are good models of frontal lobe architecture based on ANNs, although here it is more about the functional relationship between ANNs than the ANNs themselves.

Thus, the tasks described above would involve a complex interplay between the frontal lobes (which are forming hypothesis and testing them against memory) and the temporal lobes, which are storing the memories (evidence) and offering up associations of memories, based on signals received from the frontal lobes.

One things is reasonably clear though from research: we don't store memories of events as whole pictures. Rather, we store elements of a sensory event and functional relationships between them which can be recalled to form a model of the event that occured.

Quote:
I'm thinking that because of things like this, the neural network in the human brain must be composed of a series of subnetworks that "talk" to each other.

The problem is that it is not this clear cut. Yes, there are certainly sub-networks, but they aren't the same sort of sub-networks as you'd find in say a power grid, or a computer network, where relationships between the elements are fixed in the wiring. Rather, they are regions that have correlated function, where relationships and correlations depend on both the specific data being fed to the network and the time history of data. That is, the structure of the network (and hence what could be considered a sub-network) depends on the data activating that region of the brain and the propensity to activate certain regions in certain ways is conditioned over time (although there is certainly an amount of pre-conditioning determined by genetics, since we can look at individuals (using say fMRI) and see similarities in their activation regions for the same task).

If you want to read up on this sort of thing, then I would suggest looking at the literature on neuropsychology.

Cheers,

Timkin

Quote: Subsequently, any latter observation of the location of the person when they took the order does not tell you about the location of the person during that first experience of them taking your order.

Ahh, so if one piece of knowledge becomes flawed, any subsequent knowledge that directly relates to that knowledge is essentially flawed as well. But, this subsequent knowledge still does get recorded; you're just unsure about it.

So if on another day I were to find out that the person in the second window doesn't or doesn't always take the order what would happen to the connections previously made in my brain? Would they be inhibited or lessened, or would new connections be made which would replace the existing connection, and the old ones just wouldn't be used?

So do you think it's possible to model a human brain via software (ignoring the hardware requirements)? Do you think a functional brain model (just for storing and gaining knowledge) would require a simulation of the different physical parts of the brain, such as the frontal lobe and the temporal lobe?

Quote: We don't store memories of events as whole pictures. Rather, we store elements of a sensory event and functional relationships between them which can be recalled to form a model of the event that occured

That makes sense. That's why when we recall images we can't visualize them perfectly. We must just store the main characteristics, then we fill in the gaps.

Quote: The problem is that it is not this clear cut. Yes, there are certainly sub-networks, but they aren't the same sort of sub-networks as you'd find in say a power grid, or a computer network, where relationships between the elements are fixed in the wiring.

I completely agree with you. The subnetworks should not be clear cut. For example, if someone lived in the jungle all their life and never saw another human, they would not have a subnetwork dedicated to human faces or their voices. Then if one day that person was to meet another human, they would probably try to understand the idea of another person by comparing the person's characteristics with existing knowledge. Eventually over time, as the person interacted more and more and got used to the idea of another human, subnetworks would probably form relating to the other human. What do you think? Do you agree though that each subnetwork would have to store a different type of knowledge, maybe like the tables in a relational database?

Quote: Thus, the tasks described above would involve a complex interplay between the frontal lobes (which are forming hypothesis and testing them against memory) and the temporal lobes, which are storing the memories (evidence) and offering up associations of memories, based on signals received from the frontal lobes.

So essentially, the frontal lobe acts as an inference engine while the temporal lobe acts as a knowledge base. Short-term memory must simply consist of a set of variables. And basically, the temporal lobe must be a self-organizing network of self-organizing networks.

Thanks for your response.

[Edited by - chadjohnson on November 4, 2005 12:03:28 AM]