On the theoretical difficulties of telepathy.

[nhnifong]

I was reading this paper by Yoshua Bengio, and a certain passage stood out to me.

...let us ï¬Ârst focus on the mechanics of communicating good synaptic conï¬Âgurations from one brain to another. Because we have a huge number of synapses and their values only make sense in the context of the values of many others, it is difï¬Âcult to imagine how the recipe for deï¬Âning individual abstractions could be communicated from one individual to another in a direct way (i.e. by exchanging synaptic values).

I've always though that verbal communication was a painfully narrow and lossy medium, and that one day, we would have the technology to bypass it by allowing people to share more direct representations of their thoughts via neural-computer interfaces.

However, this quote casts doubt on not only that possibility, but the possibility of even plain verbal communication. Somehow, since birth, every person has inferred the meanings of tens of thousands of symbols and can use them fluently to communicate.

If we were given a new interface, with an an excessively high bandwidth, for example, the power to instantly communicate a mental image, some form of language would be necessary to represent it between brains because we would all have a different distribution of synaptic strengths and activation levels that could represent it. The burden would be on either the people communicating, to learn the shared language, or the computer system to learn each individual's unique neural structure and translate between them.

I suspect that there will be a tradeoff between the speed of communication and the time needed to learn the language used.

[Kerbface]

I believe recently there was an experiment where mice had wires attatched to their brains and were placed in identical boxes full of puzzles in different countries. The light in the second one was off. Now the first mouse was made to solve the puzzles and it's brain signals were sent across the internet to the second mouse. This second mouse completed all the puzzles in the same way in the same order.

So maybe there is more consistency between the connections in our brain than one might think.

[WestAir]

I believe recently there was an experiment where mice had wires attatched to their brains and were placed in identical boxes full of puzzles in different countries. The light in the second one was off. Now the first mouse was made to solve the puzzles and it's brain signals were sent across the internet to the second mouse. This second mouse completed all the puzzles in the same way in the same order.

So maybe there is more consistency between the connections in our brain than one might think.

Perhaps, but I would be extremely hesitant to consider a mouse brain an adequate comparator to the many different lobes of the human brain that has mastered oral communication, among other things.

Let's take yourself, and a person who speaks only Spanish as an example. Because of what we've learned from association (Pavlov's dog, anyone?) we know that whenever someone shows YOU the letter C, the letter A, and the letter T together you're going to think of a cat and picture a feline. The synapses in YOUR brain are hard-coded and wired to associate those letters (from the bundle of neurons that learned language and communication) all the way to your memory of animals (from the bundle of neurons that remembers what a cat is).

Your Spanish friend does not have the same neural connections. At all. And because any successful telepathy technology MUST be able to communicate one simple idea (a cat) to two individuals, it must be able to read and interpret the trillions of synaptic connections your brain has already created, analyze it, then correctly transmit it to the 2nd person in a decoded format their unique synaptic web can comprehend. Nightmare doesn't begin to describe the challenge here. You'd need a computer that could literally translate every neural connection of multiple brains at the same time, decode and decipher information, then re-assemble those messages in formats the receiving brain can comprehend. With computer technology that good, the human brain might as well be obsolete in comparison.

[K^2]

It's a complicated problem, but it's not as bad as all that. Thinking and communication have some additional difficulties, so lets consider a simple case. Say you have a neural network that is designed to identify printed characters. (Part of the OCR application, for example.) How does it go about that? Well, the first stage is identifying features on the image. That could be a loop, a slant line, etc. I'm oversimplifying a bit, but bear with me. Now, the top layer(s) of the network are going to consider the features and decide which character these features match the best. But to identify the features you can also use the neural net. So you end up with a multi-layered network where the lower layer(s) identify features, and higher layers translate that to a probability distribution for specific characters.

Now, suppose you have two different networks with slightly different architecture, but both are designed and trained to solve this particular task. What's going to happen is that both networks are going to learn similar feature sets. This isn't something that you'd be able to just take a look at and say, "Oh, both of these networks were trained to identify characters." But there are going to be significant correlations. So much so that I can write a third neural network that will take the feature vector of the low level(s) of the first network and derive a feature vector for the top layer(s) of the second network, so that the second network can still identify the characters seen by the first network.

This is not entirely trivial, and it will require some degree of training, but the nice thing about it is that you don't have to re-train any of the networks that are communicating. It's the translating network that's going to do all the learning and adjustment. And now the parallels with communication between two individuals should be obvious. Our thought process is also based on various abstract features. These can be rather complex and nested. You use individual strokes to identify letters on the page, but you also use the more abstract notion of the letter to identify a page of printed text even if you don't recognize the language or the script. But in the end, you are still translating features. And when you communicate verbally, you translate these into sentences. Well, so long as feature sets are similar enough, that is, so long as two individuals have similar cultural experience, you should be able to devise a neural network that rides between the thought and speech and translates abstract notions into something another person's brain can interpret almost directly, bypassing speech processing.

[Brotoro]

...I've always though that verbal communication was a painfully narrow and lossy medium, and that one day, we would have the technology to bypass it by allowing people to share more direct representations of their thoughts via neural-computer interfaces.

However, this quote casts doubt on not only that possibility, but the possibility of even plain verbal communication...

I think that Yoshua Bengio is probably correct in that the synaptic configurations of the abstractions held in each brain are likely to be quite different (because things get stored in there in weird overlapping ways...some thoughts in my head consistently trigger other unrelated memories, so unrelated data are obviously held in overlapping storage)... but I don't think that means your hope for transmitting verbal or visual information between brains is hopeless. All of our brains get the same kinds of input from the auditory and optic nerves (even if they do end up storing the results in different internal configurations), so I think that direct communication between brains will have to interface to the parts of the brain that handle those standard inputs. It should be easier with communication of "verbal" thoughts since we also have parts of the brain that take our abstract thoughts and put them into words...which can then be fed into the auditory input areas of the receiving brain (just bypassing the flappy tongue and vibrating ear parts). Visual images will be harder, since we don't process visual output in the same way we process language into speech output.