Ask HN: Can a neural network where the neurons are humans become AGI?
* The idea is that you can make an AGI out of a network of humans. The individual humans within the network needn't know how they are contributing to the overall intelligence of the AGI.
* The communication between neighboring humans in the network can happen via sending emails, over hand help radios, yelling or even dockerized containers. The optimization algorithm determines to whom these messages are sent and with what "weight".
* The inputs and outputs need not be trivial, they can be arbitrarily complex, such as numbers, theorems, songs or questions being posed and answered. These inputs and outputs need only be "differentiable" either numerically or even in a "fuzzy" sense in order to allow the communication to be optimized.
* An example could be a neural net of mathematicians who together produce a better mathematician who is able to solve a problem no one single mathematician within the network is able to solve. This neural net of mathematicians could be better than a group of mathematicians discussing amongst themselves. In a group discussion, human dynamics come into play, where one of the mathematicians becomes the leader and needs to act as a centralized consolidator and arbiter of information. Whereas in the neural net the algorithm/weights determines how the mathematicians communicate with each other.
[1] https://news.ycombinator.com/item?id=27738029
[2] https://en.wikipedia.org/wiki/China_brain
4 comments
[ 2.7 ms ] story [ 13.3 ms ] thread* Another example could involve a neural net of software developers trying to build an application. Where over many iterations of building several applications, the algorithm learns that certain sub tasks are best given to certain developers (i.e. neurons) and certain sub tasks need to be given to multiple developers since it might be a difficult or an error prone problem to solve.
* The neural network architecture needn't be limited to just MLPs, they can be recurrent, convolutional or anything else.
* Obviously the problem I am unable get around is how the inputs and outputs are to be defined for given problem. Given that the neurons are humans who can solve quite general problems themselves, the inputs and outputs needn't be limited to just numbers.
* The optimization algorithm determines to whom these messages are sent and with what "weight".
To state some of the obvious:
- A neural net of just one human is AGI
- A company or any organization of humans can also be considered to be AGI. However the focus of my question is whether we can capitalize on the recent developments we've made in the neural networks architectures and optimization techniques and apply it to a neural net of humans.
I understand that the information is different or more complex. But so long as this information follows some predefined rules then it lends itself to optimization algorithms such as root finding and gradient descent.
For example in an MLP, a common activation function that is used is a sigmoid function whose values falls in the range of [0, 1] and the function is differentiable and therefore you can run backpropagation to train the MLP.
Similarly in a human-MLP, the activation function is a human who is capable of understanding (input) and producing (output) multiple kinds of complex information types. Now let's say the humans produce a payload/output and the algorithm passes on this information onto the humans in the next layer based on the weights between the humans.
To give you a real example of how it can be used, let's used a modified version of the game of charades. Imagine there's a 2x2 layered human-MLP i.e. 2 humans in layer 1 and 2 humans in layer 2. There are no teams and there's one moderator who whispers the word to draw to the humans in the first layer. It is the job of the humans in the second layer to guess correctly.
* The humans in the first layer each draw the image of the word with their own artistic style.
* The algorithm decides depending on the weights and some randomness which image goes to which human in layer 2.
* The humans in the second layer receive this images anonymously and they make a guess.
* The algorithm decides what the final output is after weighting the outputs from the humans in layer 2.
- How does backpropagation take place here? I don't know. To begin with we can use Word2Vec to determine magnitude of error in the guess vs the right answer.
- How do you find the derivative of the human's activation wrt to the weights? I don't know. But I'm sure you can come up with interesting new strategies for it.