I have been lately researching and working with ANNs and I found that, despite of being used widely on our machines, we are unable to implement such networks in other paradigms: at a molecular level and so on. While in this case Boolean Networks show a promise.
Though this question may make no sense at first, it is perhaps an interesting problem to consider:
How to convert a neural network to boolean network?
Simply constructing a neural network by connecting basic logic gates AND, NAND, NOR, XOR, XNOR, YES et cetera.
I want to research on the topic, kindly help me as to where should I begin with because I am surely missing some important keywords here.
[Kindly disregard the training/learning process]
Thanks in advance
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
3
I'd recommend starting by looking at artificial neural nets where the units already take binary values. For example, check out Hopfield nets and the McCulloch–Pitts model (the original ANN). The Ising model may also be of interest. Although it was originally a model of ferromagnetism, it has connections to neural networks, and is of interest in neuroscience and physics.
Although the units in these models take binary values, the connection strengths are continuous, so implementing them with digital logic would require something like replacing each unit with a local network of logic gates. This will only be an approximation because real numbers can't be truly represented using a finite sequence of discrete operations--almost all of them are uncomputable. Of course, this also applies to digital computers, which we use to approximate computational models (e.g. ANNs) involving real numbers.
answered Jul 30, 2016 at 6:26
-
$\begingroup$ Hi @user20160, what I am facing problem is the "thresholding mechanism". How to bit blast Sigmoid function to a complex and intricate network of logic gates. In fact, how should I represent binary in first place? And then thridly, how to pass the output signal of neuron like state machine to all other neurons? $\endgroup$ Jul 30, 2016 at 6:47
-
$\begingroup$ That's part of the motivation for using simple units like Hopfield neurons. They have binary outputs, so just run a wire from the output of one unit to the input of the next. Also, no sigmoids necessary, so you can use standard digital logic circuits like adders (en.wikipedia.org/wiki/Adder_(electronics)). There are multiple standard ways to represent numbers in digital electronics (en.wikipedia.org/wiki/Binary-coded_decimal). Fixed-point representations will be much easier than floating point. $\endgroup$ Jul 30, 2016 at 7:14
-
$\begingroup$ Hi, Does ising model require some activation function? Could you please give me links for the same to study ising model for this project! Thanks in advance. $\endgroup$ Aug 4, 2016 at 6:49