nvm implements a Neural Virtual Machine (NVM). This is a neural network that emulates a symbolic machine using distributed representation and local learning. Human-authored assembly programs for the symbolic machine can be represented and executed by the neural network.
nvm has been tested using the following environment, but it may work with other operating systems and versions.
- Fedora 29
- Python 2.7.15
- numpy 1.15.4
- scipy 1.0.0
- matplotlib 2.2.3
- Clone or download this repository into a directory of your choice.
- Add the
srcsub-directory to your PYTHONPATH.
First, decide on the registers you want for your NVM instance. For example:
>>> register_names = ["r0", "r1"]
Next, write some NVM assembly programs for the instance. Each program should be given a name and stored in a dictionary with its name as the key. For example, the following program implements and invokes a sub-routine that computes logical-and of the two register contents, using the primitive comparison instruction:
>>> programs = {
... "myfirstprogram":"""
...
... ### computes logical-and of r0 and r1, overwriting r0 with result
...
... nop # do nothing
... sub and # call logical-and sub-routine
... exit # halt execution
...
... and: cmp r0 false # compare first conjunct to false
... jie and.f # jump, if equal to false, to and.f label
... cmp r1 false # compare second conjunct to false
... jie and.f # jump, if equal false, to and.f label
... mov r0 true # both conjuncts true, set r0 to true
... ret # return from sub-routine
... and.f: mov r0 false # a conjunct was false, set r0 to false
... ret # return from sub-routine
...
... """}
Now we can construct an NVM instance to run this program. We will use a convenience method that automatically scales the network size to accommodate the program.
>>> from nvm.nvm import make_scaled_nvm
>>> my_nvm = make_scaled_nvm(
... register_names = register_names,
... programs = programs,
... orthogonal=True)
The orthogonal=True flag uses orthogonal activity patterns to represent symbols, which greatly reduces the size requirements. Now we can assemble the program into the NVM instance, and load it so that it is ready to be executed. When loading we can also specify initial activity for each register when execution begins.
>>> my_nvm.assemble(programs)
>>> my_nvm.load("myfirstprogram",
... initial_state = {"r0":"true","r1":"false"})
The underlying representations for true and false are distributed activity patterns that we can access from the net field of the nvm instance. Each pattern is a numpy array.
>>> my_nvm.net.layers["r0"].shape
(8, 1)
>>> my_nvm.net.activity["r0"].T
array([[ 0.9999, -0.9999, 0.9999, -0.9999, -0.9999, 0.9999, 0.9999,
0.9999]])
Each layer has a coder we can use to look-up the human-readable symbol represented by a pattern:
>>> v = my_nvm.net.activity["r0"].T
>>> my_nvm.net.layers["r0"].coder.decode(v)
'true'
The nvm instance also has a convenience wrapper for this:
>>> my_nvm.decode_state(layer_names=register_names)
{'r0': 'true', 'r1': 'false'}
The program is already loaded; all we need to do to emulate it is run the network dynamics until an exit opcode is reached:
>>> import itertools
>>> for t in itertools.count():
... my_nvm.net.tick()
... if my_nvm.at_exit(): break
...
The invocation my_nvm.at_exit() is a convenience wrapper for:
>>> (my_nvm.net.layers["opc"].coder.decode(my_nvm.net.activity["opc"]) == "exit")
True
We can see how many time-steps of network dynamics were used:
>>> t
106
We can check the final register states now that the program is finished:
>>> my_nvm.decode_state(layer_names=register_names)
{'r0': 'false', 'r1': 'false'}
r0 was correctly overwritten with the logical-and of the initial register values, true and false.
We can test other implemented functionalities by running below 2 different tests from terminal:
python tests.py
python dsst.py
The following scripts in the src/experiments directory were used to generate experimental results and figures for the 2019 NVM paper:
rehebbian_experiments.py(Figure 9)rehebbian_experiments2.py(Figure 10)rehebbian3dproj.py(Figure 11)list_program.py(Figure 12)random_program.py(Figures 13-15)
The repository has been improved since that paper was submitted, but the code state at time of submission is preserved in release v1.0.