Ai Dreams Forum
Member's Experiments & Projects => AI Programming => Topic started by: Zero on September 22, 2018, 09:45:16 am
-
Generally speaking, do you think that an AGI's mind has to be a network of some sort, or are there other program structures that would fit?
-
Generally speaking, do you think that an AGI's mind has to be a network of some sort,
Good question… but a broad question.
Network - a group or system of interconnected people or things
Most AGI schemas are built around the passing/ sharing of information between disparate module types that share a common data type/ structure. Obviously both the interconnectivity of the modules running on a single machine, as well as the Ethernet connecting the physically separate machines can be construed as a network
Even the layers of a typical CNN are usually treated/ written as separate structures, and depending on their complexity and the recourses required often run on separate machines, connected by an Ethernet or some other high bandwidth transmission network/ protocol.
So it’s both the high interconnectivity required by the current AGI theories and the lack of computing recourses in a single machine that necessitate a network schema.
or are there other program structures that would fit?
Using the broad definition of network, any group of objects/ modules that pass/ share information is a network, separate modules/ functions that pass information through a predefined matrix is a network, even the computing structures used in liquid, reservoir nets can be called networks.
Even the human brain has networks... the definition is just too broad to escape lol.
So be it a consequence of either the broad definition of the word or the concept of interconnectivity the answer to your question in my opinion is… yup.
:)
-
Good answer... let's narrow the question a bit. What I had in mind was networks, as in *software* networks: sets of nodes which are structurally identical, and computing the same kind of functions, with the data being not in the individual nodes, but rather in the topology according to which they're connected.
Is my english english? ;D
-
Think of a context free grammar as a rules by a network is capable to compose. Think of a specific syntax tree as an instance of such a network. Parts of the syntax tree are data with more or less complicated structure.
It doesn't have to be a context free grammar. It may be JSON or any other structure capable of what we need to express.
-
Ivans one sounds more interesting than the ones I do, Any way to link I->O is what I use, so it includes spreadsheets, neural networks, or just simple equations. and im forming a model, to be used any way I see fit.
-
Oh! I see… I’d had four cups of coffee in a row… tends to make me waffle/ rant/ literal lol.
We humans are the problem; we have such a narrow/ limited methodology/ terminology for describing complex problem spaces.
The essence of an AGI is the re use/ re combination of known base knowledge/ variables into ever more complex representations. As this is the essence of an intelligent system, it is also our downfall.
Because the network topology as served us so well with other problems, we tend to want to re use/ think in terms of networks. It’s the main ‘goto’ technique we utilize… 99% of all AGI theories are built around the network principle.
What we actually need is a new type of math or language to adequately describe intelligence… we need to break tradition and known methods… think outside the box.
Hence my AGI program runs on simulated wetware computer architecture, the logic and syntax of the program is totally alien/ incompatible with modern processors. It’s not just about the connections, but where the nodes/ neurons are located within the connectomes 3D structure that affects processing.
I’m waffeling again… time for a brain dump lol
An intelligent/ processing system has to have some kind of base unit of knowledge that all other knowledge is comprised from.
The base unit has to be in a format that can be re-combined to represent other types of knowledge… at a meaningful processing level.
Ideally this knowledge unit should be the same format as the sensory input unit, the learned knowledge should exist in the system in the same sequence/ state it was received/ input. Intelligence/ rules should be both derived and applied to/ from data in this single format.
Obviously for a digital computer it’s 0 &1, the lowest base unit/ of information that can be re-combined to represent more complex concepts, sound, images, logic, etc.
This ‘base unit of knowledge’ concept is the current limit of our human understanding of the AGI problem space. We simply don’t know of any other way to represent information or process it… big things are made of smaller things… big knowledge comes from smaller knowledge… a + b = c.
I’ve forgotten what the question was… Oh! Yeah… lol.
For a schema to be able to re-combine low level representations of knowledge there has to be some structure/ method/ path ways/ logic/ order that represents the available options… a network, matrix or tree (which is also a network).
To achieve a required threshold of data complexity… the simpler the network, the more complex the data needs to be… and vice versa…
The only exceptions to these rules that I can think of ATM are natural laws of reality/ nature and physical mediums… like water.
End brain dump…
:)
-
An intelligent/ processing system has to have some kind of base unit of knowledge that all other knowledge is comprised from.
A cons cell (http://aidreams.co.uk/forum/index.php?topic=13207.0)??
-
To describe the whole propositional logic, we just need two operators like `not` + `or`. All other operators may be derived from the first two like `not (A or B) = (not A) and (not B)` or `(not A) or B = A -> B`. The rest of the job does underlying inference mechanism in a form of resolution rule. Such a power in only two logic operators. And the resolution rule guarantees that whatever can, will be concluded.
With a simple addition of sequencing we get predicates, and with bounded inference and scoping we get even quantifying over variables and predicates, which lifts propositional logic to higher order logic. Some new system, I'll probably write about it soon. A work name is: "bounded propositional logic with sequencing".
-
@Zero
I would think that’s the kind of concept yes… but you would need a much lower level of abstraction for a true AGI. The base unit of data has to be able to accurately describe any kind of experience/ information that the AGI could experience, that obviously includes spatial and temporal events. Words don’t actually describe data… in the absence of human interpretation.
If an AGI is going to exist in our reality and be compatible with the human condition, it will require senses on par with our own. We are as humans, the sum of our experiences, and we experience reality through our senses. We understand, remember, recall and dream in the format and resolution of our senses.
If your goal is to create an AGI you need a data format/ structure that can adequately describe/ represent anything/ everything the AGI can experience. Everything from propositional logic too the memories of a loved one, the smell of freshly cut grass too how to kick a ball or grasp a cup… one single data format.
@Ivan
Cool… That’s a good example of how we use hierarchical structures to solve/ portray complex concepts. Reminds me of/ similar to the NOR gate being functionally complete; they can be combined to create any other logical function.
I look forward to "bounded propositional logic with sequencing".
:)
-
Yep you need a net. It's also efficient too.
-
Given the current state of AGI research, I think it's a bit early to be saying that AGI has to be any particular architecture. At the moment, the only AGI accessible to our analysis and study is the biological brain -- one data point* in the entire possible solution space! And we don't even fully understand how it works.
*Or maybe a small cluster of data points, if the brains of humans, parrots, cephalopods, etc. can be said to have meaningful architecture differences.
If you want my completely speculative opinion, though, I favor approaches that feature a collection of specialized modules. These would be networks in the broad sense, because the modules communicate with each other ... but the modules aren't identical. They use algorithms and data structures (and, in the ideal case, maybe even hardware) optimized to their particular task within the intellect.
-
If you want my completely speculative opinion, though, I favor approaches that feature a collection of specialized modules. These would be networks in the broad sense, because the modules communicate with each other ... but the modules aren't identical. They use algorithms and data structures (and, in the ideal case, maybe even hardware) optimized to their particular task within the intellect.
yes. i have a version of that -> one neural network module per rule, the brain being a fixed amount of rules... for building its internal structure as well as finding its motor.
-
If your goal is to create an AGI you need a data format/ structure that can adequately describe/ represent anything/ everything the AGI can experience. Everything from propositional logic too the memories of a loved one, the smell of freshly cut grass too how to kick a ball or grasp a cup… one single data format.
I tend to agree on the idea of a common data structure for the entire system.
Several existing data formats are able to describe anything. Strings are one of them. They can contain images. S-expressions are another example. Why do you suggest that they can't adequately represent everything the AGI can experience?
We are biological beings, but AGI will be digital beings. Do you think the smell of cut grass is something they can experience the way we do?
If you want my completely speculative opinion, though, I favor approaches that feature a collection of specialized modules. These would be networks in the broad sense, because the modules communicate with each other ... but the modules aren't identical. They use algorithms and data structures (and, in the ideal case, maybe even hardware) optimized to their particular task within the intellect.
I favor these approaches too.
What's your opinion about initiatives like ROS? Do you think open platforms, where community-contributed modules can be plugged, are a possible path to AGI?
You mention specialized, "hardcoded" algorithms. Being frozen, don't they limit the potential evolution of the system?
-
Maybe a system that uses length of conductors instead of conductor architecture. You could have a brain made from a lot of unconnected conductors lined up together like bristles on a brush. Each one would self adjust it's length based on feedback from the input signal (eye) or feedback from the thing it outputs to (muscle). It might act more like a jellyfish than an AGI though.
-
You could emulate this... :)
-
AGI should be something that NASA can confidently send to Mars. They like things like PLEXIL (https://en.wikipedia.org/wiki/PLEXIL).
Ed: ...or that Skynet can confidently send to 1984 ;D
-
You mention specialized, "hardcoded" algorithms. Being frozen, don't they limit the potential evolution of the system?
When I said specialized, I didn't mean frozen. Such algorithms can make provision for learning and self-adjustment, though it is likely that they would not be *completely* flexible -- there would be some fixed core. In the limiting case of a fully developed AGI that has equivalent skill to a human programmer, it would be capable of changing and re-optimizing its own algorithms to any degree desired.
It may be worth mentioning that the "G" in AGI is a relative term. Human intelligence is the most generalized biological intelligence we know of, but even it should not be regarded as fully general. For instance, the sensory processing abilities of our brains may be adapted to the senses we have. Different brain regions handle different senses, suggesting some underlying structural optimization. If you hooked our brains up to something exotic (like the electroreception organs of a shark), it is unclear how well they'd do at handling that information.
There's probably a tradeoff to be made between complete plasticity/generality and effectiveness at any given task.
-
So, these algorithms are rather like agents, able to adapt (to some extent) their behavior to their environment to better fulfil their mission.
Agents can have a data axis and a meta axis. The meta input is for modifying the agent's behavior from the outside. The meta ouput is for the agent's activity log, and other things.
(https://aidreams.co.uk/forum/index.php?action=dlattach;topic=13399.0;attach=9398;image)
About plasticity, we know that regions of the brain that are specialised in a particular type of activity can be reassigned to other activities. But that makes sense only in highly parallel systems.
-
Maybe a system that uses length of conductors instead of conductor architecture. You could have a brain made from a lot of unconnected conductors lined up together like bristles on a brush. Each one would self adjust it's length based on feedback from the input signal (eye) or feedback from the thing it outputs to (muscle). It might act more like a jellyfish than an AGI though.
Jelly fish are amazing.
-
I don't know much about Jelly fish, but ants are cool. An ant colony has only one brain. But the brain cells are distributed among the ants. Sort of.
-
yeh that sounds quite undefeatable, a hive of telepathic robots, blue toothing each other.
-
I used to consider that networks of identical nodes was the most direct path to AGI, because of the plasticity it allows. But I'm more and more convinced that a thought stack as central mind structure is a better idea. Stack oriented languages really feel like thinking, don't they? The top of the stack is what you're currently thinking. Often, you put the current thought in "standby mode", in the background, while you solve another little thought train. Once done, you get back to your previous thinking process.
The thing is that stack should be supplied with high-level stuff by an interpretation/recognition mechanism. I don't know exactly how to do it. There's a notion of expectation and importance: you tend to focus on relevant things that are unusual/unexpected...
-
Stack oriented languages really feel like thinking, don't they?
I can see how it would feel that way, but your consciousness is at the top of a very complex hierarchy, you’re trying to figure out how the processor works by just looking at the monitor.
I think a thought is an accumulation/ compounding/ focusing of numerous lower level facets/ concepts. Every thought is made of smaller facets, bits of knowledge that guide the main thought patterns’ creation.
From my understanding of a stack (which could be wrong of course) it’s the topmost item/ concept that’s relevant at any given time.
I don't know exactly how to do it.
Ok so a stack is module that contains items, the items are ordered according to their relevance regarding the current thought, the topmost being the output.
Each item in the stack has a set of triggers/ inputs that move it up or down, these inputs are provided by lower stacks in the hierarchy. It’s a hierarchy driven by the outputs of lower hierarchies.
Let’s consider a stack that simply provides the left hands grasp state. The fingers have joints, so each joint gets a stack that represents the angle of each joint; the outputs from the joint sensors are fed into a finger position stack. The pattern of outputs from the joint stacks brings an item to the top of the finger position stack representing the whole finger position. Then the output patterns from the five finger stacks are fed into the overall had position stack, driving the correct overall posture to the top… hmmmm.
It’s like a high level modular representation of a how neural nets work. It takes a collection of neurons with several outputs to provide a varied output pattern. So each neuron group would represent an item in the stack, the pattern they all receive would only fire one output neuron; this represents the item at the top of the stack… similar to a Perceptron.
There's a notion of expectation and importance: you tend to focus on relevant things that are unusual/unexpected...
You could have global governance stacks that take their inputs from all over the hierarchy, their output could be used as bias, to guide the lower stacks, providing top down influence.
A stack is a logical module that provides an output dependent on its inputs.
If thinking about the problem space in this manner helps then run with it, you only have to design one stack module and a connection schema… cool. You could build some extremely complex architectures with this, and might get more insight into the problem space as you progress.
:)
-
Wow, very nice! O0
You're always a lot more "bottom-up approach" than me, korrelan. That's a good structure you're describing here. Would I be right to call it... a network of stacks?!! ::)
We could also imagine a multi-dimensional stack, like ranch said. Some kind of reversed tree: the top of the stack is the root of the tree. Under the unique top-most element, there's more than one stack-branches, which themselves have more than one stack-branches... It opens several unusual possibilities!
-
Wow, very nice! O0
You're always a lot more "bottom-up approach" than me, korrelan. That's a good structure you're describing here. Would I be right to call it... a network of stacks?!! ::)
We could also imagine a multi-dimensional stack, like ranch said. Some kind of reversed tree: the top of the stack is the root of the tree. Under the unique top-most element, there's more than one stack-branches, which themselves have more than one stack-branches... It opens several unusual possibilities!
Yes! I disbanded from it because I was being retarded and thought I could put infinite bits in one, the demo lasted about 5 pops then it died in digital noise.
-
When in doubt, give it a try! That's the way to do it.
-
That's a good structure you're describing here. Would I be right to call it... a network of stacks?!!
That was my interpretation of the structure/ schema… you were describing, I was just trying to jog your creativity. You can call it anything you like, Stackoids, Netstacks, Stacknets or a network of stacks if you like.
I was driving back yesterday and had few more thoughts; you will probably require different types of stacks. So besides just the pins moving up/ down depending on the stacks input, a type were any pin is moved straight to the top of it’s stack and output if the key is injected and another were if all pins are triggered in a stack it fires, etc. more types will become clear as you develop.
The schema also has some very distinct advantages.
It is very compliant with a graphical interface, drag, drop and link, etc.
Any type of data can be exchanged between stacks, not just strings.
It can easily receive imported data types, from visual modules, etc.
It’s compatible with a parallel architecture and so can be run on multiple cores/ machines.
Anyone can write compatible plugin modules, with an understanding of the data protocols.
Keep us up to date on your progress…
:)