Digital hippocampus, how to build one
In case you don’t know, hippocampus is that thing in your brain that helps you memorise things. It moves things from short-term memory to a long-term ones. Not all I mean, only those that really matter.
What matters, when and why?
We are really bad at forming general rules like the one about “what really matters” when it comes to input data. In consequence, we can’t write a proper algorithm and we struggle with building AGIs.
AGI stands for Artificial General Intelligence. The one that does not require us to build models on every occasion, like for autonomous cars or sorting rubbish in recycle centres. The algorithm that learns itself, grows and becomes smarter and smarter — exponentially — over time.
The Problem
…meaning, inability to collect, stack and spot patterns among various pieces of information were a subject of my studies.
As a solution I want to propose a very simple algorithm that could be used to draw general conclusions, expanding AI ability to generalise and making it applicable on many fields.
It’s based on the premise that information is collected constantly (illustrated as a constant input of data on the lowest pyramide tier). Also it’s input agnostic and has natural skill to spot semantics for every kind of data. Might be useful to discover natural patterns and semantics in complex arrays of data.
You can find a script that runs very simplified version of the idea here: https://www.spruce-mobile.com/samp2/pyramide/
Now you see it
It seems incredibly simple isn’t it? So what value does it have when used in real life scenarios?
Here is an example of a simple prototype build several years ago for MarieAI — my AGI project.
https://spruce-mobile.com/#aiResearcher — it has fragments of articles selected by the algorithm.
It was performed on several “pyramids” of words where initial ones served as filters — to remove most common words that have no unique value. Words like “the”, “a”, “of”, “with” etc.
Process was completely unsupervised and as you can see under the link, tested on several languages where no prior data about the content was provided.
What problems does it solve?
(aside from functioning as Digital Hippocampus)
Such simple pattern spotting solutions can be used to:
1. Evaluate content as well as measure it’s uniqueness, association.
2. Serve as pattern filtering, providing feedback for the algorithm which chunks of data should be stored and which ignored — it equips AI in ability to forget.
Above that, I wanted to propose it as a solution for Von Neumann Bottleneck for AI long term memory, as chips designed to store information that way can also perform essential computations while doing it. New architecture of AI chips could “forget” noise and store only meaningful information and all the operations would be limited to moving values between blocks.
Very similar to what the human hippocampus is doing.
https://www.techopedia.com/definition/14630/von-neumann-bottleneck
Patterns manifest themselves during a process of storing
You might think that this data needs to be processed again to be useful. Yes if you wish so. The system however allows for memory blocks to be a solution. More data alters memory blocks and no additional pattern recognition algorithm is required. Patterns manifest themselves in a process of storage.
This application is an accidental discovery (didn’t know such a silly problem exists but it’s due to my lack of knowledge in microchip architecture).
It was made however, simply as a result of design principles that the whole process was a subject for. Universal pattern recognition methods are something that belongs to the human world and should reach…
Beyond tech civilization
The whole idea had several principles in mind, helping us to progress with knowledge and information beyond computers and tech civilization. Fundamental principles:
1. It needs to have a natural ability to survive 700 years. With this requirement comes a necessity that the algorithm needs to be simple enough to be operational and transferable from CPU to simple devices, outside microchips. It means one should be able to build it from sticks and rocks. AI Algos need to be simple enough to work with sticks and rocks. Why? Because you never know what is coming.
2. It needs to be data agnostic, allowing for observation of any kind of data and comparing them together. From articles, through medical records up to stock market data and heartbeats,
3. It needs to be simple enough, so a person of IQ 80 could operate it and not screw it. It means it needs to be easy like a child cards game,
4. Expanding on the previous one — it needs to be operational by a civilisation that has no means to understand it. So it can survive past the point we communicate like we do now,
5. Even most complex models — like weather patterns — need to be compressed to a set of pieces of paper — let’s call it a book — that can survive past “digital apocalypse”. That way knowledge collected by one civilisation can be passed to the next one to use and make predictions without any available device — except the book.
Is the solution too simple?
You might wonder if it has any use and does it hold the water when data is poured? Well, it does! It’s used in some projects of mine and shortly it will be launched as a AGI v.01- something that will grow on it’s own.
Keep in mind it was built to be unsupervised so while creating quite intuitional pyramids across all branches it still governs itself, meaning some understanding is very different from ours and depends on what data and in which order are served.
Does it have any flaws? No!
If you truly understand what are the effects of this process then it has no flaws — it just has a process that provides constant, predictable results for pattern recognition allowing everything else to fade away. Mechanism built on this principle, with the ability to grow many pyramids like that, simply creates over time, millions piramides, interacting with each other based on provided data. It clusters information and creates branches for further clustering and it grows like that. Simple enough to pass all the requirements set above.
This is of course one of the many steps to AGI but quite necessary one, so I thought I’ll share with you.I hope you can see it from a different angle and it won’t take you four years to build your 1st AGI model like it took me.
About Me
Hi, I’m Marcin. Former game developer and algorithms designer.
Currently working on a novel, general purpose algorithm, you can use offline on any device. My dream is to deliver solution that could be trained, up trained and work as a swarm, solving insanely complex problems.
This is how MarieAI (https://marieai.com/) was born.
Concept of a Digital Hippocampus
More about algorithms and pattern recognition
Data Layering technique, explained here.
How to compress infinite amount of data using Digital Hippocampus, explained here.
Find me on LinkedIn: https://www.linkedin.com/in/marcin-rybicki-qa/
