distributed connection plattform & hardware
Three patent-pending hardware solutions for
- the reduction of memory usage and energy consumption
- the "short-term memory limitation" and
- a routing mechanism between experts
provide the basis this project proposal builds upon:
A (massively) distributed online and offline connection system powered by decisive-ai.
Local, portable devices process offline, connect selectively peer-to-peer or to one of the central connection plattforms.
In this project, we aim to tackle three interrelated issues to enable AI technologies to achieve unprecedented levels of performance and capability:
1) *Reducing memory use and energy consumption*
while increasing speed will facilitate the more cost-effective scaling of LLMs (Large Language Models) or enable them to deliver the same performance with reduced environmental and financial impacts. We plan to accomplish this through DBI’s method of matrix multiplication, which eliminates zero-values at the input of each processing pipeline stage (patent-pending).
2) *The “short-term memory” limitation* of LLMs
poses serious challenges, particularly when dealing with dynamic, data-intensive problems. DBI’s stream compression offers a robust solution to many of these issues (patent-pending).
3) *Mixture of Experts*
is a field of increasing focus. DBI's mapping mechanism provides a way to create a (massively) distributed system of specialised language models. This system connects both humans and machines together. It consists of two parts: Central connection plattforms - similiar to current social media plattforms and chat rooms (patent-pending) - and local digital assistants.
These new technologies are part of a processing ecosystem called "decisive-ai". It combines the power of current AI, which is based on floating point operations, with the precision of DBI ("DataBase Intelligence"), which is based on digital structure operations.
The word "digital" implies distinct states, which can be compared against.
The word "structure" implies a set of several values.
*Reducing memory use and energy consumption*
The common basis between AI and DBI is matrix processing.
Each matrix element consists of a triple of values: The two indices and the value itself.
Despite the value being a floating point it can take on the two distinct values of zero and one. These have a special meaning both in the digital world and in the world of mathematics. If a value of zero is regarded as
- a condition not fulfilled
- a multiplication not to be calculated
- a DBI structure to be omitted
a matrix processing pipeline can be buildt that drops these elements on each processing stage, leading to a novel lossless compression method (patent pending).
As each processing step requires energy, this compression method leads to a reduction of energy.
With estimations of sparsity in the range of 90% the resulting reduction of energy consumption is impressive.
While the standard way of processing a matrix multiplication A * B = C is to calculate the elements of the resulting matrix C one after another, this patent pending approach groups all multiplications per element of Matrix B (or A respectively, as the info graphic is showing). In such processing pipeline, each sequentially incoming non-zero element of Matrix B (or A in the graphic) creates several dependent calculations. An incoming zero element of B (graphic: A) is dropped.
This way, at each stage of the processing pipeline zero-values gets dropped, with little or even no decrease in processing speed.
*The “short-term memory” limitation*
decisive-ai introduces a new memory design, which eliminates the differentiation between data and adresses. Data strings of varying length can be used instead of addresses leading to a superior compact processing model. (further info under NDA).
*Mixture of Experts* - a novel mapping method
The third innovation basis is a novel approach to model dynamic relations between data. This approach, called DBI ("database intelligence"), provides a way to efficiently and dynamically map between several entities, for example several databases, without previously agreeing on common keys.
In, e.g. relational databases along the state of the art a primary key can be composed of several table columns, thus using the actual table data for mapping (instead of a unique, but not human friendly number). In this novel approach a unique key based on the actual data is created and updated dynamically. Still both data storing and retrieval is fast due to a specialised hardware (patent pending).
As a DBI structure can consist of different data types, similar to the definition of different data types for different columns in a database table, a Matrix element or even a whole Matrix can be stored in a DBI structure and its elements processed interdependently and in parallel (up to the extent of parallelism the hardware provides).
A project proposal
The imagined goal for 2030: A personal digital assistant for personal, i.e. offline-use, peer-to-peer use, i.e. local connections, plus central connection plattforms to connect in bigger groups or establish new connections, i.e. internet.
Those multiple personal digital assistants can be regarded as multiple experts dynamically connecting when needed or desired.
Imagine a timeline from now to the point in time (2030), at which this imagined goal will have reached a valuation of, e.g. one billion EUR, there's a chain of actions that led to this success.
Each chain or action link has a point of decision at its start.
For this successful chain of actions to reach from now to 2030 a correct decision must have been made at each points of decision in between.
At each point of decision, besides the option taken, at least one additional option must exist - it wouldn't have been a point of decision otherwise.
If each point of decision was noted in a separate column a database entry would be created. This entry, or set of parameters, or record is one valid combination, that leads from "now" to the desired (evaluation) outcome.
Another combination, that is another chain of actions, might have led along a different path, but would still be moving from the same starting point to the same desired end point.
As a practical example, in a metropole city like Berlin the public transport system provides several "chain of actions" aka "traveling routes" from the start to the destination location.
There's no requirement, that each chain link is decided upon, or executed by the same entity or person.
A Berlin bus driver told me once that he has been trained on a limited couple of bus lines (prior to the mass market adoption of navigation devices). That does make sense as a human has a limited capacity to remember, so specialisation is needed.
So, during a successful journey from A in the north to B in the south the traveller passes several fields of expertise (or bus lines or action links or decision points - the reader can see, where this explanation is heading).
DBI ("database intelligence") is a technology developed to combine predicting / statistic-driven technologies with deterministic / decision-driven approaches, hence the name "decisive-ai" (patent-pending).