The TEKnology‎ > ‎

FrANN™

Fractal Artificial Neural Network

FnnTEK's FrANN™ technology is a completely different approach for the development of truly intelligent systems. FrANN™ is an emergent computational architecture, not a deterministic one. Until now, all other AI efforts, even those based upon neurophysiology, have depended upon the human researchers to understand how each individual function of the brain works. The FnnTEK approach is to understand a much more basic building block and allow the higher functions to emerge from the complex non-linear interactions of those basic blocks.

The visual system is a great example. Currently, in the most advanced machine vision systems, researchers have taken apart the human visual cortex (physically, via MRI, PET and other scanners) and AI researchers create theories on how those structures work. They then translate this into deterministic algorithms or abstracted neural networks that mimic those functions and produce systems that can perform certain types of visual identification. These systems are much more functional than traditional AI techniques, but still limited in their applicability. The weakness here is the scientist’s insistence on understanding how the system works prior to building it; they fail to appreciate one of nature’s most powerful tools - emergent behaviors from complex adaptive systems.

FnnTEK’s approach is to create a large collection of software neruons for a particular sensory / manipulator environment. Our growth model produces internal structures (e.g., Lateral Geniculate Nucleus) that function in the computer as their biological equivalents do in the real world. The resulting network is an emergent computational architecture that exhibits high levels of reasoning within its environment, the ability to respond to first time exposure to data, and other ‘creative’ solutions; in other words, true cognitive behavior.

What is the process

Design

The first thing that happens is understanding the goals of the system. This controls the design for the interfaces to environment for the specific FrANN™ operation. The parameters involve the specification of the detectors, effectors, and pain / pleasure feedback loops that will control the system. For this example, we will use a simplified description of the process used in defining the DifMAN™ system.

Problem Statement

DifMAN™ is to manage periodic re-balancing of portfolios with a goal of consistently exceeding the best human fund managers. The system will create annual, bi-annual, quarterly, monthly, weekly, and daily portfolio balancing recommendations, autonomously generate these findings, communicate in English via written material.

Detector Determination

The system will use only open source publicly available data. The following list represents the primary senses of the system:

  • Stock pricing data after the daily close of markets
  • EDGAR filings
  • Yahoo Financial News Feed
  • Yahoo News Feeds
  • WWW Crawling
  • 6 Yahoo Portfolios - Simulated with $5k initial cash investment each for each of the listed time frames
  • ROI computations for each Portfolio for each time frame

Effector Determination

Initial operation is to generate buy / sell / hold recommendations for each security under consideration for the portfolio that corresponds to the time-frame under consideration.

  • Equity positions range from 0 to n shares, where n x price + commission is <= available cash to invest
  • Change in equity positions generate buy / sell number of shares, a zero change generates a hold recommendation
  • Buy / Sell recommendation includes textural reasoning behind the sale / purchase / hold recommendation
  • Findings are presented in colloquial English via publication of a periodic newsletter

Feedback Loop Determination

The general causes of pain for the system are 1) Loosing Portfolio Value, 2) Not Trading Often Enough, 3) Poor Performance relative to publicly listed mutual funds, 4) Insufficient Research, and 5) Insufficient communication of reason for purchase. The reasoning behind 1 and 3 is obvious. Number 2 is to prevent a 'buy and hold' strategy that already out-performs most mutual fund managers. Reason 4 is to prevent random selection (which also usually out-performs human managers). Reason 5 is to stimulate the system to learn how to communicate its reasoning.

Growth

  • Once the detectors, effectors, and pain receptors are programmed, the neurons are grown into the environment. How this happens is based on our trade secrets. These IP positions include stimulating fractal growth, interfacing software neurons to the real world, creating meaningful feedback, and pre-operation training of a digital mind as it is growing.
  • Once the system is grown, the digital mind (a text file!) is removed from the growth systems and placed into an operational environment where the system may be exposed to real data

Training and  Operation

  • The only difference between Training and Operation is the actions of the effectors. Usually, during training the effectors do not actually function, instead there is a simulated environment. In this specific instance, the system would not actually send newsletters to subscribers; they would only go to the trainers. Operation would be that point where the system was allowed to send the generated newsletters directly to end users.