Architecture

From aHuman Wiki
Revision as of 12:48, 21 June 2015 by Admin (Talk | contribs) (Automated page entry using MWPush.pl)

Jump to: navigation, search
High Level Architectural Design

Home -> Architecture


High Level Architecture v3

Overall Architecture

aHuman architecture is based on biological model. See BiologicalLifeResearch.

aHuman framework contains below high-level modules:

Module Name Implements Notes
ACTUATORS aWee, aYou, aPets, aHuman set of Sensors and Effectors for different targets Can interact with external world; there can be different sets of actuators, aPets or aWee can be test/trial one for prototyping aHuman
MIND mind intelligence and feeling in biological model includes brain, feeling components and all tracts
PLATFORM platform generic software application functions contains library components shared across different aHuman executable targets or specific auxiliary executable targets
TOOLS tools generic auxiliary executables using platform and libraries, multiple utilities combined in one executable, like cvs/svn/p4
LIBRARY library ready-to-use libraries can be project-generated library or external library; other modules are using only binaries and headers

Possible executables:

  • any actuators set, combined with MIND, PLATFORM and LIBRARY components
  • TOOLS combined with PLATFORM and LIBRARY components
  • executable LIBRARY components

Resulting high-level architecture model is below one:

hla-v3.jpg

Mind HLA

Mind is a sole part of aHuman and has its own layered HLA:

  • See MindArchitecture
  • mind component hierarchy can be sliced to present optimal level of functional items - set of mind regions
  • consider all the mind regions and links as produced by specific set of mind services, where one service is aimed to create specific set of circuit patterns in order to implement specific mental function
  • mind area is a group of mind regions and their links which are produced by specific mind service
  • neurolink implements excitatory and inhibitory neurotransmitters and neuromodulators, resulting in different timing and mode of signal propagation
Layer Key Components Implements
mind target MindTarget, MindSensor, MindEffector Mind target framework and lifecycle. Provides facade interfaces to specific target, isolates of mind and platform design
mind service MindService, MindNet High-level mind construction and service lifecycle implementation, factory for specific mind components, establish and maintain wide neural networks. Handles mind birth, growth and wake/sleep cycle
mind area MindArea and specific areas in unified aHuman mind Lifecycle of mind areas, regional nature and structure of unified mind. Sole mind architecture based on underlying technologies of next layers
mind region MindRegion and specific region types of cytoarchitecture Connectivity architecture of specific mind tissue types - cortex, nuclei and nerve system. Based on low-level neural components
mind matter NeuroSignal, NeuroPool, NeuroLink and specific implementations according to neurotransmitter systems Neuron types, neural networks, neurotransmitters. Effective representation of low-level mind state
auxiliary mind objects MindAreaLink, MindRegionLink, MindActiveMemory, MindMessage, MindSpace, MindLocation, MindMap Principal mind connectivity and thought flow. Effective execution of network signal propagation. Geometry of the mind - one more real dimension in brain modeling

Platform HLA

Platform contains multiple services and split on HLA level to components as well:

Component Main Class Implements Depends on
services ServiceManager base types, plain and template classes, logging, exception handling, service and platform lifecycle -
environment EnvService Handling environment, including configuring, deployed instance, operating system and other available resources -
objects ObjectService Unified objects and object persistence, object types -
math MathService Specific mathematical calculations -
persisting PersistingService Persistence of memory in filesystem space objects
threads ThreadService multi-threading environment, CPU management, performance monitoring objects
messaging MessagingService Internal decoupling mechanism to arrange manageable numerous interactions among components objects, threads
media MediaService Communicating with external world using standard protocols with non-aHuman services or internal aHuman protocol with aHuman tools objects, threads, messages
scheduling SchedulerService Running configured tasks regularly objects, threads, messages
statistics StatService Collecting performance statistics objects, scheduling, messages
adminapi AdminApi API to access internal aHuman components and run tests or specific operations media
testing TestService Testing and debugging support objects, messaging

HLA State and Development

  • rates: 1-5
Layer/Module Design Rate Code Rate Advances and TBD
ACTUATORS 2 focus 2 focus adv: unified sensor model, filesyswalker, tbd: overt and covert attention, digital-neural protocol
MIND
mind target 2 1 focus adv: separated by design, tbd: full lifecycle implementation in HLAv3
mind service 3 1 focus adv: biomodel, tbd: layered architecture, transition to HLAv3
mind area 2 focus 1 adv: cortex types, tbd: neocortex area design
mind region 2 focus 1 focus adv: HTM, reqs, features, tbd: complete neocortex region model
mind matter 2 2 focus adv: sourceforge cortex, fann, tbd: trial neocortex implementation
mind space 1 1 adv: locations, tbd: -
mind stream 2 focus 2 focus adv: -, tbd: link hierarchy, HTM links
PLATFORM 3 3 adv: life cycle, component-based code, HLAv3 codebase, tbd: -
TOOLS 2 2 adv: aiconsole, tbd: -
LIBRARY 2 2 adv: opencortex, fann, genrtn, chat, opencv tbd: -

Limitations

Feasibility Study gives us below limitations for aHuman design:

  • don't try to replicate human design exactly
  • use high-quality model of NN with 1Kb memory per neuron
  • use NNs for personality areas and knowlets for others
  • use 500K neurons in operating memory (0.05% of HI)
  • use 30K focused neurons (6% of AI NNs)
  • develop knowlets as plug-ins
    • use up to 500M in operating memory for activated knowlets (0.5% of HI)
    • remaining inactive knowlets can use up to 100Gb disk space
  • separate feeling from emotions
  • treat feeling as internal motivation system
    • use broken Markov network as a "pain" body signal
    • use sensor data consumed to adjust Markov network as "feed" body signal
    • use expected action in Markov network as internal motivator
    • curiosity is exploration action in Markov network
    • derive new "uncertainty" body signal as mismatch between expected/suggested action and actual action
  • don't create Markov networks initially, create HMM factories to construct networks
  • create learning cycle based on body signals
    • construct body reward function from body signals
    • create system of feedback recognition, train it using body reward function
    • create feedback reward function as output of feedback recognition system
    • use feedback reward function to learn all active Markov networks

Ideas inspired by coding

  • mind consists of mind areas
  • mind area consists of neural networks of various types
  • single AI component is responsible for specific area and defines how it is constructed from networks - shape, used network types, sizes and inter-area links
  • brain holds mind
  • brain defines how areas are connected
  • brain contains thread pool, executing elementary operations
  • elementary operation is either network input/output calculation or link walkthough
  • any operation leads to learning - either inside one network, or inside link
  • networks can be of small typical size - e.g. 300 neurons, and thread pool can be 100 threads executing their operations and then sleeping for a while
  • networks porcessed in the same time mean "mind focus"
  • after processing focus is changed

Stale Models