Revision as of 08:26, 2 March 2016

aMatter Requirements

@@Home -> ProjectPlanning -> aMatterRequirements

Contents:

2016 book of work

• target research - split specific and non-specific mind - done
• split mind into target, core, lifecycle, specific and integration modules - done
• copy aHuman core model to aWee - done
• create aWee target model - done
• implement filewatcher - primary sensor analogous to human eye (with focus point, focus depth, projection to neuromatrix - to be able to feel file system) - in progress
• design lifecycle set for software alive creature (aSoftLife, state control and autonomic control modules) - in progress
• reduce aWee core model to functionally transparent, improve external circuit coverage
• define aWee specific and integration models
• biological research - define set of neural tissue types
• biological research - describe logic of neural tissue types
• create aWee dynamical model
• setup running aWee model, define runtime metrics to measure proof of the concept

Overall Features

• to be refined later

Mocked Functions

• M-01. aMatter has primitive predefined set of effectors producing representation in external world based on predefined set of low-level commands provided by hardcoded motor strategies
• M-02. aMatter limits effectors actions to ones explaining internal representations by means of hardcoded symbolical language
• M-03. aMatter has primitive predefined hierarchy of behavioural strategies, on leaf level directly connected with effectors commands

Cognition

• C-01. aMatter receives information using predefined set of sensors
• C-02. aMatter recognises received information in real-time mode and calculates recognition metric R reflecting percentage of successfully recognised sensors inputs in given environment
• C-03. aMatter generalises unrecognised inputs so that R monotonously increases for the same static environment
• C-04. aMatter forms growing set of internal entities, so that specific subset of internal entities, when being in active state, can be treated as a representation of specific external data from sensors, disregarding whether previously perceived or internally inspired
• C-05. aMatter is able to forget internal entities, if not activated for a long time, so that the same input triggers another set of internal entities after a while
• C-06. aMatter forms growing set of associations between internal entities activated about the same time

Feeling

• F-01. aMatter collects information from predefined set of embodiment signals equivalent in purpose with human being, with body treated as related operating system process with all its inherent features and properties
• F-02. aMatter collects predefined set of uncertainty metrics from expectation flow of behavioural strategies

Goals Achievement

Perception/Self-Learning features

Perception

• Effective Signal Processing
  • MindSensorArea - receptive field averaging and inhibition
  • ThalamusArea - create relay sensory nuclei - done
  • ThalamusArea - create inhibitory NeuroPool - done
  • Connect PerceptionArea feedback and inhibitory NeuroPool - done
  • Connect internal InhibitoryLink from relay NeuroPool and inhibitory NeuroPool - done
  • Implement realistic inhibitory properties (150ms inhibition vs 20 ms excitatory non-firing interval)
• Functional value
  • implement sampling
  • implement subsampling

Self-Learning

• Mock cortex implementation - to allow development of other components
  • create feed-forward NeuroPool - done
  • create spatial pooler - process feed-forward signal from ThalamusArea - done
  • create temporal pooler - process fixed-size sequences and derive predicted spatial pooler item - done
  • create feedback NeuroPool - done
  • apply temporal pooler prediction to feedback NeuroPool and generate cortex feedback signal - done
• Cortex implementation
  • Columnar processing
  • Hierarchical Processing
  • Infinite Temporal Prediciton
  • Focus Processing
  • Attention Processing
  • Event Driven Implementation
• Functional value
  • implement high-probability predictive sampling and subsampling

Cognition as Meaningful Sensor Control

• saccadic scene scanning, spacial into temporal-spatial approach
• novelty, motavation, attention
• virtuality

Demonstrate Feeling Feature

• TBD

Real-Time Neural Networks

• Neural Structures
  • NeuroPool - accumulate arriving action potentials in membrane potential - done
  • NeuroPool - time-based dissolving of membrane potential - done
  • NeuroPool - postpone firing to minimum time interval after last firing - done
• Signal Processing
  • ExcitatoryLink - project excitatory signal to NeuroPool - done
  • ExcitatoryLink - generate excited signal from projection - done
  • NeuroSignal - store only activated source items - done