Skip to content
Version 3.0 Implementation of Meaningful Based Cognitive Architecture in MBLS-3
Python Jupyter Notebook
Branch: master
Clone or download
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.gitignore
HLAI18 PowerPoint (in PDF) Presentation of MBLS.pdf Add files via upload Sep 1, 2018
LICENSE Initial commit Aug 31, 2018
Schneider_MBLS.pdf Add files via upload Dec 13, 2018
image_from_slidedeck_2.jpg Add files via upload Oct 10, 2018
mbls_brain_arch_0001.py Add files via upload Feb 3, 2019
readme.md
requirements.txt Add files via upload Oct 3, 2018
tester_notebook.ipynb Update tester_notebook.ipynb Apr 5, 2019

readme.md

MBLS-3.0 Revised Granular Simulations

Note: Due to the complexity of a fine-grained simulation of the MBCA, simulations have been re-organized in 2019 as a nested spectrum of simulations ranging from coarse-grained, demonstrating overall principles of the MBCA, to finer-grained, demonstrating more authentic MBCA components, simulations of the MBCA. Estimates for posting of working code which can be copied and/or forked is given below. Also, as a temporary stopgap measure PyTorch1.0 is now being used for sub-symbolic simulations, although there are issues with it modeling the behavior of a network composed of HLN (Hopfield-like Network) units rather than neurons as the basic unit.

Version 3.0 Meaningful Based Learning System (MBLS)

Meaningful Based Cognitive Architecture (MBCA)

The 'Meaningful Based Learning System' (MBLS) is an implementation of the Meaningful Based Cognitive Architecture (MBCA).

What Project Does

The MBLS creates an AGI (artificial general intelligence) via synergistically integrating the sensory processing abilities found in neural networks with many of the symbolic logical abilities found in human cognition. The basic unit of the MBLS is a reconfigurable Hopfield-like Network unit (HLN). HLNs can reconfigure with each other in an attempt to achieve a practical maximal 'meaningfulness' (defined as the reciprocal of the Shannon entropy of the HLNs). A portion of the HLNs are configured for hierarchical sensory processing, others as causal memory (including holding of multiple world views), and others organized as logic/working memory larger units which can process vectors from other parts of the MBLS, in accordance with intuitive physics, intuitive psychology, intuitive scheduling and intuitive world views stored in the preconfigured (instinctual) core goals module.

See below for key concepts that I wish to demonstrate in MBLS-3.

image of child from slidedeck

How to Install & Dev Environment

This is a Python-based project. Installation is intended to be very straightforward.

At the time of this writing, the code uses Python 3.6 with modules imported from the Standard Library or PyPI.

For development no specific IDE or other tools are required.

Installation and dev details are given at the beginning of the source code.

Licensing

All code of the MBLS-3.0 Project is open source, as per the license on this GitHub page. As well, all modules imported into the code are open source, as per similar open source licenses.

Papers describing the Meaningful-Based Cognitive Architecture and the MBLS:

(A prepublication of the Procedia article is on this GitHub page for download.)

Schneider, H., Non-Hybrid Meaningful-Based Learning System Using a Configurable Network of Neural Networks. Proceedings of the 2018 International Conference on Artificial Intelligence pp 96-102; Aug 2018.

Schneider, H., Meaningful-Based Cognitive Architecture. Proceedings of the 2018 Annual International Conference on Biologically Inspired Architectures (BICA 2018) held on August 22, 2018 as part of the Joint Multi-Conference on Human-Level Artificial Intelligence (HLA-18).

https://www.respekt.cz/spolecnost/muze-pocitac-trpet-psychozou

https://vimeo.com/297054113?ref=em-share

Schneider, H., Meaningful-Based Cognitive Architecture. Procedia Computer Science, 9th Annual International Conference on Biologically Inspired Cognitive Architectures, BICA 2018,edited by Samsonovich, A.V., Volume 145, 2018, Pages 471-480. https://www.sciencedirect.com/science/article/pii/S1877050918323974

https://github.com/howard8888/MBLS-3.0/blob/master/Schneider_MBLS.pdf

PowerPoint slides (in PDF format) of my presentation at the BICA section of the HLAI-18 in Prague, Czech Republic can be downloaded from this GitHub page, and give an idea of the key concepts behind the Meaningful-Based Cognitive Architecture and the MBLS implementation.

https://github.com/howard8888/MBLS-3.0/blob/master/HLAI18%20PowerPoint%20(in%20PDF)%20Presentation%20of%20MBLS.pdf

Excerpts of paper:

2018 Annual International Conference on Biologically Inspired Cognitive Architectures

Meaningful-Based Cognitive Architecture

Howard Schneider

Sheppard Clinic North, Toronto, ON, Canada

howard.schneider@gmail.com

Abstract

An overview is given of the cognitive architecture of the biologically inspired meaningful-based learning system (MBLS). The basic element of the MBLS is a reconfigurable Hopfield-like network (HLN) which can rapidly connect to other HLNs depending on the level of abstraction which yields a practical maximal “meaningfulness,” defined as the reciprocal of the Shannon entropy of the HLNs. Without any external memory the MBLS synergistically processes external data (and internal data – “thoughts”) with sensory processing abilities found in neural networks and some of the symbolic logical abilities found in human cognition. In practical applications the MBLS offers near-simultaneous pattern recognition and comprehension. In modeling the development of psychotic disorders in humans, the MBLS predicts that in many patients the etiology stems from the fragility of the working memory and the integration of additional reasoning mechanisms during adolescence.

Keywords: cognitive architecture, neural networks, cortical minicolumns

1 Introduction

At the time of this writing, despite the human-like performance of artificial neural networks (ANNs) in pattern recognition and reinforcement learning [1,2], such neural networks, trained with a very small quantity of examples, cannot causally make sense of their environment or information at the level a four-year old child can [3,4]. .....

....

Version History

version 1 - Winter 2017 - Initial transfer from whiteboards to Python simulation. No ANN libraries. Hand coded in Python 2.7. Shannon diversity used to assess meaningfulness values. (No GitHub repository ever created -- local files.)

version 2- Spring 2018 - Refactored to Python 3.6. Hand coded. No ANN libraries. Reciprocal of Shannon entropy now used to assess meaningfulness. MBLS modules still simulated by Python methods rather than being simulated by some of the collection of connected simulated HLNs. Causal memory now extensively used. Full development prematurely terminated for transition to Version 3 to allow collaboration and to more scientifically address concepts that the Meaningful-Based Cognitive Architecture proposes. (No GitHub repository ever created -- local files.)

version 3 - Summer 2018 -work pending - Code refactored into a more professional appearance to allow future team collaboration with the code. HLNs are being ramped up to 50,000 from a few dozen in previous versions, with the goal of obtaining quantitative experimental data to better validate key concepts behind the Meaningful-Based Cognitive Architecture. (GitHub repository is MBLS-3.0 )

intermediary version 3 note -- February 2019 -- PyTorch being used to simulate the subsymbolic HLNs in the what we are now calling the "MBCA" -- Meaningful-Based Cognitive Architecture. PyTorch to be swapped out by actual simulation of HLNs.

Version 3 Revised Granular Simulations -- Nano, Micro, Mini and full MBCA Versions -- Spring 2019

Due to greater levels of complexity, despite efforts to follow best programming practices and the 'Human Oriented Programming' the project espouses, a new revision of the project has been started. It is not sufficient to simply get the code working and participate in a simulation of finding a lost hiker in the forest, but the code must be understandable enough to demonstrate principles of the MBCA and to be continuously modifiable in this regard.

In Version 3Revised of the project is split into a nested hierarchy of differing details.

version 3Revised -- Nano - Smallest 'nano' size, Largest granularity of simulation - Implements fundamental principles of the MBCA with less than 1000 lines of code in a very transparent and understandable fashion. Short cuts are taken to accomplish this goal, eg, a Standard Library fuzzy pattern matching module is used rather than PyTorch in the higher levels of the project. All the principles of the MBCA can nonetheless be demonstrated in this version. EST DATE AVAILABLE TO RUN, COPY & FORK: June 1, 2019

version 3Revised -- Micro - Modest 'micro' size, Better granularity of simulation than the Nano simulation - Less than 10,000 lines of code. Off the shelf neural networks used for the subymbolic portions of the architecture, at present PyTorch is being used in Version 3. EST DATE AVAILABLE TO RUN, COPY & FORK: August-November 1, 2019, depending on release of pre-final versions

version 3Revised -- Mini - Substantial 'mini' size, Good granularity of simulation - PyTorch possibly swapped out for custom HLN subsymbolic library at this level of simulation. EST DATE AVAILABLE TO RUN, COPY & FORK: Jan 1, 2020

version 3Revised -- Full MBCA - Large code base, team programming required, Fine granularity of simulation. EST DATE AVAILABLE TO RUN, COPY & FORK: TBD

Key concepts to demonstrate in MBLS-3:

  • To show that psychosis (retrieval of incorrect vectors not associated with the reality at hand and cognitive dysfunction) is inevitable if Working Memory without error-correcting circuitry or software is used and becomes stressed.

  • To show that an intervention that reduces the likelihood of psychosis in the MBLS will by analogous intervention reduce the likelihood of a young person of developing chronic schizophrenia.

  • 50,000 proposed HLNs for MBLS-3 are not sufficient to implement the presentation example (eg, see slides which can be downloaded from this GitHub page) of the MBLS Search-and-Rescue Robot, but are enough to show a simplified Search-and-Rescue Robot simulation.

  • Better implementation of causal memory.

  • Better use of causal memory in the Working Memory to allow one-example learning

  • Better use of causal memory in the Working Memory to allow one-example extrapolation

  • Show separation of Procedural Memory and Declarative Memory as most Cognitive Architectures do, is not necessary

  • To show meaningfulness feature extracts more data out of an input sensory vector than not using the feature

  • To see if backpropagation of MBLS-2 between the HLNs can be replaced by one-shot reinforcement learning in MBLS-3

  • To better define minimal intuitive physics required

  • To better define minimal intuitive psychology required

  • To better define minimal intuitive planning required

  • To better define minimal intuitive meta-learning required

  • To better define minimal human culture required to ensure safe operation, direct wiring to emotional center

  • To better interweave subsymbolic and symbolic processes in the MBLS

  • To better allow meaningfulness to be used adjunctly in the symbolic processes, rather than just as input sensory vector processing

  • To better implement reinforcement learning via consciousness/goals/emotional centers

  • To better implement the autonomic modules and account for incorporation within the a physical system, such as the MBLS Search-and-Rescue robot

  • Effect of the autonomic modules in response to surprise negative stimuli

  • More work on the pre-configuration Python algorithms

  • More work on the Python learned/stored algorithms to better control library functions

  • More realistic Python quasi-back propagation

  • Implement the discrete learning features made available in the previous version/ implement semantic nets

  • Better use of the dual functions of the feedback vector to the HLNs -- meaningfulness and learning of the pattern at hand

  • Better use of the system-wide meaningfulness feedback vector

  • Fine tuning of the vector convergence circuitry in the HLN simulation so that the Hopfield nets remain usable in a large network

  • Start transition from the Python high functional simulation of the Working Memory to simulation by individual HLNs

  • More dynamic algorithms for reconfiguration of the HLNs rather than the very static algorithms used in previous version

  • Start implementing the functions of the sequential/error-correcting module and interweave with the rest of the MBLS

  • Consider implementing meta-prototyping structures versus intuitive algorithms

  • To better define minimal algorithms for world building

  • To better define minimal algorithms for logic operations on world models from causal memory

  • Given all the other structures, to better define minimal algorithms for common sense-like ability

  • To decide which levels of sensory hierarchy to feed directly into Working Memory, and which to not, ie, V1 vs V4 situation

  • See if can start implementing some simple language features with 50,000 HLNs

  • Lifelong learning -- one item per HLN

  • Lifelong learning -- start holding multiple items per HLN

  • Intuitive goals

  • Sequences of thoughts in MBLS (ie, feed output of working memory back in as input to the MBLS)

  • Self-awareness -- define better causal recording of Working Memory for better problem solving

  • Self-awareness -- define better causal recording of Working Memory for better transparency

  • Self-awareness (consciousness center/emotional/reward module) solving class imbalance problem and resulting overfitting to correct

  • To better describe the architecture of the MBLS from a conventional classification. Empirically show the MBLS must act as a Turing machine when the Working Memory is used (ie, as per the CT thesis a person with pen and paper could come up with the same results), but at other times can act as a super-Turing machine (ie, person with pen and paper could not compute the results of the MBLS), with, of course, the caveat that this description may not be valid, ie, are such machines even theoretically possible according to a more expert reading of the CT thesis.

You can’t perform that action at this time.