Concepts in MeTTa

chevron-icon
RFP Proposals
Top
chevron-icon
project-presentation-img
Expert Rating 4.0
Soumil Rathi
Project Owner

Concepts in MeTTa

Expert Rating

4.0

Overview

There is a need for components within the larger PRIMUS framework relating to knowledge / concept representation and concept generation. To fill that need, this proposal presents a new framework for concept storage, representation, and generation. It also includes implementing a FCA framework (with fuzzy and paraconsistent combined) within Hyperon for real-time concept learning and adaptability to previously unseen objects and attributes. Moreover, it implements an LLM-based evaluative metric to judge the novelty and creativity of the concept generation algorithms, and also proposes areas of exploration in the future to expand on the capabilities of the knowledge representation framework.

RFP Guidelines

Experiment with concept blending in MeTTa

Complete & Awarded
  • Type SingularityNET RFP
  • Total RFP Funding $100,000 USD
  • Proposals 6
  • Awarded Projects n/a
author-img
SingularityNET
Oct. 4, 2024

This RFP seeks proposals that experiment with concept blending techniques and formal concept analysis (including fuzzy and paraconsistent variations) using the MeTTa programming language within OpenCog Hyperon. The goal is to explore methods for generating new concepts from existing data and concepts, and evaluating these processes for creativity and efficiency.

Proposal Description

Project details

This proposal presents a framework to enable dynamic and adaptive knowledge representation (concept addition, storage, and generation) in MeTTa. It does this through a combination of multiple modules which run in parallel and have been detailed below.

Storage:

  1. The framework will store concepts within MeTTa in alignment with FCA - using objects and attributes. 

  2. It is important to note that any concept generation framework must have explicit relationships between concepts, which requires the attributes to be recursive in nature

  3. To allow for dynamic concept generation, a concept should have syntactical attributes that tell the system how they would be placed in a blended concept ie. a Book can be about another concept (meaning), a King can be the king of another concept (control). These base relations/attributes have to be defined within the concepts.

Concept Addition: 

  1. The system will dynamically receive new objects and attributes, and integrate them within the current knowledge context. 

  2. In the case of new objects, the attributes of the object will be identified, and thus, the object will be added to an existing concept or create a new concept in the lattice

  3. In case any new attributes are found, fuzzy and paraconsistent FCA algorithms will be used to identify the performance of old objects on the new attributes, thus allowing adaptive concept formation. This would be continuous process, thus allowing for any prior errors to be efficiently corrected.   

Concept Generation: 

  1. Concept generation will be highly linked to the relational attributes for all existing concepts and objects. Based on how different ideas and concepts can blend together, they will be combined to generate new concepts. 

  2. These continuous concept generation algorithms will be running parallelly, allowing the model to dynamically create new concepts as it learns more. 

  3. The input contexts will be chosen based on their relational attributes from varied “mental states”, thus allowing for novel solutions. 

Evaluation:

A multi-agent LLM-based system will be created to evaluate the creativity of this system. It will be designed as follows:

  1. There will be a “Tester” LLM that will be given the same challenge as the concept generation system

  2. There will be three “Judge” LLMs (to get an average) that have to rate the creativity of both the Tester LLM as well as the concept generation system. These judges will have multiple instances opened with the following format:

    1. They will first be asked to evaluate both responses individually (without seeing the other response). 

    2. They will then be asked to compare both responses, score both, and decide a winner (being able to see both)

  3. Key points: These tests will be anonymous (judges won’t know which response is which) and each test will be conducted 5 times and averaged to increase reliability of judging.

Scope for Exploration:

While not explicitly part of the RFP, I believe that knowledge representation and generation is a critical component of PRIMUS and of the overall Hyperon framework. As such, it is important to evaluate paths where this framework can be extended and improved.  

One potential idea for this is exploring hypervectors as a method to represent objects and attributes. This allows for much higher generalization, also allowing perception engines to be created that can learn ideas from the environment more generally, allowing for a closer step to general intelligence. 

As such, this proposal also includes an exploration into hypervectors and other promising extensions, believing future improvements to be essential.

Open Source Licensing

Apache License

Proposal Video

Not Avaliable Yet

Check back later during the Feedback & Selection period for the RFP that is proposal is applied to.

  • Total Milestones

    6

  • Total Budget

    $60,000 USD

  • Last Updated

    5 Dec 2024

Milestone 1 - Concept Design

Description

The first step is create the basis for FCA by defining things like objects and attributes within within MeTTa thus also including concepts to be setup. Apart from having attributes that simply define them each object will also have attributes that define how they can interact/blend with other objects. Each of these "relational attributes" are essential to novel concept generation and must be defined explicitly. This milestone covers the definition of a concept in MeTTa as well as the initialization of these concepts and the definition of the "relational attributes" for objects.

Deliverables

The deliverable for this milestone will be the codebase so far. This codebase will include the ability to create a concept in MeTTa. The types of "relational attributes" will be setup for individual objects. Functions for adding new objects or attributes to a concept will be created.

Budget

$10,000 USD

Success Criterion

A successful implementation of this milestone will be a system that can be able to store created objects and attributes. This system will also allow concepts that link these objects and attributes to be setup. By this milestone the "relational attributes" will also be setup allowing the system to link an object with an attribute signifying its syntactical role in concept creation.

Milestone 2 - Developing FCA algorithms

Description

Realtime and adaptive concept learning and generation is possible with FCA algorithms in place that can handle new objects and attributes. These algorithms will have to be both fuzzy and paraconsistent to work with complex real world data. So this milestone covers the development of FCA algorithms for learning about new objects and attributes.

Deliverables

The deliverable for this milestone will be the codebase for the system so far. This system will be able to learn new objects and attributes after being initialized and will be able to integrate these accurately within the existing concept lattice. This means using FCA algorithms to identify the links between existing objects and new attributes as well as new objects and existing attributes.

Budget

$15,000 USD

Success Criterion

A successful implementation of this milestone will result in a system that can perceive new attributes as well as new objects and be able to fit them within the existing concept lattice. An example of this would be a situation where a system that has been given information about animals and their attributes (size; color; fur etc) sees a new attribute for flight and can identify whether the existing animals possess it or not based on FCA.

Milestone 3 - Concept Generation

Description

Concept generation is the primary focus of this proposal and will be covered in this milestone. This milestone sets up rule-based algorithms to identify compatible concepts across different "mental spaces" for blending them together based on the "relational attributes" previously set up.

Deliverables

The format for the deliverable for this milestone will be the codebase. This codebase now will include rules in MeTTa to identify compatible concepts as well as algorithms to actually blend different input concepts to create novel ideas. The attributes of these novel ideas will be based on the role of each input concept in the final blended concept and will be defined dynamically with additional algorithms.

Budget

$15,000 USD

Success Criterion

A successful implementation of this milestone would have an agent that can generate novel concepts and ideas realtime. This agent would be able to view different unrelated concepts and blend them based on relational attributes. Eg. Given two concepts Book (which can be about something else) and AI (which is a topic in itself), the agent would be able to combine them for the specific and novel concept (Book about AI).

Milestone 4 - Evaluation

Description

It is important to setup a fair and reliable evaluation system to ensure that the designed system is functioning as intended. This will be covered through a multi-agent system within in this milestone. This milestone will also include developments of a few demonstrations of the system to showcase and evaluate its capabilities.

Deliverables

The deliverable for this milestone will be the codebase for the multi-agent LLM evaluation framework. It will also include the codebase for the developed demonstrations of this system. This milestone also includes a final report on the results of this evaluation detailing the creativity and novelty of the generated concepts.

Budget

$10,000 USD

Success Criterion

A successful implementation of this milestone will include well documented demos that can be understood by the rest of the community. It will also include a detailed report about the creativity of the models as well as a maintained codebase for the evaluation framework.

Milestone 5 - Implementation

Description

After developing the concept generation system this milestone will implement it within the existing Hyperon system. This implementation will follow the other requirements within the RFP including documentation and an API

Deliverables

The deliverables for this milestone include the codebase being added to Hyperon and complying with the requirements of the RFP. It also includes the detailed documentation for this codebase.

Budget

$5,000 USD

Success Criterion

The success criteria for this milestone is a successful deployment of the framework on the Hyperon Atomspace and wider codebase. A successful implementation would also include a detailed documentation understood by the wider community such that the framework is highly extensible.

Milestone 6 - Experimentation

Description

This milestone will include explorations into ideas such as hypervectors to ideate on future improvements. Other such technologies will be researched to identify paths towards more generality within the concept framework.

Deliverables

The deliverable for this milestone will be a report on ideas of exploration beyond the current concept framework that can contribute to the wider AGI approach. The report will also evaluate these ideas depending on their potential.

Budget

$5,000 USD

Success Criterion

A successful implementation of this milestone will have a well written and detailed technical report that can serve as a basis for future exploration into this component.

Join the Discussion (0)

Expert Ratings

Reviews & Ratings

Group Expert Rating (Final)

Overall

4.0

  • Feasibility 5.0
  • Desirabilty 4.0
  • Usefulness 3.5
  • Expert Review 1

    Overall

    4.0

    • Compliance with RFP requirements 5.0
    • Solution details and team expertise 4.0
    • Value for money 4.0
    A competent proposal fully addressing the requirements

    The proposal seems competent and fully addresses the requirements. It doesn't add or demonstrate any special insight or have any new ideas, but it totally makes sense.

  • Expert Review 2

    Overall

    4.0

    • Compliance with RFP requirements 5.0
    • Solution details and team expertise 4.0
    • Value for money 3.0

    A solid and comprehensive approach that presents "a new framework for concept storage, representation, and generation [....] includes implementing a FCA framework (with fuzzy and paraconsistent combined) within Hyperon [....] implements an LLM-based evaluative metric to judge the novelty and creativity of the concept generation algorithms, and also proposes areas of exploration in the future to expand on the capabilities of the knowledge representation framework."

feedback_icon