Soumil Rathi

Project Owner

Concept Blending in MeTTa

Expert Rating

n/a

Type SingularityNET RFP
Funding Request $30,000 USD
RFP Guidelines Experiment with concept blending in MeTTa

Overview

In this proposal, I will first implement a regular information-theoretic concept blending algorithm. Then, I will design and implement a new concept blending algorithm based on uncertain (fuzzy, paraconsistent) FCA. To judge the creativity and logical coherence of each of the algorithms, I will implement a multi-agent LLM judge system and evaluate both algorithms using this system to identify the better one. Finally, I will then implement the better concept blending algorithm into the Hyperon framework and write all necessary accompanying documentation.

Project Tags:
Community and Collaboration

RFP Guidelines

Experiment with concept blending in MeTTa

Complete & Awarded

Type SingularityNET RFP
Total RFP Funding $100,000 USD
Proposals 11
Awarded Projects 1

SingularityNET

Apr. 14, 2025

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. Bids are expected to range from $30,000 - $60,000.

Proposal Description

Our Team

I have previous experience with AI (neurosymbolic AI) research and development:

Research

Approaches to AGI -> https://arxiv.org/pdf/2202.03153.pdf
- Wrote this 3 years ago about how I saw OpenCog as the best path to AGI

Projects

Built a cognitive system with hebbian learning and heirarchical actions that reliably use 50+ apps
- Code: https://github.com/SoumilRathi/sentient/tree/main
Made a CSS styling agent with graph-based memory
- Code: https://github.com/SoumilRathi/Flaire

Project details

In this description, I will be elaborating on each aspect of the proposal one-by-one:

The information-theoretic algorithm

This algorithm takes in any two concepts with a set of weights for both truth values (true and false) for each attribute (for fuzzy and paraconsistent inputs).

When combining attributes between two concepts for a new blended concept, the algorithm looks through to find attributes with high entropy. These are the ones that carry the most information about the object, and are important to preserve. Then, the algorithm employs certain heuristics to match the new concept with the right attributes. For example, any attributes with contradictions can be removed, and the ones with high agreement can be specifically stored.

Finally, you could calculate the attribute weights (for both the true) for the new blended concept based on the weights of the original concepts.

This step can then be repeated for forming new attributes, creating a cycle and working well as a continuous concept blending algorithm.

The Uncertain Concept Blending algorithm

This is an alternate algorithm that will also be implemented and compared with the current information-theoretic implementation.

Given a set of objects, attributes, and the matrix indicating their respective weights, the Uncertain Concept Blending algorithm, or UCB algorithm, would use a clustering algorithm to find objects with similar attributes. It could then create a new concept with those attributes (and other ones from the parent objects). A similar clustering algorithm could then be used for forming new attributes as well, leading the cycle to continue repeatedly.

This cycle would thus lead to repeated formation of new concepts, thus being a suitable concept blending algorithm for Hyperon.

To make this algorithm compatible with fuzzy and paraconsistent FCA, the clustering algorithm must handle uncertain inputs. With that in mind, I will now be elaborating on my design for the clustering algorithm.

The clustering algorithm:

Given a set of objects S, and a weight vector w for each of those objects (corresponding to the relation between the object and each attribute), here is how the clustering algorithm would work:

Initialize by placing each object in its own cluster.
Compare clusters pair-by-pair and combine the ones with the highest similarity, based on a given similarity function.

Now, this similarity function has to be designed to work well with fuzzy and paraconsistent FCA.

Given a weight vector carrying the truth and false values for each attribute for each object, the similarity function between two objects could be the average of the difference between the truth and false values for each attribute, or some similar measure comparing the weights of both values.

3. Now, we can keep iterating until the similarity is below a particular threshold. This threshold is the value that can be modified to increase/decrease creativity.

Thus, this algorithm will allow Hyperon to work with both fuzzy and paraconsistent FCA, as well as get the ability to modify its creativity.

The LLM Judge:

The judging system will be designed to ensure that the evaluations for the answers are as accurate as possible:

There will be 5 LLM judges (each will be a different model to ensure no model bias). Each of the LLMs would be prompted to score the blended word on both novelty and logical consistency.
Each of these LLM calls will be run in parallel to ensure minimum latency
The 10 (2 values for each of the 5 LLMs) accumulated values will then be averaged to get an unbiased and accurate score.

The LLM responses, specifically for comparing the two algorithms, will be supplemented by human checks for novelty and logical consistency. The intention is to evaluate using the LLM system, but human oversight will be used to ensure that the evaluation is working as intended.

Open Source Licensing

Apache License

Activity Summary

Milestones

3

Total

Discussion

0

Total Comments

Reviews

0

Total Posted

Project Team

1

Total People

Total Milestones
3
Total Budget
$30,000_USD
Last Updated
23 May 2025

Milestone 1 - Planning

Description

This milestone involves creating the entire research plan for this proposal. This includes a detailed outline and timeline estimates for the work to be done. This milestone also includes creating rough implementations of the two concept blending algorithms—the information theoretic and uncertain FCA algorithms. As part of planning the milestone will also contain a rough “toy” implementation of a concept blending function in MeTTa and the Hyperon Atomspace—this is to setup the integration between the concept blending algorithm and Hyperon and ensure it works as intended.

Deliverables

The deliverable for this milestone would be a document containing the research plan for this proposal as well as a showcase of the rough implementation of the two cluster blending algorithms. Another deliverable here is an early version of the MeTTa concept blending algorithm integrated within the Hyperon Atomspace.

Budget

$6,000 USD

Success Criterion

The success criterion here is the successful creation of a reasonable and realistic plan, as well as a good implementation of the cluster blending algorithms. Plus, the success criterion of the “toy” implementation of the algorithm is that the algorithm functions without errors and can integrate successfully with the overall Hyperon Atomspace.

Milestone 2 - Implementation

Description

This milestone marks the implementation of each of the modules mentioned in the proposal—the information theoretic algorithm the uncertain FCA algorithm and the LLM based evaluator. These modules will be implemented in Python—originally—to enable faster development and more convenient testing. The final implementation of course will be in MeTTa. Once the modules are implemented both the algorithms will be compared using the LLM based judge system to evaluate their effectiveness in concept blending.

Deliverables

The deliverable here will be the implementation of both algorithms and the LLM based evaluator as well as the final report on the results of the comparison containing the information regarding which is the better algorithm for concept blending. The report will contain the complete data from 10 iterations of the LLM judge system to ensure an accurate and robust evaluation.

Budget

$12,000 USD

Success Criterion

The success criterion for this milestone is the successful implementation of the two concept blending algorithms as well as the LLM based evaluation system. Another success criterion here is the delivery of the final report of the comparison between the two algorithms, and thus, the knowledge of which algorithm is more suitable to implement.

Milestone 3 - Submission

Description

As the final milestone this one contains the implementation of the final chosen algorithm in MeTTa and integrated with Hyperon’s Atomspace. Naturally the system will be designed professionally to meet the Non-Functional Requirements such as making it modular and extensible as well as well-commented and documented. This milestone also includes the completion of all accompanying documentation including writing a report on how to use and extend the software and creating an interface to let the community use the concept blending algorithms.

Deliverables

The deliverable for this milestone is the implementation of the final concept blending algorithm into Hyperon’s Atomspace a report explaining how to use and extend the software an interface to use the concept blending algorithms any other required documentation. The intention behind the documentation is to promote community usage and development of the concept blending algorithms and they will be designed to promote that.

Budget

$12,000 USD

Success Criterion

The primary success criterion for this milestone is a successful implementation of the final concept blending algorithm into Atomspace. Another success criterion here is adoption of the interface and documentation by the SingularityNET community, with the these services being intuitive and easy to understand by community members. With a successful implementation and reception, this milestone—and this project—can be marked complete.

Join the Discussion (0)

Expert Ratings

Reviews & Ratings

No Reviews Avaliable

Check back later by refreshing the page.

The weighted average of the 4 perspectives Overall

0.0
Each RFP defines a maximum allowed budget, but teams can differentiate their proposal by offering a solution with a lower budget or a wider scope.Value for money

0.0
This rating indicates compliance to 'Must haves' but also adaptation of 'Nice to haves' and Non-functional requirements defined in the RFP.Compliance with RFP requirements

0.0
RFPs will offer varying degrees of freedom. This rating indicates the quality of the team's specific solution ideas, the provided details, and the reviewer's confidence in the team's ability to execute.Solution details and team expertise

0.0

Review Headline

0 /50 chars

Review Summary

0 /5000 chars

The weighted average of the 4 perspectives Overall

0.0
Each RFP defines a maximum allowed budget, but teams can differentiate their proposal by offering a solution with a lower budget or a wider scope.Value for money

0.0
This rating indicates compliance to 'Must haves' but also adaptation of 'Nice to haves' and Non-functional requirements defined in the RFP.Compliance with RFP requirements

0.0
RFPs will offer varying degrees of freedom. This rating indicates the quality of the team's specific solution ideas, the provided details, and the reviewer's confidence in the team's ability to execute.Solution details and team expertise

0.0

Review Headline

0 /50 chars

0 /5000 chars

Warning: Adding final group rating for this project will prevent expert users from adding new or editing existing reviews

Reviews and Ratings in Deep Funding are structured in 4 categories. This will ensure that the reviewer takes all these perspectives into account in their assessment and it will make it easier to compare different projects on their strengths and weaknesses. Overall (Primary) This is an average of the 4 perspectives. At the start of this new process, we are assigning an equal weight to all categories, but over time we might change this and make some categories more important than others in the overall score. (This may even be done retroactively). Feasibility (secondary) This represents the user\'s assessment of whether the proposed project is theoretically possible and if it is deemed feasible. E.g. A proposal for nuclear fission might be theoretically possible, but it doesn’t look very feasible in the context of Deep Funding. Viability (secondary) This category is somewhat similar to Feasibility, but it interprets the feasibility against factors such as the size and experience of the team, the budget requested, and the estimated timelines. We could frame this as: “What is your level of confidence that this team will be able to complete this project and its milestones in a reasonable time, and successfully deploy it?” Examples:

A proposal that promises the development of a personal assistant that outperforms existing solutions might be feasible, but if there is no AI expertise in the team the viability rating might be low.
A proposal that promises a new Carbon Emission Compensation scheme might be technically feasible, but the viability could be estimated low due to challenges around market penetration and widespread adoption.

Desirability (secondary) Even if the project team succeeds in creating a product, there is the question of market fit. Is this a project that fulfills an actual need? Is there a lot of competition already? Are the USPs of the project sufficient to make a difference? Example:

Creating a translation service from, say Spanish to English might be possible, but it\'s questionable if such a service would be able to get a significant share of the market

Usefulness (secondary) This is a crucial category that aligns with the main goal of the Deep Funding program. The question to be asked here is: “To what extent will this proposal help to grow the Decentralized AI Platform?” For proposals that develop or utilize an AI service on the platform, the question could be “How many API calls do we expect it to generate” (and how important / high-valued are these calls?). For a marketing proposal, the question could be “How large and well-aligned is the target audience?” Another question is related to how the budget is spent. Are the funds mainly used for value creation for the platform or on other things? Examples:

A metaverse project that spends 95% of its budget on the development of the game and only 5 % on the development of an AI service for the platform might expect a low ‘usefulness’ rating here.

A marketing proposal that creates t-shirts for a local high school, would get a lower ‘usefulness’ rating than a marketing proposal that has a viable plan for targeting highly esteemed universities in a scaleable way.
An AI service that is fully dedicated to a single product, does not take advantage of the purpose of the platform. When the same service would be offered and useful for other parties, this should increase the ‘usefulness’ rating.

Soumil Rathi

Project Owner

Project Owner + Developer

View Profile