Evolving Knowledge Structures

chevron-icon
RFP Proposals
Top
chevron-icon
project-presentation-img
Expert Rating 4.2
evolveai
Project Owner

Evolving Knowledge Structures

Expert Rating

4.2

Overview

The best use of the evolutionary process is to reward and evolve meaningful structure within complex systems. The proposed research will explore the development of new evolutionary processes for evolving DNNs with intrinsic functional and semantic decomposition. This will support rapid development of novel DNN architectures that are composable, extensible and able to be grounded to the real world, while being naturally transparent rather than opaque.

RFP Guidelines

Evolutionary algorithms for training transformers and other DNNs

Complete & Awarded
  • Type SingularityNET RFP
  • Total RFP Funding $40,000 USD
  • Proposals 8
  • Awarded Projects 1
author-img
SingularityNET
Aug. 12, 2024

Explore and demonstrate the use of evolutionary methods (EMs) for training various DNNs including transformer networks. Such exploration could include using EMs to determine model node weights, and/or using EMs to evolve DNN/LLM architectures. Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is an example of one very promising evolutionary method among others.

Proposal Description

Project details

Introduction

The rapid advancement of Deep Neural Networks (DNNs) has driven significant improvements in areas like computer vision, natural language processing, and autonomous systems. Despite their success, DNNs often operate as "black boxes," with their internal structures and decision-making processes opaque. This limits trust and adoption, particularly in high-stakes domains like healthcare, finance, and autonomous systems. Additionally, designing new DNN architectures involves a time-consuming, resource-intensive trial-and-error process. While directly evolving all the weights of large networks is computationally impractical in this researcher's opinion, evolving the architecture itself offers a promising alternative. This approach enables the evolution of modular components that can leverage prior training, leading to more efficient and adaptive learning without sacrificing the broad generalization capabilities of DNNs.

In this proposal, we introduce an evolutionary approach to DNN architecture that incorporates three levels of evolution: composition and decomposition, evolution guided by prior training, and evolution of weights at multiple levels. While the primary focus of this research is on prototyping modular architectures and evolving their components, we will also explore the feasibility of grounding these architectures in real-world knowledge. Grounding will be treated as a feasibility study with limited scope.

The approach involves evolving modular components through crossover, mutation, and selection, leveraging prior training to speed up convergence, and restricting weight evolution to specific modules and sub-problems. These evolving parts will operate at different "speeds," with some sub-populations evolving quickly to specialize in narrow tasks, while others evolve more slowly to maintain general adaptability. We will demonstrate this approach using the MeTTA language within the Hyperon framework.


Background and Motivation

DNNs have achieved impressive results but face challenges that hinder their broader adoption:

  • Opacity: The decision-making process of DNNs is often not interpretable, limiting their trustworthiness in high-stakes fields.
  • Complexity and Rigidity: Designing new architectures is a resource-intensive process, and once trained, networks are difficult to adapt.
  • Lack of Grounding: Many DNNs lack real-world connections, making them less useful for practical applications.

Evolutionary algorithms have been successful in optimizing complex systems, yet their application to DNNs has been limited. Most existing methods treat DNNs as monolithic structures, neglecting modular decomposition and semantic grounding. Our approach aims to address these challenges by evolving DNN architectures with both functional and semantic decomposition, enabling the creation of more interpretable, adaptable, and efficient networks grounded in real-world knowledge.


Research Objectives

This research aims to develop evolutionary algorithms for evolving DNN architectures with intrinsic functional and semantic decomposition. The specific goals are:

  1. Development of Evolutionary Algorithms for DNNs: Design evolutionary algorithms that promote modularity and semantic clarity, breaking complex tasks into interpretable components.

  2. Intrinsic Functional Decomposition: Evolve modular DNN components that can decompose tasks into reusable modules, allowing for task-specific optimization and promoting cross-domain reuse.

  3. Semantic Grounding of DNNs: Explore methods for grounding the evolution of network components in real-world knowledge, using knowledge bases, domain-specific constraints, or sensory data to ensure that modules are both functionally specialized and semantically relevant.

  4. Composability and Extensibility: Create DNN architectures that are composable and extendable to adapt to new tasks or data sources, facilitating scalable AI solutions.

  5. Transparency and Interpretability: Ensure that evolved architectures are inherently transparent, with modular components that enhance understanding of the model's decision-making processes.


Methodology

Design of Evolutionary Algorithm

The evolutionary algorithm will be designed to support modular, interpretable networks that decompose complex tasks into specialized submodules. The algorithm will operate at three levels:

  1. Composition/Decomposition: Evolve the overall composition of the network, determining how smaller components combine to form larger architectures.

  2. Bootstrapping with Prior Training: Leverage pre-existing knowledge (e.g., back-propagation, pre-trained models) to inform the evolutionary process, speeding up convergence without sacrificing generalization.

  3. Localized Weight Evolution: Evolve network weights within specific modules or sub-populations. This will involve intra-modular weight evolution (optimizing within a module) and extra-modular weight evolution (optimizing between modules). This approach makes the process feasible and tractable, limiting weight evolution to smaller, manageable regions, and reducing computational costs.

Exploration of Grounding Mechanisms

While the primary focus is on evolving modular components, we will explore the feasibility of integrating semantic grounding into the evolutionary process. This exploration will be early-stage, aiming to determine how real-world knowledge could be embedded into DNNs. Key areas include:

  • Automated Semantic Grounding: Investigate how external knowledge sources and contextual embeddings can ground components in real-world data.

  • Context-Aware Evolution: Explore how self-supervised learning and active learning techniques could allow modules to refine themselves based on real-time feedback, optimizing their alignment with task objectives.

  • Imputing the Primary Focus of Modules: Explore mechanisms to automatically determine the primary focus of a module within the data space and leverage that focus to ensure that modules specialize in distinct tasks (e.g., object detection, syntactic structures) and evolve in a way that optimizes network performance while avoiding redundancy.


MeTTA and Hyperon Framework

We will demonstrate the proposed evolutionary strategies using the MeTTA language within the Hyperon framework. MeTTA will allow us to define and evolve modular DNN architectures with the flexibility to support composition and decomposition of network components. Evolutionary operators such as crossover, mutation, and selection will be implemented to support modularity, semantic grounding, and functional diversity. MeTTA will provide a robust environment to define modular tasks and objectives, guide evolutionary processes, and track the evolving system's progress in real-time.


Evaluation Framework

The effectiveness of our approach will be assessed using both traditional metrics (e.g., accuracy, computational efficiency) and novel metrics focused on modularity, composability, and adaptability. A benchmark will be used from domains such as natural language processing or autonomous robotics to conduct experiments. Additionally, we will assess the feasibility of real-world grounding by exploring how well modules align with real-world knowledge and task objectives. This will guide the evolutionary process to ensure that the system evolves in a way that is both efficient and practical for real-world applications.


Conclusion

By leveraging the Hyperon framework and the MeTTA language, this research will advance the development of modular, composable, and interpretable DNNs. The outcomes will offer insights into how such architectures can evolve to meet complex, dynamic tasks, providing a foundation for the future of transparent, adaptable AI systems.

Notes on references:

"Let's Evolve Intelligence, not Solutions" shows recent work exploring how we might compose AIs within an evolutionary framework involving specialized sub-populations, as well as discussing how we might ground the learning of an AI to the real-world.

"A meta-model perspective and attribute grammar approach to facilitating the development of novel neural network models," shows work on using a neural markup language to support the evolution of the architecture, behaviors and weights of large, complex neural networks that are modular and specialized for subtasks as determined by the evolutionary process itself.

"A Genetic-Algorithm-Based Reconfigurable Scheduler" shows work on specifying the elements and constraints of a reconfigurable evolutionary algorithm for a wide range of scheduling problems via a specialized control language.

Open Source Licensing

BSD - Berkeley Software Distribution License

Any code developed will be made available to the public.

Links and references

"Let's Evolve Intelligence, not Solutions" Genetic Programming Theory and Practice Workshop XX https://drive.google.com/file/d/1i71SvtvfXlGP53627JqSj8Qb6XHNv7Ff/view

"A meta-model perspective and attribute grammar approach to facilitating the development of novel neural network models" https://drive.google.com/file/d/1TPQ4NG5fJhl2b7Gj9ikfQevyLaXVsZPD/view 

"A genetic-algorithm-based reconfigurable scheduler" (Studies in CI 49) https://drive.google.com/file/d/1tMr0TVm1ZRt3Oyz3jWLmb6SUcuYrjPQq/view 

Proposal Video

Not Avaliable Yet

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

  • Total Milestones

    2

  • Total Budget

    $40,000 USD

  • Last Updated

    8 Dec 2024

Milestone 1 - Interim report

Description

The interim report will present the results of the research conducted and methods developed with a focus on evolving architecture.

Deliverables

Report on progress

Budget

$20,000 USD

Success Criterion

Show clear translation of research concepts into MeTTa, with results from preliminary experiments demonstrating feasibility of operators for decomposition and composition of DNNs and leveraging prior training.

Milestone 2 - Final Report and demonstration

Description

Final report and associated demonstration summarizing the research performed with an additional emphasis compared to the first milestone on localized evolution of network weights and the exploration of grounding methods

Deliverables

Report and demonstration of working code performing evolution of DNN within Hyperon.

Budget

$20,000 USD

Success Criterion

Evolutionary process within Hyperon shows tangible benefits in either evolution of architecture or localized evolution of weights, ideally in both. Establish potential feasibility of grounding methods, though prototyping and testing of semantic grounding are out of scope due to computational resources required.

Join the Discussion (0)

Expert Ratings

Reviews & Ratings

Group Expert Rating (Final)

Overall

4.2

  • Feasibility 4.8
  • Desirabilty 4.5
  • Usefulness 4.3
  • Expert Review 1

    Overall

    3.0

    • Compliance with RFP requirements 4.0
    • Solution details and team expertise 3.0
    • Value for money 3.0
    Interesting multi-level formulation

    There doesn't seem to be a strong backing of the proposed direction, nor the detail and evaluation criteria desired for the scientific nature of the RFP. The proposal does bring up a valuable research direction in decomposition -> localized training -> re-composition.

  • Expert Review 2

    Overall

    5.0

    • Compliance with RFP requirements 5.0
    • Solution details and team expertise 5.0
    • Value for money 5.0
    Solid unique approach

    Unique and ambitious approach to evolving modular, interpretable, and grounded DNN architectures. If executed well, has the potential to significantly advance the transparency, adaptability, and scalability of AI systems while contributing meaningfully to AGI development.

  • Expert Review 3

    Overall

    5.0

    • Compliance with RFP requirements 5.0
    • Solution details and team expertise 5.0
    • Value for money 5.0
    This is an exceptional proposal that goes beyond what's asked for in a very intelligent way

    Using EDAs for architecture discovery and semantic grounding makes sense... expanding the task in this way makes it a lot to bite off but the proposer seems to understand this too and will take a step by step approach. This seems like a great researcher to bring into the Snet / Hyperon fold..

  • Expert Review 4

    Overall

    4.0

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

    Excellent and detailed proposal addressing and going beyond the RFP. Good focus of DNN limitations and how Ems can help with them. Also good discussion of traditional metrics as well as novel metrics focused on modularity, composability, and adaptability. Could have presented better and more detailed milestones and additional team information.

feedback_icon