Foundations of a Developmental Design Paradigm for Integrated Continual Learning, Deliberative Behavior, and Comprehensibility

TL;DR

This work proposes a novel, developmentally inspired AI design that integrates a gradient-free Modeller, a goal-directed Planner, and a Behavior Encapsulation mechanism to tackle the intertwined challenges of continual learning, information reuse, comprehensibility, and deliberative planning. The Modeller builds a discrete topological representation of environment structure using state variables (BSVs, DSVs, CSVs) and learns through local variation and selection, yielding continual learning without destructive forgetting. The Planner derives executable action networks from the Modeller’s model, while Behavior Encapsulation extracts stable subpolicies in a hierarchical, reusable form, enhancing interpretability. The framework is extended to high-dimensional observations via State Polynetworks (SPNs) and Network Refinement with rerelation (MNR), demonstrated on MNIST-based shape tasks and validated in simple FSM experiments that show robust continual learning, planning, and interpretability, with a thoughtful discussion of computational viability and directions for future work. Together, these components offer a unified approach to overcoming major ML limitations by learning structured environment models that support symbolic-style planning and explainable behavior, including potential extensions to more complex visual domains and hierarchical representations.

Abstract

Inherent limitations of contemporary machine learning systems in crucial areas -- importantly in continual learning, information reuse, comprehensibility, and integration with deliberate behavior -- are receiving increasing attention. To address these challenges, we introduce a system design, fueled by a novel learning approach conceptually grounded in principles of evolutionary developmental biology, that overcomes key limitations of current methods. Our design comprises three core components: The Modeller, a gradient-free learning mechanism inherently capable of continual learning and structural adaptation; a planner for goal-directed action over learned models; and a behavior encapsulation mechanism that can decompose complex behaviors into a hierarchical structure. We demonstrate proof-of-principle operation in a simple test environment. Additionally, we extend our modeling framework to higher-dimensional network-structured spaces, using MNIST for a shape detection task. Our framework shows promise in overcoming multiple major limitations of contemporary machine learning systems simultaneously and in an organic manner.

Figures (19)

• Figure 1: Sample formation of a CSV in a continual manner. The relationship to be modelled is $Y = X0\ and\ !X2$ ("!" denotes "not"). Black and orange arrows represent positive and negative sources for CSV $C0$ respctively. $Xi$ can be interpreted either as single or grouped SVs. (a) Initial state with no relation formed between $X0-3$ and $Y$. (b) $X0, X1 \rightarrow Y$ observed. Positive connections hypothesizing both $X0$ & $X1$ are required for Y are formed. (c) $X0 \rightarrow Y$ is observed. $X1$ is deduced unnecessary for $Y$. (d) $X0, X2, X3 \rightarrow !Y$ observed. $Y$ is hypothesized to be suppressed by $X2$ and $X3$. (e) $X0, X2 \rightarrow !Y$ observed. $X3$, seen unnecessary for suppression of $Y$, refined. Correct structure learned and is stable from now on.
• Figure 2: Example of upstream conditioning, continuing from Figure \ref{['fig:csvform']}. Assume that the unconditionality flag of $C0$ is set following an observation that $(X0,\ !X2)$ did not result in its activation (see main text). (a) $X0, !X2, X4, X5 \rightarrow Y$ observed. $C0$ is observed to be active. A new CSV $C1$ conditioning $C0$ formed. Note that $(X4, X5)$ alone will not predict activation of $C0$ if $C0$’s sources are not also active. (b) New conditioners are also subject to the CSV processes: Here, the source $X5$ of $C1$ has been refined, and new conditioners $C2$ and $C3$ are formed. Multiple conditioners represent alternative paths: In this case, $C0$ is expected to be active when sources of either $C1$ or $C2$ is active. Any logical function can hence be incorporated in a conditioning pathway in a minimal and ongoing manner without destroying past knowledge.
• Figure 3: Illustration step-by-step upstream generation of action network, operating on different SV types (see the text for details). $BX$, $CX$ and $GX$ stand for BSV, CSV and GSV nodes respectively, (A) for activation, (0) for nonactive state. Black arrows are positive sources and precondition targets, green arrows are constituent (dashed) and constituency (solid) relations. The node that is extended at each step is highlighted in red. (a) Step 1. CSV $C0$ is opened. (b) Steps 2-4. Each step opens up one of the sources of previous step. Possible interrelations (e.g. $B2$ for $C1$, $G0$) do not need reopening if they already exist.
• Figure 4: Illustrative example for the aim of behavior encapsulation process. To the left are two action networks that represent two alternative pathways, split from the unified AN generated by the planner (node names are placeholders and can be any SV type or target effect). We want to encapsulate the pathways between X and Z. For that; all paths that are reliably present in both networks are identified and a new encapsulated AN (EAN) is formed with them (right). Each encapsulated edge (dashed) in EAN includes copies of subnetworks that corresponded to this path in the original AN variants; which can be further encapsulated in subgroups via a recursive call (e.g. edge (D0,Y) would include two pathways; first one formed only of E0, the second of C2 and E1). The EAN on right can be regarded as the subpolicy for realization of Z from X.
• Figure 5: Illustration of network refinement with rerelation. In (c), highlighted edges are created through rerelation. Paths $(A,D)$ and $(A,C)$ exist in both networks but are mediated by different intermediaries (B and K respectively), leading to refined intermediaries and new edges. Similarly, path $(Z,C)$, mediated by $(Y,X)$ in the source and $(L)$ in the refiner, is refined. Edge $(A,Z)$ is removed as it lacks a corresponding path in the refiner SPN. Edge $(A,B)$ is preserved unchanged, as it appears in both networks, despite differing successors of B. (Node positions are illustrative and irrelevant to refinement.)

Theorems & Definitions (6)

• Definition 1
• Definition 2
• Definition 3
• Theorem 1
• Definition 4
• Definition 5