Self-evolving tensor-network computing system with structural mutation

Abstract

A computing system is disclosed in which a tensor-network structure mutates autonomously during execution. A structural mutation unit modifies tensor-network topology concur- rently with tensor contraction execution without halting computation or requiring host processor intervention.

Description

[0001] SELF-EVOLVING TENSOR-NETWORK COMPUTING SYSTEM WITH STRUCTURAL MUTATION

[0002] Technical Field

[0003] The present invention relates to computing systems and hardware accelerators, and more particularly to tensor- network based computing systems capable of autonomously modifying their computational structure during execution.

[0004] Background Art

[0005] Tensor networks are widely used to represent high-dimensional data and computational models by decomposing such data into interconnected tensor nodes. Existing tensornetwork implementations, whether realized in software libraries or specialized hardware accelerators, assume that the structure of the tensor network remains fixed during execution.

[0006] Some systems permit adaptive parameter tuning, rank truncation, or contraction order optimization. However, such adaptations are performed either offline, between execution runs, or as part of multi-epoch training or optimization processes. Structural modification of the tensor-network topology during an active execution pass is generally not supported.

[0007] Accordingly, there exists a need for a computing system capable of autonomously evolving the structure of a tensor network during execution without halting computation.

[0008] Summary of Invention

[0009] The invention discloses a self-evolving tensor-network computing system comprising a tensor-network execution engine, a network state memory, and a structural mutation unit.

[0010] The structural mutation unit autonomously modifies the tensor-network topology during execution, including creation, deletion, splitting, or merging of tensor nodes and modification of contraction edges. Structural mutation is performed concurrently with tensor contraction execution without requiring a global execution halt, pipeline drain, or restart. Technical Problem

[0011] The technical problem addressed by the invention is the inability of existing tensornetwork computing systems to modify their computational structure during execution without interrupting or restarting computation.

[0012] Solution to Problem

[0013] The problem is solved by providing a structural mutation unit that operates autonomously within the computing system and modifies tensor-network topology during execution based on internally generated execution metrics. Structural mutation is performed without host processor intervention and without halting tensor contraction execution. Structural mutation is coordinated using local execution boundaries and atomic topology metadata updates that allow affected contraction pipelines to resume execution without requiring a global execution halt or restart.

[0014] Structural mutation may be initiated and completed while one or more contraction pipelines remain active, without introducing a global synchronization barrier across the execution engine. In certain embodiments, mutation operations are localized to affected tensor nodes or contraction edges, allowing unrelated contraction pipelines to continue execution uninterrupted.

[0015] For contraction pipelines affected by a structural mutation, execution is resumed from a locally consistent boundary without requiring a global execution halt. In such embodiments, the execution engine allows an affected pipeline to complete an in-flight contraction operation or reach a predefined quiescent state within the same execution pass before rebinding to the updated topology metadata. Upon rebinding, the affected pipeline resumes tensor contraction execution using the modified tensor-network topology and associated execution state of the affected pipeline, without requiring reinitialization or restart of the overall execution pass.

[0016] Advantageous Effects of Invention

[0017] The invention provides one or more of the following advantageous effects:

[0018] • Enables autonomous evolution of tensor-network structure during execution;

[0019] • Eliminates the need for execution halts or restarts during structural adaptation;

[0020] • Improves computational efficiency and adaptability across varying workloads;

[0021] • Supports hardware realization without reliance on external control. Brief Description of Drawings

[0022] No drawings are provided for the present invention.

[0023] Description of Embodiments

[0024] In one embodiment, the computing system comprises a tensor-network execution engine configured to perform tensor contractions on a tensor-network defined by tensor nodes and contraction edges.

[0025] The system further comprises a network state memory storing topology metadata representing the tensor nodes and contraction edges. The topology metadata is modifiable during execution.

[0026] A structural mutation unit is configured to monitor execution metrics generated internally by the tensor-network execution engine and to modify the tensor- network topology during execution. Structural mutation may include splitting tensor nodes into multiple lower-rank nodes, merging tensor nodes, inserting or deleting tensor nodes, or inserting or deleting contraction edges.

[0027] The execution metrics used to trigger structural mutation are derived from internal signals, counters, or state variables generated by the tensor-network execution engine during operation. Such metrics may include, for example, measures indicative of tensor rank distribution, resource or energy utilization, or information-flow characteristics within the tensor network. The specific mathematical formulation, normalization, or measurement technique used to derive a given metric may vary across implementations and does not affect the execution semantics of concurrent tensor contraction and structural mutation as described herein.

[0028] Structural mutation is performed concurrently with tensor contraction execution such that no global execution halt, pipeline drain, or execution restart occurs.

[0029] In one embodiment, consistency between tensor contraction execution and structural mutation is maintained using dual-buffered topology metadata. The network state memory maintains an active topology metadata table accessed by the tensor-network execution engine and a shadow topology metadata table modified by the structural mutation unit during execution.

[0030] Upon completion of a structural mutation operation, the structural mutation unit commits the updated topology metadata by performing an atomic pointer update that designates the shadow topology metadata table as the active table. The atomic nature of the update ensures that contraction pipelines unaffected by the mutation continue execution without interruption or perception of delay.

[0031] For contraction pipelines directly affected by a structural mutation, the execution engine allows the affected pipeline to complete an in-flight contraction operation or reach a predefined local quiescent state within the same execution pass. Upon reaching such a boundary, the affected pipeline rebinds to the updated topology metadata and resumes tensor contraction execution using the modified tensor-network topology and associated execution state, without requiring reinitialization or restart of the overall execution pass.

[0032] Structural mutation operations may be limited to a specific subgraph of the tensor network, allowing unrelated subgraphs and contraction pipelines to continue execution concurrently. The mechanisms described herein avoid global synchronization barriers, pipeline drains, or execution restarts.

[0033] In one embodiment, the system is implemented in hardware. In other embodiments, the system may be simulated or emulated on a general-purpose computing platform while preserving the execution semantics of the hardware system.

[0034] The structural mutation unit may be implemented using a variety of control mechanisms, including a finite-state machine, a micro-coded control sequencer, hardwired control logic, or a hybrid hardware-firmware implementation. The structural mutation unit operates autonomously based on execution metrics generated internally by the tensornetwork execution engine, without host processor intervention. The specific implementation mechanism does not affect the execution semantics of concurrent tensor contraction and topology mutation as described herein.

[0035] Examples

[0036] In one example, a tensor- network execution engine performs continuous tensor contractions while a structural mutation unit monitors rank entropy and energy-related metrics. When a rank entropy metric exceeds a threshold, a tensor node is split into multiple lower-rank nodes during execution. When an energy-related metric falls below a threshold, tensor nodes are merged.

[0037] In one illustrative execution scenario, an initial tensor-network topology comprising a plurality of tensor nodes interconnected by contraction edges is loaded into the network state memory. During execution of tensor contractions by the execution engine, the structural mutation unit continuously monitors a rank entropy metric associated with one or more tensor nodes. Upon detecting that the rank entropy metric of a given tensor node exceeds a predefined threshold during an active contraction pass, the structural mutation unit initiates a node-splitting operation. The tensor node is split into multiple lower-rank tensor nodes, and corresponding contraction edges are rewired in the shadow topology metadata table. The updated topology metadata is committed by an atomic pointer swap while tensor contraction execution continues on unaffected pipelines, such that no global execution halt, pipeline drain, or restart occurs. Industrial Applicability

[0038] The invention is applicable to hardware accelerators, high-performance computing systems, machine learning inference engines, scientific simulation platforms, and other computing systems requiring adaptive computational structures.

[0039] Reference Signs List

[0040] Not applicable.

[0041] Reference to Deposited Biological Material

[0042] Not applicable.

[0043] Sequence Listing Free Text

[0044] Not applicable.

[0045] Citation List

[0046] Patent Literature

[0047] None.

[0048] Non Patent Literature

[0049] None.

Claims

AMENDED CLAIMS received by the International Bureau on 22 April 2026 (22.04.2026)1. A computing system comprising: a tensor-network execution engine comprising one or more contraction pipelines configured to perform tensor contractions on a tensor-network; a network state memory storing tensor-network topology metadata including tensor nodes and contraction edges; and a structural mutation unit implemented in hardware, wherein the structural mutation unit is configured to autonomously modify the tensornetwork topology during a single execution pass corresponding to a single inference step or a single simulation timestep, wherein said modification comprises creating, deleting, splitting, or merging tensor nodes or rewiring contraction edges, wherein structural mutation is performed concurrently with tensor contraction execution such that no global execution halt, pipeline drain, or execution restart occurs, and wherein the structural mutation unit operates without host processor intervention, wherein:(i) contraction pipelines unaffected by the topology modification continue execution without interruption;(ii) contraction pipelines affected by the topology modification are permitted to reach a local quiescent boundary corresponding to completion of an in-flight contraction operation, and upon reaching said boundary rebind to updated topology metadata and resume execution within the same execution pass without reinitialization; and(iii) the system operates without a global synchronization barrier across contraction pipelines during topology modification and commit.[0001][0002]STATEMENT UNDER ARTICLE 19 (1 )[0003]The amendments to Claim 1 introduce additional limitations relating to execution behavior of contraction pipelines during topology modification.[0004]In particular, the amended claim specifies that contraction pipelines unaffected by topology modification continue execution without interruption, while contraction pipelines affected by topology modification are permitted to reach a local quiescent boundary before rebinding to updated topology metadata and resuming execution.[0005]The amended claim further clarifies that topology modification and commit are performed without a global synchronization barrier across contraction pipelines.[0006]These amendments are supported by the description, in particular the sections describing concurrent execution, dual-buffered topology metadata, atomic commit, and local quiescent boundary rebinding.