Selective Live Training Architecture (SLTA): Capability-Selective Continual Adaptation for Deployed Large Neural Networks

Abstract Large neural networks are traditionally separated into two fundamentally different operational states: training and inference. During training, model parameters are modified through an optimization process, whereas during inference, model parameters are typically frozen. This separation makes deployed models difficult to continuously adapt to new information, skills, errors, and environmental feedback. This paper proposes the Selective Live Training Architecture (SLTA), a deployment-oriented architecture that enables a neural network to perform inference while selectively updating only the parameters associated with explicitly selected capabilities. When live training is disabled, the model operates as a conventional frozen inference system. When live training is enabled, the model continues performing inference while a capability controller determines which parameter groups are permitted to receive updates. For example, an SLTA-based multimodal model could enable live learning for programming while keeping mathematical reasoning, vision, general knowledge, and unrelated parameters frozen. The architecture introduces a Capability Parameter Registry (CPR), selective parameter masks, a live learning controller, candidate update validation, replay, and rollback mechanisms. The central SLTA update is formulated as Wt+1=Wt+M⊙ΔWt,W_{t+1}=W_t+M\odot\Delta W_t, where WW represents the model parameters, ΔW\Delta W represents a proposed learning update, and MM represents the capability-selective parameter mask. SLTA is compatible with full-weight selective training, parameter-efficient fine-tuning methods such as LoRA, adapter modules, quantized base models, and mixture-of-experts (MoE) architectures. The paper presents the proposed architecture, mathematical formulation, algorithms, experimental methodology, ablation studies, resource analysis, and comparison with existing adaptation approaches. Because SLTA has not yet been implemented and experimentally evaluated, all numerical results presented in this paper are explicitly identified as hypothetical estimates and research targets rather than measured results. Index Terms— continual learning, live training, selective fine-tuning, parameter-efficient fine-tuning, LoRA, mixture-of-experts, adaptive neural networks, catastrophic forgetting, lifelong learning, capability isolation Research Status and Disclosure This work presents the Selective Live Training Architecture (SLTA) as a proposed research architecture. The architecture and its mathematical formulation are presented as a research proposal and have not yet been experimentally validated. All numerical values identified as estimated, hypothetical, or target results are not experimental measurements. The authors make no claim that the individual components of SLTA—including continual learning, parameter-efficient fine-tuning, parameter masking, sparse adaptation, replay, rollback, or mixture-of-experts specialization—are novel when considered individually. The proposed contribution is their integration into a capability-selective live adaptation architecture. The novelty claim of this work is therefore primarily architectural and system-level: SLTA combines runtime-selectable learning, capability-specific parameter selection, explicit parameter masking, candidate-update validation, and rollback mechanisms into a unified deployment-oriented framework for continual adaptation.