A data maturity-deployment scale framework is introduced, arguing that engineering AI should be understood not as a search for a single best model, but as an evolving intelligence strategy that co-adapts with system maturity, operational context, and deployment scale.
Abstract
Machine learning (ML) has driven major advances across engineering domains, from early data-driven approaches to deep learning and, more recently, foundation models. Despite rapidly increasing model capabilities, the practical impact of ML in engineering systems remains limited. We argue that this gap arises from a fundamental misalignment between prevailing ML paradigms and the realities of engineering environments. Modern ML methods, largely developed for data-rich and statistically stable domains, are fundamentally ill-suited for engineering systems characterized by sparse, noisy, and context-dependent data, uncertain ground truth, non-stationary dynamics, strict deployment constraints, and strong physical and domain knowledge requirements. To address this mismatch, we first identify four core challenges: data and benchmark limitations, distribution shift and generalization, resource and deployment constraints, and insufficient integration of domain knowledge. Then we propose a corresponding four-shift framework: (i) from data quantity to engineered data foundations, (ii) from static learning to adaptive and robust learning, (iii) from model scaling to resource-efficient and deployment-aware intelligence, and (iv) from general-purpose models to domain-specific, knowledge-informed foundation models. Building on this framework, we outline four research directions toward Engineering Intelligence: data foundation intelligence, reliable and generalizable intelligence, sustainable and deployable intelligence, and knowledge-centric intelligence. Finally, we introduce a data maturity-deployment scale framework, arguing that engineering AI should be understood not as a search for a single best model, but as an evolving intelligence strategy that co-adapts with system maturity, operational context, and deployment scale.
Deep neural networks excel in various tasks but struggle to generalize across evolving data distributions, leading to significant performance degradation under domain shifts. Domain incremental learning (DIL) addresses this challenge by enabling models to continuously adapt while retaining prior knowledge. Among existing DIL approaches, the parameter-isolation paradigm achieves state-of-the-art performance. However, these methods often adopt a one-size-fits-all approach to adapt to new domains, resulting in either insufficient learning capacity or redundant parameters. In this work, we propose BPG, a unified framework that addresses both challenges through two complementary components: BPG-Adapter, which dynamically determines each domain's adapter hidden dimension based on domain-specific feature separability, and BPG-Inference, a soft domain mixture strategy that integrates multiple domain-specific models at test time, mitigating domain ID misselection. Experimental results on DomainNet, CDDB, and CORe50 demonstrate that BPG consistently outperforms uniform adapter-based approaches and hard domain selection strategies, achieving state-of-the-art average accuracy while reducing forgetting to as low as 0.22% on DomainNet.
Qiang Wang, Songlin Dong, Shaokun Wang et al.· 0 citations
Deep learning has enabled the wide-spread deployment of computer vision beyond laboratory settings. Consequently, computer vision has become the main bridge for intelligent systems observing the real, physical world. Yet, a fundamental barrier persists: the substantial volume of precisely labeled training data required for effective supervised learning. Especially in industrial settings, where public datasets rarely match domain-specific requirements, the combined effort of data capture, cleansing, and annotation can account for up to 80% of project costs, constraining deployment agility and accessibility. This data hunger manifests distinctly across tasks. For example, semantic segmentation can demand pixel-perfect masks, while anomaly detection often requires extensive normal examples and, in some cases, defective examples, which are rare by nature. Some failure cases might even be entirely unknown while the inspection pipeline is being built. This dissertation develops and evaluates methodologies that reduce dependency on extensively labeled datasets. The central research objective is to demonstrate that strategic approaches, leveraging sparse supervision, pre-trained multimodal knowledge, and transferable embeddings, can achieve competitive performance while substantially reducing annotation requirements, thereby making deep learning more practical for resource-constrained deployment.
Three empirical investigations address complementary aspects of data efficiency. First, a “Lazy Labels” technique enables U-Net architectures to learn semantic segmentation from sparse point or scribble annotations by zero-masking unlabeled pixels during loss computation, requiring modifications only to label tensor handling. Second, Multimodal Large Language Models (MLLMs), including GPT-4V and Qwen-VL-Chat, are evaluated for visual anomaly detection through a training-free Visual Question Answering (VQA) framework, using natural language prompts with zero to a few reference images. Third, a comparative study contrasts pre-trained foundation model embeddings (MegaDescriptor) against task-specific training for visual re-identification, examining both frozen and fine-tuned transfer learning under standard and one-shot conditions. The sparse segmentation experiments demonstrate that models trained with point annotations covering ≈1% of pixels achieve performance within 1− −3 F1-score percentage points of fully supervised baselines, while producing spatially coherent dense predictions. For anomaly detection, GPT-4V achieves F1 = 0.92 in a one-shot configuration on MVTec AD without task-specific training, approaching specialized detector performance. In re-identification, fine-tuned pre-trained embeddings reach 91.6% top-1 accuracy under full supervision and notably outperform from-scratch training in data-limited one-shot scenarios, validating their superior data efficiency.
The findings demonstrate that introducing prior structure, through strategic sparse supervision, pre-trained multimodal knowledge, or transferable representations, consistently compensates for limited labeled data across diverse vision tasks. While each approach operates within specific constraints, including sparse label distribution, sensitivity to foundational capabilities, and domain similarity, they collectively demonstrate viable pathways for reducing the data bottleneck. The work contributes observations and validated techniques for practitioners: zero-masking for direct sparse supervision, prompt-based MLLM deployment for rapid baselines, and transfer learning strategies optimized for data scarcity. These methodologies advance the practical deployment of computer vision in contexts where extensive annotation remains prohibitive.
Deep learning has achieved state-of-the-art performance across domains such as computer vision, NLP, and healthcare, but it often requires large datasets and high computational resources. Pretraining and fine-tuning have emerged as effective strategies to improve performance, reduce training cost, and enhance generalization. Pretraining learns transferable representations from large-scale data, while fine-tuning adapts models to specific tasks with limited labeled data. This paper provides a comprehensive study of various pretraining methods (supervised, unsupervised, self-supervised) and fine-tuning techniques, including full-model and parameter-efficient approaches. A unified framework is proposed to integrate both processes in a deep learning pipeline. Experimental findings show that pretrained models outperform those trained from scratch in terms of accuracy, convergence speed, and robustness. The study also highlights challenges, trade-offs, and emerging trends such as foundation models and multimodal learning, emphasizing the importance of these techniques in advancing deep learning systems.
Kwame Nkosi· International Journal of App...· 0 citations
Recent advances in deep learning have enabled remarkable success across computer vision tasks using massive datasets, yet the persistent challenge of domain shift-where training and testing distributions differ-continues to degrade model performance. While Domain Generalization techniques aim to learn invariant representations across multiple domains, their reliance on diverse training domains limits applicability in settings where multiple source domains are not available. This survey examines Single Domain Generalization (SDG) approaches that leverage targeted inductive biases to enhance out-of-domain (OOD) performance using a single source domain, addressing key questions regarding their effectiveness across various domain shifts, different model architectures such as CNNs and Vision transformers, and the impact of shape bias for OOD generalization. These questions are crucial as they directly inform the development of robust models that can reliably operate under real-world conditions, guiding the research community in overcoming limitations of current techniques. Our findings indicate that although SDG methods improve performance under highly domain distinguishable stylized shifts, they exhibit limited robustness to background, correlation and corruption shifts on larger datasets, and increased shape bias does not consistently yield enhanced OOD performance, thereby underscoring the need for further research into developing more resilient and generalizable models. For reproducibility, the implementation and code used in our benchmark are available on our official https://github.com/Kavindya98/SDG-Benchmark GitHub repository.
K. Imbulgoda, Ruwan Tennakoon, W. Chuah et al.· International Journal of Com...· 0 citations
Generalizing machine learning models to environments that differ from their training distribution remains a critical hurdle, particularly when data from the target domain is entirely or partially unavailable. We propose Generative Meta-Learning with Human Feedback (GMHF), a novel framework that bridges this domain gap by leveraging expert intuition to guide data synthesis. Grounded in a theoretical analysis of generalization error, we derive bounds demonstrating that aligning the distribution of generated data with human beliefs regarding the target physics significantly mitigates risk. GMHF operationalizes this insight by employing a Conditional Neural ODE (cNODE) as a generative digital twin, coupled with a Reinforcement Learning (RL) agent. The agent iteratively refines the latent physical parameters of the generated trajectories based on feedback, effectively steering the meta-learner toward the unobserved target distribution. Empirical validation on a nonlinear Duffing oscillator shows that GMHF substantially reduces deployment loss as expert reliability increases, and that the divergence between generated and target data falls under reliable feedback, directly corroborating the divergence-minimisation mechanism predicted by our theory. Further experiments on a non-dynamical probabilistic model confirm that the framework extends beyond ODE-governed systems, establishing human-AI collaboration as a rigorous catalyst for robust generalisation under distribution shift.
Reinforcement Learning (RL) provides a principled framework for sequential decision-making in complex environments. However, state-of-the-art Deep Reinforcement Learning (DRL) algorithms typically require large amounts of training data and often fail to generalize beyond small-scale training scenarios, even on standard benchmarks.
We propose a neuro-symbolic DRL approach that incorporates background symbolic knowledge to improve both sample efficiency and generalization to more challenging, unseen tasks. Specifically, partial policies learned in simple domain instances, where high performance can be achieved reliably, are transferred as structured priors to accelerate learning in more complex environments, eliminating the need to tune DRL parameters from scratch.
Our method represents partial policies as logical rules in the Answer Set Programming (ASP) formalism and performs online reasoning to guide training through two complementary mechanisms: (i) biasing the action distribution during exploration, and (ii) rescaling Q-values during exploitation.
This integration of ASP reasoning with DRL enhances interpretability and trustworthiness while accelerating convergence, particularly in sparse-reward settings and tasks with long planning horizons, without introducing significant computational overhead.
We empirically evaluate our approach on challenging variants of gridworld environments under both fully and partially observable settings. Results demonstrate consistent performance improvements over a state-of-the-art reward machine baseline.
Celeste Veronese, Alessandro Farinelli, Daniele Meli· Proceedings of the TwentyThi...· 0 citations