Source-linked AI summary
Transfer Learning for Evolving Domains
Ricardo Ribeiro Pereira, Jacopo Bono, Hugo Ferreira, Pedro Ribeiro, Pedro Saleiro, Pedro Bizarro, Carlos Soares
TL;DR
Classical transfer-learning settings capture fixed snapshots, although deployed systems often accumulate target instances and labels over time. The paper formalizes this evolving path as TrED, with evaluation across a trajectory of models, and concludes that existing methods do not yet optimize the whole trajectory.
Problem
Method
TrED specifies an environment-fixed data-availability process, a learner-chosen protocol, and cumulative evaluation of the resulting model trajectory.
Source: §3.3 · p. 6, §5 · p. 16Results
Most existing methods address a single regime, and even strong candidates including foundation models do not yet optimize the whole trajectory.
Source: §5 · p. 16Takeaways & Limitations
TrED is a well-posed but unsolved problem, providing a precise criterion for comparing candidate solutions.
Source: §5 · p. 16The classical-regime sequence is an idealized picture because availability changes continuously and transition boundaries are not sharp.
Source: §3.4 · p. 8
Abstract
from arXiv · showhide
Transfer learning explores how to leverage knowledge from various tasks or domains (sources) to enhance predictive performance in related tasks or domains (targets). Typically, transfer learning research is segmented into several isolated sub-areas (such as domain generalisation, domain adaptation, or multi-domain learning), each making distinct assumptions about target data availability, namely how much data and how many labels are available at training time. However, in many real-world applications, data availability is not fixed but evolves over time, as instances and labels are progressively collected from a new domain. Each of the classical settings then describes only a snapshot of a trajectory that a deployed system must traverse in full. We formalise this trajectory as a transfer learning problem in its own right, Transfer Learning for Evolving Domains (TrED), specified by a data availability process fixed by the environment, a learning protocol that the method is free to choose, and an evaluation criterion that scores the whole trajectory of models rather than a single one. Within this formalism, the classical settings are recovered as regimes that a learner may pass through, rather than as separate problems that TrED concatenates. We then examine the transfer learning literature to identify mechanisms that are promising building blocks for a solution, and find that most methods are tailored to a single regime and that even the strongest existing candidates do not yet optimise the whole trajectory. We argue that TrED is a well-posed and unsolved problem, and an important direction for future research.
1 Introduction
Transfer learning usually treats target-data conditions as fixed, but deployed systems often gain target instances and labels over time. The paper formalizes this evolving path as TrED and argues that existing methods do not yet optimize the full trajectory.
1 Introduction
- Transfer learning reuses knowledge from source tasks or domains to improve performance on related targets, under differing assumptions about available data and labels.Obtaining labels can be costly and time-consuming, as in expert labelling for money-laundering detection.Source: §1 · pp. 1–2
- Evolving target-data availability arises in market expansion, regulatory change, and disease outbreaks, where data collection and labels accrue over time.Examples include new-country services, privacy-related changes to data collection, and COVID-19 diagnostic models refined with limited disease-specific data.Source: §1 · p. 2
- Classical transfer-learning settings each capture a snapshot, but deployed systems must traverse the path as target instances and labels accumulate.The path moves from no target data through target data with few or no labels to considerable labelled target data.Source: §1 · p. 2, Figure 1 · p. 3
- TrED formalizes changing data availability as a transfer learning problem judged across the trajectory, while existing methods do not yet optimize that whole path.The paper proposes a data-availability process fixed by the environment, a learner-chosen protocol, and evaluation over the full trajectory.Source: §1 · p. 2
2 Background
The background defines the transfer-learning concepts and scope used in the paper, then distinguishes four classical settings by target-domain data and label availability. These settings provide the fixed-condition cases that TrED later treats as regimes.
2.1 Definitions and notation
- Domains are characterized by feature spaces and marginal distributions, while tasks are probabilistic label relationships; equivalent tasks may share the underlying problem despite differing feature spaces.The paper adopts P(Y|X) to represent label uncertainty, noise, and possible concept drift.Source: §2.1 · pp. 3–4
2 Background
- Transfer learning uses source-domain and task knowledge to learn a target predictor that better approximates the target label distribution than target-only learning.This paper focuses on different domains with equivalent tasks.Source: §2.1 · p. 4
2.2 Classical TL settings
- DG uses no target data, UDA uses unlabelled target data, SDA uses a small labelled target set, and MDL trains across labelled domains including the target.All four settings assume a large pool of labelled source data; MDL optimizes across domains rather than transferring to one target.Source: §2.2 · p. 4
3 Transfer Learning for Evolving Domains (TrED)
Classical transfer-learning settings capture snapshots of target-data availability, while deployments evolve as instances and labels accumulate. TrED formalizes that evolution as a problem evaluated over a learner’s model trajectory.
3 Transfer Learning for Evolving Domains (TrED)
- TrED treats evolving data availability as a transfer-learning problem, with classical settings as regimes rather than separate problems to concatenate.Its instance specifies an environment-determined availability process, a learner-chosen protocol, and an evaluation criterion over the trajectory.Source: §3 · pp. 4–5
3.4 Classical TL settings as regimes of TrED
- Domain generalisation, unsupervised and supervised adaptation, and multi-domain learning arise as regimes along TrED, but their transition boundaries are unobservable and not sharp.The proposed sequence is an idealized picture of availability states, not a precise account of every deployment.Source: §3 · p. 5, §3.4 · p. 8
3.1 Evolution of data availability
- The availability process is environment-controlled and records when instances are collected and when their labels become available, including potentially variable label delays.The learner chooses how to use available target data, not when target labels arrive.Source: §3.1 · p. 5
3.2 Learning protocol
- The learner chooses update times and per-step losses, may adapt its schedule to observed state, and can revisit the full accumulated history.Possible protocols include fixed-interval retraining, switching among algorithms, or continuously updating a model as target labels accumulate.Source: §3.2 · p. 6
3.3 Evaluation
- Evaluation scores the active-model trajectory over the deployment horizon, rewarding learners that reach low target error sooner.The target-risk criterion is distinct from the per-step losses the learner can minimize.Source: §3.3 · p. 6
- Retrospective risk can be estimated only after labels arrive, so current-performance monitoring or model selection must use delayed or surrogate signals.The paper also leaves open how to combine cumulative risk, label-budget performance, and update cost into one criterion.Source: §3.3 · p. 7
3.5 Relation to other learning paradigms
- Online learning is single-stream and typically one-pass, whereas TrED permits retaining and revisiting data; continual learning instead assumes past data may be unavailable.Online transfer learning is closer, but restricts source access to pretrained models and target updates to a single pass.Source: §3.5 · p. 8
- Test-time adaptation restricts source access and uses unlabelled target data, while gradual domain adaptation studies shifting distributions; TrED holds the target distribution fixed.The paper treats combining evolving availability with distribution shift as a natural extension outside its present scope.Source: §3.5 · pp. 8–9
4 Paths to a TrED solution
The paper maps possible approaches to TrED, from a regime-based switching baseline to literature mechanisms and broader-coverage candidates. It identifies foundation models as the strongest existing candidate while noting that limitations remain.
4.1 The stage-switching baseline
- Section 4.1 examines switching among off-the-shelf methods for the current classical regime, explaining why this reasonable baseline is neither trivial nor guaranteed optimal.section 4.1Source: §4 · p. 9
4.2 Existing transfer-learning mechanisms
- Section 4.2 surveys popular transfer-learning mechanisms, asking when they are useful and what must change for them to address a whole trajectory rather than one snapshot.section 4.2Source: §4 · p. 9
4.3 Broader-coverage methods
- Section 4.3 highlights methods that already suggest coverage across a broader range of the TrED continuum.section 4.3Source: §4 · p. 9
4.4 Foundation models
- Section 4.4 presents foundation models as the strongest existing candidate for a complete TrED solution and maps their capabilities and remaining limitations across availability regimes.section 4.4Source: §4 · p. 9
4.1 The stage-switching baseline
The stage-switching baseline delegates successive data-availability regimes to specialized methods. Its timing is difficult under delayed labels, and its piecewise regime structure may exclude better trajectories; the authors frame beating it as a well-posed, unsolved problem.
4.1 The stage-switching baseline
- The stage-switching baseline assigns a specialized method to each data-availability regime and switches at transition times, making it a strong TrED baseline built from mature methods.It selects update times and per-step losses by delegating each interval to a regime-specific algorithm.Source: §4.1 · p. 9
- A switch must rely on a proxy because transition times are continuous and delayed labels can hide target risk; mistiming may make early-label SDA worse than the UDA method it replaces.An oracle that knows transition times could switch optimally, but a deployed learner cannot.Source: §4.1 · p. 9
- A regime switcher only expresses piecewise-constant sequences of named methods, although the best whole-trajectory solution may instead use one model with an objective that changes continuously.A possible objective anneals from domain invariance toward a target-specific objective.Source: §4.1 · pp. 9–10
- Even smoothly weighting the four regime solutions cannot reach a model outside their fixed span, so lower cumulative risk than the best stage-switching baseline remains a well-posed, unsolved problem.The authors characterize the stage-switcher as a greedy, oracle-dependent baseline that serious methods must beat.Source: §4.1 · p. 10
4.2 Building blocks from the TL literature
The review groups transfer-learning mechanisms by whether they modify data, model architecture, or learning objectives, and assesses how each can support a learner as target data availability changes. Mechanisms that adapt gradually across regimes are promising building blocks, but none alone spans the full trajectory.
4.2 Building blocks from the TL literature
- The review groups techniques by whether they act on training data, the model, or the learning objective, assessing their requirements and usefulness across the trajectory.Several methods combine techniques from multiple families.Source: §4.2 · p. 10
4.2.1 Data transformation
- Instance weighting and feature mapping both align source and target distributions, but weighting is cheaper to update while learned mappings typically require refitting.Marginal alignment can use unlabelled target data; conditional alignment may require target labels, and TrAdaBoost down-weights source instances that hurt target performance.Source: §4.2.1 · p. 10
- Pseudo-labelling can bridge label delays and shift continuously from generated to real labels without changing the model or objective.Its early predictions may be unreliable, so confidence or agreement criteria can help avoid training on errors.Source: §4.2.1 · pp. 10, 12
- Feature expansion separates shared from domain-specific representations and works from domain generalisation onward, but its representation grows with the number of domains.The shared-versus-private idea can instead be implemented in model architecture for greater scalability.Source: §4.2.1 · p. 12
4.2.2 Model adaptation
- Shared-shared modular networks are portable from source-only training through later regimes, while private components require domain labels or particular alignment conditions.Hybrid-shared designs may balance early shared-feature performance with later specialization; Figure 2 depicts the recurring architectures.Source: §4.2.2 · p. 12, Figure 2 · p. 12
- Ensembles can add domains without retraining existing members, but their memory and inference costs grow with domain count, motivating member retirement or merging.Combination rules can be tuned as target data arrives, and source members can regularize a mature target model.Source: §4.2.2 · p. 13
4.2.3 Learning objective
- Objective-level tools include distribution alignment, meta-learning, and self-supervised losses; their target-data needs, instability, and label reliability vary across regimes.Loss terms can be added or reweighted; self-supervised losses provide training signal early and can complement other mechanisms later.Source: §4.2.3 · p. 13
- Promising TrED building blocks degrade gracefully as availability changes: pseudo-labels absorb real labels, shared backbones reduce per-domain data needs, and loss emphasis can shift gradually.These mechanisms avoid discrete switches in the relevant label mixture or loss emphasis.Source: §4.2.3 · p. 14
4.3 Methods that adapt to different regimes
Some transfer-learning methods already describe adaptations across adjacent target-data regimes, but the reviewed examples address only the unlabelled-to-labelled transition. Other transitions along the TrED trajectory remain rarely addressed.
4.3 Methods that adapt to different regimes
- Adaptations of pseudo-labelling and supervised loss terms show that a method can extend from unlabelled to labelled target data, but other trajectory transitions remain rarely addressed.The reviewed examples replace pseudo-labels with real labels or add a supervised loss when target labels become available.Source: §4.3 · p. 14
4.4 Foundation Models
Foundation models offer adaptation mechanisms across target-data regimes, making them the strongest existing candidate for one system spanning the TrED trajectory. Yet they remain a starting point: their mechanisms do not decide transitions, exploit new signal in unlabelled target data, or optimize the whole trajectory.
4.4 Foundation Models
- Foundation models are a strong candidate for one system spanning TrED, but the mechanisms reviewed remain building blocks rather than a complete solution.They can operate in some form under every data-availability regime.Source: §4.4 · p. 14
4.4.1 Current literature on FMs
- The reviewed foundation-model landscape spans language and vision models, specialised models for data types and applications, and increasingly multimodal systems.Examples include models for tabular data, time series, audio, medicine, and finance, as well as multimodal systems that process multiple data types.Source: §4.4.1 · pp. 14–15
4.4.2 Applying FMs to TrED
- Foundation models can use zero-shot prompting before target data, distillation with unlabelled target data, in-context learning with a few labels, and fine-tuning or retrieval as labels accumulate.In-context learning is constrained by context size; distillation can yield a smaller student model with lower inference cost and latency.Source: §4.4.2 · p. 15
- Foundation models do not yet optimize the whole trajectory: transition timing, learning from new unlabelled signal, cumulative risk, and pretraining leakage remain unresolved.Honest evaluation must establish that the target domain was genuinely unseen during pretraining.Source: §4.4.2 · p. 15
5 Conclusion
Transfer learning usually treats domain settings as separate snapshots, although data and labels from new domains often accumulate gradually. TrED formalises this evolving trajectory, and the reviewed literature contains no existing method that optimises it throughout.
5 Conclusion
- TrED formalises transfer learning as evolving target-data availability, with performance judged across the trajectory rather than at one snapshot.Its specification fixes the data-availability process by the environment, leaves the learning protocol to the method, and uses a cumulative evaluation criterion.Source: §5 · p. 16
- Classical transfer-learning settings are regimes a learner may pass through, not separate problems that TrED concatenates.Source: §5 · p. 16
- Most reviewed methods are constrained to a single regime, and even the strongest candidates do not optimise the whole trajectory; TrED remains unsolved.The authors describe TrED as well-posed because solutions and comparisons can be stated precisely.Source: §5 · p. 16
- Framing transfer learning as an evolving process is proposed as a path toward adaptive, resilient models for settings where data continually changes.Source: §5 · p. 16