Transfer Utility: Concepts and Applications

• Transfer Utility is a multifaceted concept defined by how transferred representations, welfare properties, or risk sharing maintain or enhance performance across domains.
• Empirical studies in machine learning show that pre-trained models and task-specific transfer methods boost downstream accuracy, with DenseNet161 reaching 83% accuracy versus 58% without transfer.
• In economics, transferable utility underpins models of surplus redistribution and equilibrium matching, while in systems it measures efficiency in data and semantic transfers.

“Transfer utility” is not a single technical doctrine. In the literature represented here, it denotes several related but non-equivalent ideas: the usefulness of transferred representations, layers, or source datasets in machine learning; the transfer of regularity, curvature, or equity properties across utility representations in economic theory; transferable utility as an equilibrium structure in matching and risk transfer; and application-specific measures of usefulness retained after semantic, masked, object, stream, or physical transfer (Huang et al., 2021, Pireddu, 2010, Galichon et al., 28 Nov 2025, Wang et al., 2023). Across these usages, a common theme is that transfer is evaluated not merely by movement or reuse itself, but by what structure, welfare, or task performance survives or improves under that transfer.

1. Main senses of the term

In the surveyed literature, the expression is polysemous. In transfer learning, “transfer utility” is usually operationalized as downstream target performance or as “transferability,” namely the usefulness of a pretrained model, source dataset, or layer for a new task. In economics, the phrase splits into at least three lines: transferable utility in matching and risk-sharing models, transfer principles such as generalized Pigou-Dalton transfers in welfare comparisons, and property transfer between primitive von Neumann–Morgenstern utility and induced expected utility. In systems and communications, the term appears in utility-oriented data movement, privacy-utility trade-offs, and semantic transfer, where the objective is the usefulness of transferred content rather than raw fidelity (Shadman et al., 2020, Dubey et al., 2020, Asikis et al., 2017).

Domain	Operative meaning	Representative work
Transfer learning	Usefulness of pretrained features, layers, or sources for a target task	(Huang et al., 2021)
Welfare and utility theory	Transfer of utility properties or equity under transfers	(Pireddu, 2010)
Matching and risk transfer	Transferable utility as a market or equilibrium structure	(Galichon et al., 28 Nov 2025)
Communications and systems	Utility retained or optimized when data or semantics are transferred	(Wang et al., 2023)

This suggests that encyclopedia treatment is best organized by technical usage rather than by a single universal definition.

2. Transfer utility in machine learning

In transfer learning, utility is typically defined by improvement on a target task under limited target supervision. A clear empirical formulation appears in “The Utility of Feature Reuse: Transfer Learning in Data-Starved Regimes” (Shadman et al., 2020). That study evaluates utility by comparing full fine-tuning, feature extraction, and training from scratch on a binary transporter erector launcher classification task with only 100 TEL images and 130 non-TEL images, plus an out-of-distribution set with 31 TEL images and 34 non-TEL images. Utility is assessed through validation accuracy, OOD test accuracy, and performance differences across transfer regimes. Across six CNN architectures, transfer consistently improves OOD accuracy; for example, DenseNet161 reaches 83% with transfer versus 58% without transfer, and SqueezeNet v1 reaches 80% versus 48% (Shadman et al., 2020). The same study argues that both feature reuse and overparameterization contribute, but that the benefit of model size levels off.

A related but linguistically specialized formulation appears in “The Utility of General Domain Transfer Learning for Medical Language Tasks” (Ranti et al., 2020). There, transfer utility is tested through radiology report classification on 1,977 labeled head CT reports drawn from a corpus of 96,303 reports. General-domain BERT and BioBERT both achieve a sample-weighted average $F_1$-score of 0.87, while randomized BERT reaches 0.39, the LSTM 0.35, and logistic regression 0.53 (Ranti et al., 2020). The paper’s central claim is that general text transfer learning may be sufficient for at least some medical NLP tasks, and that further biomedical pretraining on PubMed and PMC does not necessarily improve downstream utility when the additional corpus is not closely aligned with clinical report language. A plausible implication is that transfer utility depends not only on domain specificity in the abstract, but on the match between the transferred representation and the actual linguistic distribution of the target task.

Task-specific utility transfer also appears in “Utility-Oriented Underwater Image Quality Assessment Based on Transfer Learning” (Chen et al., 2022). That paper defines utility-oriented IQA as “the quality evaluation of an image considering its utility to complete a vision-based task,” and instantiates the task as underwater fish detection. Its Underwater Image Utility Measure transfers multi-scale features from Fish-YOLOv4 and achieves PLCC 0.8473, SRCC 0.8377, KRCC 0.6544, and $C_0$ 0.8794 on the Underwater Image Utility Database (Chen et al., 2022). Here, transfer utility is not generic representation reuse but task-conditioned utility prediction: a transferred detector backbone is useful because it already encodes the fish-relevant structures that define the downstream notion of quality.

3. Estimating and controlling transfer utility

A separate research line treats transfer utility as something to be estimated before expensive target-task training. “Frustratingly Easy Transferability Estimation” (Huang et al., 2021) formalizes true transferability as expected downstream target performance after optimal transfer learning and proposes TransRate as a practical estimator. For a transferred feature extractor $g$, the score is defined as mutual information between target-task features and target labels, approximated by coding-rate reduction: $$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$ The representation-level interpretation is that transfer utility depends on both completeness and compactness. Empirically, TransRate is evaluated on 32 pre-trained models and 16 downstream tasks, and is reported as roughly 2,160×–6,264× faster than fine-tuning depending on model and data size (Huang et al., 2021).

Whereas TransRate estimates utility, “Parameter Transfer Unit for Deep Neural Networks” (Zhang et al., 2018) learns it as a continuous control variable inside the network. The Parameter Transfer Unit replaces discrete transfer states such as random initialization, frozen sharing, and fine-tuning with two learned gates: $$\mathbf{r}_{l} = \sigma (\mathbf{W}^{r}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]), \qquad \mathbf{z}_{l} = \sigma (\mathbf{W}^{z}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]).$$ The fine-tune gate $\mathbf{r}_l$ decides whether source activations should be copied directly or transformed, and the update gate $\mathbf{z}_l$ decides how much source knowledge should replace target activations. In CNN experiments, PTU reaches 0.5612 on CIFAR-10 $\rightarrow$ CIFAR-100 and 0.7686 on ILSVRC-2012 $\rightarrow$ Caltech-256 with MobileNetV1, outperforming the reported fine-tuning baselines in two of three scenarios; in RNN transfer from MNIST to Omniglot alphabets, PTU improves over fine-tuning on all five target alphabets, with relative gains from 6.86% to 35.43% (Zhang et al., 2018).

A still more operational form appears in “Fast and Accurate Transferability Measurement for Heterogeneous Multivariate Data” (Park et al., 2019). Transmeter estimates the usefulness of a source dataset/model pair for a heterogeneous target task by jointly learning a target encoder, target decoder, reused source label predictor, and adversarial domain classifier. Ground-truth transferability is defined as

$\frac{acc_T - acc_0}{acc_0}\times 100\ (\%),$

but the method ranks sources without full transfer by scoring source-conditioned alignment models. On ten heterogeneous datasets, Transmeter attains top-2 identification accuracy of 8/10 versus 5/10 for HeMap-t, and is reported to be up to 10.3 times faster than its competitor; using Transmeter to shortlist sources before full transfer is up to 3.26 times faster than exhaustive HeMap (Park et al., 2019). This suggests a mature distinction between transfer utility itself and the meta-problem of cheaply estimating it.

4. Utility of transfer in systems, privacy, communications, and physical transport

Outside model transfer, utility is often attached to transferred content or transferred state. “SkyHOST: A Unified Architecture for Cross-Cloud Hybrid Object and Stream Transfer” (Tariq et al., 20 Mar 2026) presents a “transfer utility” in the systems sense: a unified control plane over bulk object movement and stream replication. The system uses URI-based routing to select object or stream operators, supports object-to-stream and stream-to-stream modes, and reports 76–123 MB/s for Kafka-to-Kafka replication versus 58–159 MB/s for Confluent Kafka Replicator under comparable settings; for S3-to-Kafka, the best measured throughput in the conclusion is 131.6 MB/s for 96 MB chunks (Tariq et al., 20 Mar 2026). Here, utility is operational simplicity plus competitive throughput under heterogeneous transfer patterns.

In privacy-preserving data transfer, utility is defined as retained analytic value after masking. “Optimization of Privacy-Utility Trade-offs under Informational Self-determination” (Asikis et al., 2017) studies distributed IoT data transferred to a centralized consumer under local masking. Utility is based on the global aggregation error

$C_0$0

and the utility score combines normalized mean, standard deviation, and entropy of this error. On the Electricity Customer Behavior Trial dataset with 6,435 users, 536 days, and $C_0$1 sensor values, the framework evaluates more than 20,000 privacy settings and constructs privacy-utility trajectories under homogeneous and heterogeneous sharing (Asikis et al., 2017). The paper’s theorem shows that heterogeneous masking can still preserve aggregate approximability when masking and aggregation are compatible, so transfer utility can remain high even under user-specific privacy choices.

In wireless semantic communications, “Utility-Oriented Wireless Communications for 6G Networks” (Wang et al., 2023) defines a semantic transfer utility

$C_0$2

where $C_0$3 is the semantic bit number for user $C_0$4. The proposed objective jointly minimizes latency and power while maximizing semantic utility: $C_0$5 with $C_0$6 (Wang et al., 2023). This is a particularly explicit use of the term: transfer utility is the value of conveyed meaning relative to communication cost.

A physically operational sense appears in “Improved transfer efficiency with pulsed atom transfer between two magneto-optical traps” (Ram et al., 2010). In a double-MOT system for $C_0$7, pulsed VC-MOT loading followed by a pulsed push beam yields a UHV-MOT atom number about three times that obtained with continuous VC-MOT loading and a CW push beam of optimized power. The optimized focused CW baseline reaches $C_0$8 atoms in the UHV-MOT, while the pulsed scheme is reported as $C_0$9-times larger (Ram et al., 2010). In this setting, transfer utility is effectively transfer efficiency.

5. Transfer principles in utility theory and intertemporal welfare

In economic theory, “transfer utility” can mean transfer of formal properties between utility representations. “On expected and von Neumann-Morgenstern utility functions” (Pireddu, 2010) studies

$g$0

and asks which classical assumptions imposed on expected utility $g$1 are equivalent to or implied by corresponding assumptions on the primitive Bernoulli utility $g$2. Under lower unboundedness of $g$3, Theorem 1 states that $g$4 satisfies the four properties (2)–(5) if and only if $g$5 satisfies the analogous properties (6)–(9): $g$6-regularity, strict positivity of marginal utilities, strict concavity, and closed upper contour sets (Pireddu, 2010). The paper also proves that a weaker tangent-space second-order condition transfers from $g$7 to $g$8, but not generally from $g$9 to $$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$0.

A different line concerns transfer principles over distributions of utility. “Equitable preference relations on infinite utility streams” (Dubey et al., 2020) extends strong equity and Pigou-Dalton transfer principles to infinite utility streams $$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$1. It defines generalized equity (GE), generalized Pigou-Dalton (GPD), infinite equity (IE), and weak equity (WE), allowing potentially infinitely many paired redistributions. The paper proves sharp existence and representation restrictions. Theorem 2 states that there exists a social welfare function satisfying GE and monotonicity if and only if $$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$2 contains at most 5 elements; Theorem 3 gives the corresponding threshold 7 when monotonicity is dropped; and Theorem 1 states that there does not exist any social welfare function satisfying infinite equity and anonymity on $$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$3 if $$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$4 (Dubey et al., 2020). These results concern utility transfers in welfare evaluation, not transferable utility in the matching sense.

This distinction matters. The transfer principle literature studies inequality-reducing redistributions of utility across dates or persons, whereas the expected-utility paper studies transfer of mathematical properties across representations. Both use the vocabulary of transfer, but neither concerns matching-market TU.

6. Transferable utility in matching and systemic risk transfer

The most canonical economic use of the term is transferable utility in matching and equilibrium. “Transferable Utility Matching Beyond Logit: Computation and Estimation with General Heterogeneity” (Galichon et al., 28 Nov 2025) studies one-to-one bipartite matching with separable surplus

$$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$5

where $$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$6 is systematic surplus and $$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$7 are idiosyncratic tastes. The paper removes the i.i.d. type-I extreme value assumption of Choo–Siow and allows arbitrary heterogeneity distributions $$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$8, including correlated tastes. The matching problem is reformulated as a linear program with a dual transfer matrix $$\mathrm{TrR}_{T_s \rightarrow T_t}(g,\epsilon) = R(\hat{Z},\epsilon) - R(\hat{Z},\epsilon|Y).$$9, and solved by the Repeated Restricted Optimal Assignment algorithm, a column-generation procedure with finite termination. In simulations, RROA achieves speedups up to 25x over Gurobi on larger problems; in estimation experiments, normalized root mean square error falls below 1% under correct specification and stabilizes around 7% under Gumbel misspecification (Galichon et al., 28 Nov 2025). Here, transferable utility means that match surplus can be split additively between partners and cleared through endogenous transfers.

A systemic extension appears in “Multivariate Systemic Optimal Risk Transfer Equilibrium” (Doldi et al., 2019). That paper generalizes the original SORTE framework from additive utilities $$\mathbf{r}_{l} = \sigma (\mathbf{W}^{r}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]), \qquad \mathbf{z}_{l} = \sigma (\mathbf{W}^{z}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]).$$0 to a genuinely multivariate utility

$$\mathbf{r}_{l} = \sigma (\mathbf{W}^{r}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]), \qquad \mathbf{z}_{l} = \sigma (\mathbf{W}^{z}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]).$$1

so that an agent’s valuation can depend on the whole vector of allocations. An mSORTE is a triple $$\mathbf{r}_{l} = \sigma (\mathbf{W}^{r}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]), \qquad \mathbf{z}_{l} = \sigma (\mathbf{W}^{z}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]).$$2 combining deterministic capital allocation $$\mathbf{r}_{l} = \sigma (\mathbf{W}^{r}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]), \qquad \mathbf{z}_{l} = \sigma (\mathbf{W}^{z}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]).$$3, zero-sum terminal risk exchange $$\mathbf{r}_{l} = \sigma (\mathbf{W}^{r}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]), \qquad \mathbf{z}_{l} = \sigma (\mathbf{W}^{z}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]).$$4, and a pricing vector $$\mathbf{r}_{l} = \sigma (\mathbf{W}^{r}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]), \qquad \mathbf{z}_{l} = \sigma (\mathbf{W}^{z}_{l} [\mathbf{h}^{\mathcal{S}_{l}}, \mathbf{h}^{\mathcal{T}_{l}}]).$$5. The paper proves existence, uniqueness, and a Nash Equilibrium property of the optimizer (Doldi et al., 2019). Transfer utility in this setting is neither a feature-reuse score nor a welfare transfer principle; it is an equilibrium structure for allocating and pricing risk when total surplus is jointly determined.

Taken together, these papers show that transferable utility retains a precise technical meaning in economics: a structure in which surplus or risk can be redistributed through transfers while preserving equilibrium feasibility. That sense is conceptually distinct from machine-learning transferability, privacy-utility trade-offs, and semantic transfer, even though all belong to the broader family of transfer utility concepts.