Pareto Multi-Task Learning

• Pareto Multi-Task Learning is a framework that formulates multi-task problems as multi-objective optimization to capture all optimal trade-offs among tasks.
• It employs advanced gradient-based techniques and decomposition strategies to overcome scalarization limitations and traverse nonconvex Pareto fronts.
• Recent advances include continuous parameterizations and user-controlled trade-offs that enhance predictive performance and robust front coverage.

Pareto Multi-Task Learning (Pareto MTL) encompasses a principled set of strategies for training multi-task models under explicitly multi-objective optimization criteria, targeting the set of Pareto optimal solutions that formalize all optimal trade-offs among tasks. Rather than relying on a single scalarized objective, Pareto MTL methods aim either to compute an extensive and well-distributed set of Pareto solutions across the Pareto front or to provide parametric mappings that deliver user-controllable trade-offs at inference time. These approaches are motivated by the inherent conflicts among task objectives in joint learning settings and have produced a distinct methodology underpinning modern algorithmic advances in deep multi-task learning and multi-criteria decision-making.

1. Multi-Objective Formulation in Multi-Task Learning

Pareto MTL is founded on the explicit formulation of MTL as a multi-objective minimization problem. Given $K$ differentiable (potentially non-convex) task losses $f_i(\theta)$, $i=1,\ldots,K$ parameterized by shared variables $\theta \in \mathbb{R}^n$, the vector objective is $$F(\theta) = (f_1(\theta),\ldots, f_K(\theta)) \in \mathbb{R}^K$$. This system is governed by Pareto dominance: $\theta$ dominates $\theta'$ if $f_i(\theta) \le f_i(\theta')$ for all $i$ and strictly for some $j$. The set of non-dominated points forms the local Pareto set $f_i(\theta)$0, with the Pareto front given by $f_i(\theta)$1 in objective space. This framework yields solutions representing all attainable trade-offs such that no task loss can be further improved without sacrificing another (Mahapatra et al., 2021, Lin et al., 2019).

2. Scalarization, Pareto Optimality, and Limitations

Classical scalarization, such as the weighted sum or Chebyshev (min-max) scalarization, reduces the multi-objective problem to a single-objective form. For positive user-specified priorities $f_i(\theta)$2 ($f_i(\theta)$3), linear scalarization forms $f_i(\theta)$4, and Chebyshev scalarization adopts $f_i(\theta)$5. A minimizer $f_i(\theta)$6 of $f_i(\theta)$7 is an Exact Pareto Optimal (EPO) solution for $f_i(\theta)$8, achieving losses proportional to the inverse of the user priorities (Mahapatra et al., 2021).

However, scalarization is theoretically incapable of fully tracing nonconvex or intersection regions of the Pareto front, especially in under-parameterized or non-convex settings, as it cannot represent Pareto optimal points at the intersection of distinct surface components with "gradient disagreement" (Hu et al., 2023). This necessitates the development of specialized multi-objective optimizers or decomposition-based strategies that directly target Pareto criticality (Lin et al., 2019, Sener et al., 2018).

3. Core Algorithms for Pareto MTL

3.1 Exact Pareto Optimal (EPO) Search

EPO Search (Mahapatra et al., 2021) alternates between "balance" and "descent" anchor modes to guarantee both convergence and explicit path-tracing along the Pareto front. In each step, a small-dimensional quadratic program in $f_i(\theta)$9 yields a search direction $i=1,\ldots,K$0, which is used to update the parameters. Balance mode equalizes relative losses to approach the user-specified priority ray, while descent mode reduces all objectives after proximity to the ray is achieved. The algorithm ensures (i) robust avoidance of oscillation and stagnation characteristic of simple Chebyshev-SGD and (ii) convergence guarantees (including global linear convergence under mild conditions). Its computational cost per iteration is $i=1,\ldots,K$1 and remains practical for deep models.

3.2 Constrained Pareto MTL via Subproblem Decomposition

The Pareto MTL framework (Lin et al., 2019) decomposes the original vector optimization into $i=1,\ldots,K$2 constrained subproblems, each associated with a preference vector $i=1,\ldots,K$3, representing a region of the front. For each, a restricted steepest-descent subproblem is solved under linear constraints in objective space, using a dual quadratic program with dimension $i=1,\ldots,K$4. All $i=1,\ldots,K$5 subroutines can be run in parallel.

3.3 Specialized Multi-Task Optimizers (SMTOs)

Gradient-based multi-objective optimizers such as the Multiple Gradient Descent Algorithm (MGDA) (Sener et al., 2018) and its efficient upper-bound variant (MGDA-UB) directly compute a convex combination $i=1,\ldots,K$6 of per-task gradients so that $i=1,\ldots,K$7 is minimized. These approaches provably reach Pareto-stationary points and, unlike scalarization, can traverse nonconvex regions of the front. MGDA-UB leverages shared encoder structure to reduce complexity and remains practical for high task counts.

3.4 Continuous Pareto Manifold Parameterizations

Recent advances provide continuous mappings from preference vectors $i=1,\ldots,K$8 in the simplex to model parameters, either in weight space or through low-rank adapter augmentations:

• Pareto Manifold Learning (PML) (Dimitriadis et al., 2022): Maintains $i=1,\ldots,K$9 anchor networks $\theta \in \mathbb{R}^n$0, parameterizing the Pareto front as $\theta \in \mathbb{R}^n$1.
• Efficient Low-Rank Manifolds (Chen et al., 2024, Dimitriadis et al., 2024): Augment a main network with $\theta \in \mathbb{R}^n$2 low-rank adapter pairs $\theta \in \mathbb{R}^n$3, yielding parameterizations $\theta \in \mathbb{R}^n$4. Orthogonal regularization and deterministic preference scheduling further enhance front coverage, scalability, and monotonicity.

These approaches support real-time, continuous inference control, addressing the needs of practical applications with many tasks or user-specific trade-offs.

4. Extensions for Multi-Criteria Decision-Making and Preference Elicitation

Pareto MTL methods such as EPO Search are extended to support multi-criteria decision-making paradigms:

• PESA-EPO: Employs the Pattern Efficient Set Algorithm to generate diverse preference vectors and warm-starts each from a previous Pareto point, tracing the front through contiguous arcs for a posteriori analysis. Empirically, this yields higher-quality Pareto front approximations with fewer restarts than population-based evolutionary algorithms (Mahapatra et al., 2021).
• GP-EPO: For interactive preference elicitation, a Gaussian Process models utility along inverse preference rays. Each EPO Search run returns the exact solution for a suggested preference, reducing both the number of queries and overall regret compared to GP methods relying on discretization.

5. Scalability, Empirical Performance, and Applications

Pareto MTL algorithms have been evaluated across personalized medicine, e-commerce, scene understanding, and hydrometeorology. They demonstrate:

• More uniform compliance with user-specified priorities (i.e., task losses proportional to $\theta \in \mathbb{R}^n$5).
• Strictly higher predictive performance or lower regret than scalarization or earlier gradient-manipulation techniques.
• Denser and more expressive Pareto front coverage, with the ability to expand or interpolate the solution space (e.g., via low-rank factor adjustment or adapter fine-tuning).
• Robust computational scaling and convergence properties, with empirical per-iteration costs linear in the model parameter count.

Notably, connection strength-based approaches (Jeong et al., 2024) leverage task-specific connection strengths and task priority quantification to strictly expand the attainable Pareto frontier beyond what previous gradient-projection methods achieve.

6. Theoretical Guarantees and Limitations

Pareto MTL frameworks provide provable convergence (to EPO points) under standard smoothness and multi-objective regularity conditions (Mahapatra et al., 2021) and theoretical separation from scalarization approaches, which are unable to cover intersection regions on the front in nonconvex or under-parameterized models (Hu et al., 2023). Continuous parameterization methodologies, such as Pareto manifold learning, achieve universal approximation on the Pareto front under mild continuity assumptions (Chen et al., 2024).

Limitations include the potential for coverage challenges in high-dimensional task spaces, the need for careful selection or learning of preference vectors and adapter ranks, and, in certain algorithmic realizations, reliance on specific architectural structures (e.g., shared encoders or explicit modularization).

7. Summary Table: Distinctive Pareto MTL Methodologies

Approach	Core Mechanism	Front Parametricity	Reference
EPO Search	Chebyshev scalarization, anchor-QP alternation	Discrete	(Mahapatra et al., 2021)
(Restricted) PMTL	Constrained subproblem decomposition	Discrete/multi-run	(Lin et al., 2019)
MGDA/MGDA-UB	Gradient simplex QP	Discrete	(Sener et al., 2018)
Pareto Manifold	Linear convex hull of anchors	Continuous	(Dimitriadis et al., 2022)
Low-rank Adapter PF	Backbone + parametric adapters	Continuous	(Dimitriadis et al., 2024, Chen et al., 2024)
Task-priority methods	Connection strength/priority-based scheduling	Discrete/expansion	(Jeong et al., 2024)

Each method addresses specific challenges of front coverage, computational tractability, and user-controllable trade-off specification. The field continues to evolve toward scalable, flexible, and theoretically sound mechanisms for sampling, covering, and exploiting the Pareto frontier in high-dimensional multi-task settings.