Transfer and Multitask Learning: How to Know Whether Sharing Helps

This guide, revised on July 30, 2026, begins with a comparison that many claims about artificial general intelligence (AGI) omit. If a model trained on one task learns another from 100 examples, does it outperform the same model initialized randomly with those 100 examples? And does it retain the first skill? Without those baselines, “transfers knowledge” may mean only that parameters were reused.

Transfer, multitask learning, and meta-learning all try to exploit relationships among problems, but they are not synonyms and do not guarantee improvement. Sharing can help, have no effect, or cause harm. The reader’s useful skill is to build a comparison matrix that exposes positive transfer, negative transfer, and adaptation cost.

Three settings that should remain separate

In sequential transfer, a model first learns a source task or domain and is then adapted to a target. Parts of the network may be frozen, every parameter may be fine-tuned, or modules may be added. The question is how much the target improves over learning without that source under the same budget. It also matters whether the source skill remains or adaptation degrades it.

In multitask learning, several losses are optimized during the same training process. A common architecture shares a trunk and keeps one head per task. It is not “parallelization” by definition; the essential feature is that updates from multiple tasks affect common parameters. Rich Caruana’s foundational multitask work described tasks learned in parallel through a shared representation that can let one task help another.

In meta-learning, the unit of training is a distribution of tasks. MAML learns an initialization designed so that a few updates with a small amount of data from a new task yield good performance. This demonstrates rapid adaptation within the evaluated family, not an ability to learn any task or an automatic step toward AGI.

The minimum baseline: four models, not one

Establishing transfer requires at least four conditions. The first trains the target from scratch. The second adapts the source model to the target. The third jointly trains source and target. The fourth preserves a source-only model so forgetting can be measured after adaptation. Capacity, available data, optimization steps, and hyperparameter search should be compared fairly.

Positive transfer occurs when adaptation or multitask training improves the target metric, reduces data needs, or accelerates convergence without an unacceptable source cost. Negative transfer occurs when sharing leaves the target worse than its independent baseline. An average across tasks can conceal it: if one task gains ten points and another loses eight, a positive mean does not mean both benefited.

Plot performance against number of examples and compute, rather than reporting only the final result. A source may help substantially with little data and lose its advantage as the target dataset grows; another may speed the beginning but converge to a lower ceiling. “Needs less data” is defensible only when the curve uses the same examples, splits, and selection process.

Relationships among tasks are measured, not assumed

Two tasks may share inputs and still require opposite invariances. Identity recognition may need to ignore illumination; estimating illumination must preserve it. Similar labels are not enough. Taskonomy experimentally studied transfer dependencies among 26 visual tasks and built a computational map of their relationships. The methodological lesson matters more than any particular edge: affinity is estimated through source-target experiments.

Some measures try to anticipate whether a representation will transfer before full fine-tuning. LEEP computes a score from a source model’s predictions on labeled target data and tests it against downstream performance. It is a tool for selecting candidates under the studied conditions, not a universal replacement for training and validation. A high score also says nothing definitive about future distribution shifts.

Interference: when tasks pull in different directions

In a shared network, each loss produces a gradient over common parameters. If two gradients point in compatible directions, one update may help both. If their dot product is negative, improving one locally may hurt the other. Magnitude matters, too: a task with a larger loss scale or greater difficulty can dominate even when it is not the priority.

GradNorm dynamically changes weights to balance gradient magnitudes and training rates across tasks. PCGrad projects a gradient when it conflicts with another task’s gradient. Both works exist because dropout and batch normalization do not solve multitask balance by themselves. Regularization may reduce overfitting, but it does not decide which objective should yield.

A simple audit records each task’s loss, metric, gradient norm, and gradient cosine over training. The diagnosis does not treat every negative cosine as failure; optimization may pass through useful conflicts. It does locate when a task stops progressing after another is introduced and tests whether changing weights, sampling frequency, or shared parameters repairs the damage.

Loss weights are a product decision

When total loss is a weighted sum, the weights express a trade-off. Equal numbers do not mean equal importance because metrics, noise, and learning speeds differ. A system predicting depth and segmentation may accept a trade-off that would be intolerable between fraud detection and false account blocks.

“Multi-Task Learning as Multi-Objective Optimization” frames training as several objectives and seeks a Pareto solution: one objective cannot improve without worsening at least one other. The idea supplies a durable question for any result: which trade-off points were explored, and who selected the reported one? A single aggregate score can hide alternatives better suited to actual use.

What to share: trunk, layers, modules, or nothing

Hard sharing uses the same parameters for part of the network and separate heads at the end. It saves memory and forces a common representation, but may impose similarity where none exists. Soft sharing keeps task-specific parameters and penalizes their distance or exchanges information through modules. It costs more but allows specialization.

The comparison should include independent models with comparable total capacity. If the multitask model has many more parameters or sees more examples, improvement cannot be attributed solely to shared knowledge. Ablations also matter: share up to different layers, remove an auxiliary task, and vary its data proportion. An auxiliary task never used in production may still improve the primary goal, but its contribution must be shown against the same architecture without it.

Fast adaptation is not open-ended transfer

MAML and other meta-learned methods divide tasks into meta-training and meta-test sets. That split is critical. If the two sides change only labels or superficial parameters within the same template, the result demonstrates adaptation inside that distribution. A broader claim requires holding out relevant families, domains, or mechanisms and reporting how many examples and update steps adaptation uses.

A model can also learn the format of the exam. Baselines without meta-training, out-of-distribution tasks, and sensitivity to the selected support examples are therefore necessary. If the benefit disappears when the few examples are presented differently, the capability may depend more on the protocol than on a transferable rule.

The matrix that reveals help and harm

For N tasks, rows can represent a source or shared task set and columns each target. Every cell records from-scratch performance, transferred performance, difference, examples, compute, and retained source performance. The diagonal contains independent training; off-diagonal cells reveal direction. Transfer from A to B does not imply that B helps A.

Repeat the matrix under domain shifts and across several seeds. One average is insufficient when source choice, task order, or initialization changes the conclusion. Storage, latency, and maintenance cost should also appear: a slightly weaker multitask model may be worthwhile if it replaces several systems, but that is an explicit operational choice, not a precision victory.

The transferable skill is to demand a difference against an independent baseline for every task, together with data curves, retention, and out-of-domain tests. Parameter sharing is a mechanism; positive transfer is a result. Only when the matrix shows repeated help without hidden harm can we speak of reusable knowledge—and even then, not of general intelligence by decree.