Principle:Snorkel team Snorkel Multitask Classifier Construction

Knowledge Sources	An Overview of Multi-Task Learning in Deep Neural Networks Multi-Task Learning as Multi-Objective Optimization
Domains	Multi_Task_Learning, Deep_Learning, Model_Architecture
Last Updated	2026-02-14 20:00 GMT

Overview

An architecture pattern for constructing a single neural network that serves multiple tasks through shared representations and task-specific heads.

Description

Multitask Classifier Construction assembles a unified model from multiple Task definitions. The model:

• Merges module pools from all tasks into a single shared pool
• Maintains separate operation sequences per task
• Shares parameters between tasks that reference the same module
• Routes each tasks data through its specific operation sequence during forward pass

This enables hard parameter sharing (shared lower layers) and soft sharing (task-specific heads) in a single model, which typically improves generalization through inductive bias from related tasks.

Usage

Use this principle when you have multiple related classification tasks that could benefit from shared representations. Construct the classifier after defining all tasks.

Theoretical Basis

A multi-task classifier with $T$ tasks shares parameters ${\theta }_{\text{shared}}$ while maintaining task-specific parameters ${\theta }_t$:

${\hat{y}}_t=g_t(f(x;{\theta }_{\text{shared}});{\theta }_t)$

where $f$ is the shared feature extractor and $g_t$ is the task-specific head. The combined model parameters are $\theta ={\theta }_{\text{shared}}\cup \bigcup _t{\theta }_t$.

Related Pages

Implemented By

• Implementation:Snorkel_team_Snorkel_MultitaskClassifier_Init