Principle:Snorkel team Snorkel Task Operation Definition

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

Overview

An abstraction for defining task computation graphs as sequences of named operations over shared module pools.

Description

Task and Operation Definition provides a declarative way to define how data flows through neural network modules for each task in a multi-task model. Instead of hardcoding forward passes, each task is defined as:

Module Pool: A dictionary of named nn.Module instances (layers, embeddings, etc.)
Operation Sequence: An ordered list of Operations, each specifying which module to use and where its inputs come from
Loss Function: How to compute the task-specific loss
Output Function: How to convert logits to predictions

This architecture enables module sharing across tasks: multiple tasks can reference the same module in the pool while having different operation sequences.

Usage

Use this principle when configuring the computation graph for a MultitaskClassifier. Define Operations to specify data flow and Tasks to bundle them with loss functions and scorers.

Theoretical Basis

A computation graph for task $t$ is defined as a DAG of operations:

$h_{0} = x, h_{i} = f_{m_{i}} (h_{inputs (i)})$

where $f_{m_{i}}$ is the module referenced by operation $i$ and $inputs (i)$ specifies the preceding operations whose outputs are fed in. Module sharing is achieved when $m_{i} = m_{j}$ for operations in different tasks.

Related Pages

Implemented By

Implementation:Snorkel_team_Snorkel_Task_Init

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment