Multi-Task Learning Basics | Stanford CS330 Lecture 2

Added:

Gradient Interference and the Modular Alternative to Multi-Task Learning

While Multi-Task Learning (MTL) aims to improve generalization through shared representations, it frequently suffers from "negative transfer" or "gradient interference." This occurs when the optimization paths of different tasks conflict, causing the joint model to perform worse on individual tasks than dedicated, single-task models. Critics argue that balancing task objectives and loss weights in MTL is notoriously difficult and computationally expensive. Consequently, many researchers advocate for alternative paradigms like self-supervised pre-training followed by task-specific fine-tuning, or modular architectures such as Mixture of Experts (MoE) and Parameter-Efficient Fine-Tuning (PEFT). These alternatives isolate task-specific parameters to prevent interference while still leveraging shared foundational knowledge, challenging the traditional MTL approach of joint, dense parameter sharing.

Fundamentals of supervised machine learning, including empirical risk minimization, loss functions, and evaluation metrics.
Basic deep learning architectures such as Multi-Layer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs).
Standard optimization algorithms in deep learning, particularly gradient descent and its variants like Adam.
Familiarity with multi-variable calculus and linear algebra for understanding joint gradient updates and parameter sharing.
Advanced optimization techniques in multi-task learning to handle conflicting gradients and negative transfer, such as PCGrad or GradNorm.
Meta-learning paradigms (learning-to-learn), including optimization-based approaches like MAML (Model-Agnostic Meta-Learning).
Transfer learning, domain adaptation, and domain generalization strategies to apply learned representations to novel distributions.
Few-shot and zero-shot learning frameworks, which leverage multi-task representations to adapt to new tasks with extremely limited data.
28.8K views353likes1:17:37@stanfordonlineOriginal Release: 2023-03-31

Multi-task learning involves training a single neural network to perform multiple related tasks simultaneously by conditioning on a task descriptor (such as one-hot encoding or natural language description) and sharing parameters across tasks through various architectures (concatenation, additive, multiplicative conditioning, or multi-head structures), with key design choices including how much to share parameters, how to weight different tasks in the loss function, and whether to use vanilla optimization or advanced techniques like Minimax to address challenges such as negative transfer or overfitting.