Introduction to Course, Multitask Learning, and Meta Learning

Course Introduction

Noting the return to near-normalcy but acknowledging ongoing global uncertainties.
Policies have been set to provide flexibility recognizing students' diverse circumstances.

Course Website: Primary source of information.
Communication: Use Ed for questions, linked to Canvas. Preference for public posts on Ed to benefit all, but private posts and emails are allowed for confidential matters.
Office Hours: Information available on course website; Zoom links on Canvas. Start on Wednesday.

Foundations: Modern deep learning methods for multitask and learning across tasks.
Practical Experience: Implementing methods in code (Pytorch) and understanding real-world applications beyond theory.
Scientific Process: Insights into developing algorithms, encouraging critical thinking and understanding development processes.

Topics:
- Basics of multitask learning and transfer learning.
- Meta learning algorithms: Black Box approaches, optimization-based approaches, and Metric learning.
- Advanced topics: Overfitting, unsupervised and Bayesian meta learning, few shot learning, adaptation, unsupervised pre-training, domain adaptation, and domain generalization.
Case Studies: Real applications, e.g., recommendation systems (YouTube), land cover classification, education, few shot learning in large language models.

Removal of reinforcement learning topics.
New content: few shot learning with unsupervised pre-training, domain adaptation, etc.
Introduction of a new deep reinforcement learning course in the Spring.

Lectures: In-person, live-streamed, and recorded (available on Canvas).
Guest Lectures: Scheduled, details to be announced.
Participation Encouraged: Questions during lectures help gauge understanding.
Office Hours: Mix of in-person and remote; specifics on course website.
Prerequisites: Background in machine learning, familiarity with Pytorch.
Pytorch Review: Scheduled session for review.
Assignments: Range from Pytorch warm-up to advanced homework on meta learning and fine-tuning pre-trained models.
Grading: 50% homework, 50% project. Flexibility offered with an optional homework to replace a lower score or part of the project grade.
Late Days: 6 late days allowed across assignments, with up to 2 late days per assignment.
Collaboration Policy: Allowed but must acknowledge collaborators and independently write solutions and avoid external solutions.

Type: Research-level project in groups of 1-3, ideally related to students' ongoing research.
Flexibility: Projects can overlap with other courses but with higher expectations.
Poster Session: Final presentation of projects, no late days for this.

Chelsea's Research Perspective
- Goal: Enabling real-world agents (e.g., robots) to learn diverse skills using few examples.
- Examples: Robots using tools, mimicking human tasks, understanding objects and environments.
- Challenges: Current systems learn narrow tasks requiring extensive human effort and supervision. Aim to build more generalizable systems.
General vs. Specialist Systems
- Current systems are specialists, trained on single tasks with significant data and effort required for each new task.
- Humans and generalizable systems learn broadly, applying learned skills across domains.

Beyond Robotics: Applicable in general-purpose machine learning and various fields such as personalized education, rare language translation, medical imaging, etc.
Addressing Long-tail Distributions: Handling rare cases and edge cases more effectively by leveraging shared structures and prior data.
Quick Learning: Few-shot learning enables rapid adaptation to new tasks or environments using minimal data.
Applications: Systems that can handle a broad array of inputs, like vision and language tasks, providing diverse functionalities.

Determining shared structure between tasks and understanding dependencies and generalization capabilities in multitask systems.

Task Definition: A machine learning task defined by a dataset and a loss function to produce a model (formal definitions in future lectures).
Multitask Learning: Learning multiple tasks simultaneously, with training and testing on the same set of tasks.
Meta Learning/Transfer Learning: Using data from previous tasks to learn new tasks more effectively.

Multitask learning can sometimes be reduced to single-task learning by combining datasets and loss functions, but there are unique challenges and benefits.
Emphasized that studying multitask and meta learning can lead to significant advancements in performance and applicability of machine learning systems.

Engaged students with Q&A throughout the lecture, addressing concepts, applications, and theoretical underpinnings.

Note: Continue with office hours for more detailed personalization and guidance on projects and course material.