CS 601.471/671: Self-supervised Models

Large self-supervised (pre-trained) models have transformed various data-driven fields such as natural language processing (NLP). In this course, students will gain a thorough introduction to self-supervised learning techniques for NLP applications. Through lectures, assignments, and a final project, students will learn the necessary skills to design, implement, and understand their own self-supervised neural network models, using the Pytorch framework.

Note: The course is different from 601.771 (offered in the fall semesters) which involves reading recent papers and is geared toward grad students that want to specialize in the latest developments in self-supervised models.

Prerequisites: (1) Data Structures (601.226), (2) All the class assignments will be in Python/PyTorch. If you don’t know Python or PyTorch but have experience in other programming languages (Java, C++, etc.) you can probably pick Python/PyTorch pretty quickly. (3) Calculus and linear algebra: you should be comfortable with matrix operations (matrix multiplication, transpose, inverse, dot product, gradients). (4) Probability: basic probability properties (conditionals, marginals, mean, standard deviation), distributions (gaussian, categorical). (5) Background in Natural Language Processing & Machine Learning or having finished one of the relevant courses such as Machine Learning (475.675), Artificial Intelligence (464.664), Natural Language Processing (600.465), Machine Translation (600.468), or Introduction to HLT (601.467/667).

Relevant Courses at Hopkins: This course has some overlap with "Natural Language Processing" (EN.601/665), "Introduction to Human Language Technology" (601.467/667), and "Artificial Agents" (EN.601.470/670), though the courses have different focuses.

Logistics

Classes: on Tuesday/Thursday 12 - 1:15 pm EST (room: Ames 234, or zoom meeting: 704 538 4305)
Office hours: (Daniel office hour) Tuesday and Thursday 1:30 - 2:30 pm EST (Hackerman hall, 316B), (TA office hour) Thursdays 11 - 11:50 am (in-person, Hackerman hall 3rd floor), and (CA office hour) Fridays 1 - 1:50 pm (in-person, Hackerman hall 3rd floor).
Contact: If you have any questions about the course, you can post them on Piazza.
Virtual or in-person: The class will be in-person. Though there will be recordings of each class made available online after each class on the Youtube playlist (shared on Piazza).
Changes: The instructor reserves the right to make changes to the syllabus or project due dates. These changes will be announced as early as possible.
News and announcements: All the news and announcements will be made on Piazza.
COVID: Students who report symptoms associated with COVID-19 are expected not to attend class and to isolate themselves for at least five days and until they have been symptom-free for 24 hours.
Course grade: Your grade is based on the following activities: (1) Semi-weekly assignments (50%), (2) midterm exam (20%), and (3) a final project (30%) -- same grade for all members of a team. Attendance is not mandatory (hence, 0%) but highly encouraged: participation in class is our chance to learn more effectively. Up to 3% additional credit for any actions taken to improve the course that is brought to the instructors' attention.

Key links

Piazza for discussion and announcements. Sign up, follow, ask questions, and participate in discussions!
Gradescope for submitting your assignments.

Assignments

The homework is your opportunity to practice doing the thing. The lectures and office hours hopefully provide good intuition and motivation and justification for the skills we want you to develop, but the best way to develop those skills is by trying to solve the problems yourself. The practice is far more important than the solution.

The course has ~12 weekly assignments which will improve both your theoretical understanding and your practical skills. All assignments contain both written questions and programming parts (mainly in Python). They will be released on this website, and submissions should be uploaded to Gradescope.

Here is a tentative list of topics for the assignments:

#	Focus
#1	Algebra, calculus, probability recap, implementing Skip-Gram model, classification, evaluation, comparison to basic features (unigrams, bigrams) and existing word embeddings.
#2	Understanding softmax function, classification via vector representations, playing with gradient descent.
#3	PyTorch introduction, automatic differentiation, computation graph, how to use PyTorch on GPUs, basic feedforward network and backpropagation, Word2vec as a feedforward net with automatic differentiation
#4	Neural language model with feedforward network, evaluating language modeling, count-based models, decoding language models
#5	Recurrent neural language model and evaluation; Transformers
#6	Fine-tuning LMs, prompting language models, fine-tuning them, distributed tuning.
#7	Prompt engineering, in-context learning; Retrieval-augmented language models

Submission: Assignments must be submitted via GradeScope. If you're not familiar with Gradescope, check out its documentation.
Collaboration: Study groups are allowed, but students must understand and complete their own assignments, and hand in one assignment per student. If you worked in a group, please put the names of the members of your study group at the top of your assignment. Please ask if you have any questions about the collaboration policy. Again, you must understand and complete your own assignments in your own words, and hand in one assignment per student.
Using Other Resources: We strongly encourage you to use any outside source at your disposal, provided you use your sources properly and give them proper credit. If you get an idea from an outside source, citing that source will not lower your grade. Failing to properly cite an outside source—thereby taking credit for ideas that are not your own—is plagiarism.
Appropriate Citations: You must write everything in your own words, and properly cite every outside source you use, including other students. Using ideas from other sources or people without citation is plagiarism. Copying other sources verbatim, even with proper citation, is plagiarism. Don't do that. The only sources that you are not required to cite are the official course materials (lectures, slides, and assignments).
Honor code: We expect students to not look at solutions or implementations online. We take the student Honor Code seriously. We sometimes use automated methods to detect overly similar assignment solutions.
Late days: Each student has 10 late days to use for assignments. A late day extends the deadline 24 hours. Once you have used all your late days, the penalty is 5% off the final homework grade for each additional late day. For example, if you're late for 3 days on an assignment (beyond your legitimate "late days" capacity, which is 7 days per homework and 10 days in total), you will lose 15% of the points for that assignment. The deadline cutoffs are at 12pm of each day. There are no fractional late days. If you're late for 1 hour, you lose a full day.
You can use up to 7 late days per assignment (so, if you're late on HW1, you can submit it until the release of HW2). Assignments submitted after 7 late days will not be graded (unless explicit permission is given in advance by the instructor).
Grading: Homeworks are graded by the entire course staff, directly within Gradescope. To keep grading consistent, each numbered problem is graded by a single grader, under the supervision of one of the TAs, using a detailed rubric developed within Gradescope. Under normal circumstances, all homework should be graded within 10 calendar days of submission.
Regrading: Regrade requests can be submitted directly within Gradescope and must include a brief written justification for the request. We encourage students who have questions or concerns about their grades to talk with the course staff before submitting a regrade request. However, no grades will be changed in any student's presence.

Midterm exam

There will be one in-class midterm. The midterm exam will be paper-based and during the usual class time. This midterm exam aims to evaluate students' progress and understanding of ideas presented in the first half of the semester, which will serve as a foundation for the material covered in the second half of the semester. The exam will assess students' mastery of the topics discussed in the lectures and weekly homework assignments. The exam will also provide feedback to both the student and the instructor, and identify areas that need improvement to inform further learning and teaching. The midterm will cover all material until the end of "Transformer Language Models", just before "Doing Things with Language Models". The week leading to the midterm, we will not have homework assignments.

Final project

The objective of the final project is to make use of what you have learned during this course to solve a hard problem.

The project deliverables are: (1) A project proposal, (2) a final report, (3) a final project poster summarizing the technical aspects of the project.

Topic: The topic of this project is open-ended. This project, for example, can focus on demonstrating systemic limitations of prior work or suggesting improvements on methods or benchmarks discussed in the class.
Group work: Students are encouraged to work in groups on the final project (team sizes limited to 2 or 3 people).
Project proposals: All groups will be required to submit a project proposal (due on the class calendar). The project proposal is a 2-page description of what you intend to do (experiments, datasets, methods, etc.) The instructor(s) will provide feedback on these ideas to help the teams with finding a concrete idea.
Midway progress reports: Reports elaborating discussing the progress made for each project and any remaining hurdles/goals.
Final poster presentations: All students will present their findings at a poster presentation during the final exam period.
Final report: Students should write code and carry out additional experiments and then write up the results in a standard conference paper format (template). References don't count toward the page limit. Note that longer reports are not necessarily better. Students in groups are required to include a “contributions” section concretely lists each author’s contributions (see Section 8 of this paper, for example). The final report should concisely summarize your findings and answer the following questions: 1. What approach did you take to address this problem, and why? 2. How did you explore the space of solutions? 3. How did you evaluate the performance of the approach(es) you investigated? 4. What worked, what did not work, and why?
Project grading: The goal of the project is to demonstrate the group’s understanding of the tools and challenges when using self-supervised models. Grading will reflect the quality of the approach, the rigor of evaluation, and reasoning about successes and failures. Grading will also depend on the completeness of the project, the clarity of the writeup, the level of complexity/difficulty of the approach, and your ability to justify the choices you made. Here is the grade breakdown for the projects:
- Project proposal write-up: 5%
- Quality of final report write-up, implementation and results: 70%
- Final poster and its presentation: 25%

Content Schedule

Each session will involve an instructor-led presentation on a focused topic self-supervised models. There will be weekly assignments related to class presentations, a midterm exam, and a final project.

The current class schedule is below (subject to change):

Date	Topic	Course Materials	Events	Deadlines
#1 - Tue Jan 24	Course overview Plan and expectations [slides: pdf, pptx]		HW1 released! [tex] [pdf] [colab]
⬇️ -- Self-supervised Word Representations
#2 - Thu Jan 26	Word meaning and representation [slides: pdf, pptx]	Suggested Reading: Jurafsky & Martin Chapter 6 Additional Reading: Python / Numpy Tutorial (with Jupyter and Colab) Learning Word Embeddings Dive into Deep Learning: Word Embeddings Dive into Deep Learning: Gradient Descent Efficient Estimation of Word Representations in Vector Space (Original W2V paper)
Fri Jan 27	TA Review Session (virtual over Zoom): Math background + Python [zoom link] [slides]	Time: 9 - 9:50 AM
#3 - Tue Jan 31	Word2vec objective function (continued), inspecting and evaluating word vectors [slides: pdf, pptx]	Suggested Reading: Jurafsky & Martin Chapter 6 Additional Reading: Optimization: Stochastic Gradient Descent Distributed Representations of Words and Phrases and their Compositionality (Negative Sampling paper)	HW2 released! [tex] [pdf] [colab]	HW1 due
⬇️ -- Self-Supervised Representation of Feedforward Neural Language Models
#4 - Thu Feb 2	Word2vec limitations and modeling context feedforward networks Neural nets: brief history Word2vec as simple feedforward net [slides: pdf, pptx]	Suggested Reading: Jurafsky & Martin Chapter 7 Additional Reading: Neural Networks: the Architecture Neural Networks: data and loss Dive into Deep Learning: Multilayer Perceptron Dive into Deep Learning: Practitioners Guide to Neural Networks Dive into Deep Learning: Linear Algebra in PyTorch
#5 - Tue Feb 7	Analytical backpropagation Automatic differentiation Practical tips for training neural networks [slides: pdf, pptx]	Suggested Reading: Jurafsky & Martin Chapter 7 Additional Reading: Neural Networks: Backpropagation Neural Networks: Training and empirical tips Computing Neural Network Gradients Learning representations by back-propagating errors (the original backpropagation paper) Deep Learning Tuning Playbook	HW3 released! [tex] [pdf] [colab]	HW2 due
#6 - Thu Feb 9	Language modeling, N-gram models, evaluating LMs [slides: pdf, pptx] [slides: pdf, pptx]	Suggested Reading: Jurafsky & Martin Chapter 3 Additional Reading: Revisiting Simple Neural Probabilistic Language Models Prediction and Entropy of Printed English (the foundational paper by Shannon on language compression and uncertainty) Google N-grams (very insightful trends over time)
Fri Feb 10	TA Review Session (virtual over Zoom): Backpropagation and PyTorch	Time: 9 - 9:50 AM
#7 - Tue Feb 14	Measuring LM quality Fixed-window language modeling with FFNs [slides: pdf, pptx]	Suggested Reading: Jurafsky & Martin Chapter 7 Additional Reading: Prediction and Entropy of Printed English (the foundational paper by Shannon on language compression and uncertainty) Revisiting Simple Neural Probabilistic Language Models	HW4 released! [tex] [pdf] [colab]	HW3 due
⬇️ -- Self-Supervised Representation of Recurrent Neural Language Models
#8 - Thu Feb 16	Text generation algorithms Recurrent neural networks Encoder-decoder models [slides: pdf, pptx]	Suggested Reading: CS231N course notes on RNNs Additional Reading: Dive into Deep Learning: Recursive Neural Networks The Unreasonable Effectiveness of Recurrent Neural Networks (blog post overview) Learning long-term dependencies with gradient descent is difficult (one of the original vanishing gradient papers) The Curious Case of Neural Text Degeneration
#9 - Tue Feb 21	RNN continued: ELMo Language units and subwords [slides: pdf, pptx]	Suggested Reading: Deep contextualized word representations (ELMo paper) Additional Reading: Huggingface tutorials on Tokenization	HW5 released! [tex] [pdf] [colab]	HW4 due
⬇️ -- Self-Supervised Representation of Transformer Language Models
#10 - Thu Feb 23	Self-attention Transformer [slides: pdf, pptx]	Suggested Reading: Attention Is All You Need Additional Reading: Dive into Deep Learning: Attention Mechanism The Annotated Transformer The Illustrated Transformer
⬇️ -- Large Language Models
#11 - Tue Feb 28	Encoder family (BERT, RoBERTa, ...) [slides: pdf, pptx]	Suggested Reading: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding Additional Reading: The Illustrated BERT, ELMo, and co		HW5 due
#12 - Thu Mar 2	Encoder-Decoder family (T5, BART), Decoder family (GPTk) [slides: pdf, pptx]	Suggested Reading: Exploring the limits of transfer learning with a unified text-to-text transformer (T5 paper) Additional Reading: BART: Denoising Sequence-to-Sequence Pre-training Language Models are Unsupervised Multitask Learners The Illustrated GPT-2
#13 - Tue Mar 7	Midterm exam		HW6 released! [tex] [pdf]
#14 - Thu Mar 9	Final projects [slides] Decoder family (GPTk) In-context learning [slides: pdf, pptx]	Suggested Reading: Language Models are Few-Shot Learners (GPT3 paper) Additional Reading: OPT: Open Pre-trained Transformer Language Models
Mar 9-17	Project teaming extravaganza
#15 - Tue Mar 14	In-context learning Adapting models with prompting (prompt engineering) Failure modes of in-context learning [slides: pdf, pptx]	Suggested Reading: Calibrate Before Use: Improving Few-Shot Performance of Language Models Additional Reading: Fantastically Ordered Prompts and Where to Find Them: Overcoming Few-Shot Prompt Order Sensitivity (LMs are sensitive to ICL example orders) Reframing Instructional Prompts to GPTk's Language (LMs are sensitive to phrasing of prompts) Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models	HW7 released! [tex] [pdf]	HW6 due
#16 - Thu Mar 16	Multi-step reasoning via prompts Adapting models with parameter change (head-tuning, prompt-tuning, adaptors) [slides: pdf, pptx]	Suggested Reading: The Power of Scale for Parameter-Efficient Prompt Tuning Additional Reading: Prefix-Tuning: Optimizing Continuous Prompts for Generation Prompt Waywardness: The Curious Case of Discretized Interpretation of Continuous Prompts
#17 - Tue Mar 21	No Class - Spring Break
#18 - Thu Mar 23	No Class - Spring Break
#19 - Tue Mar 28	Scaling laws Modifying self-attention for long context Retrieval augmented language models [slides: pdf, pptx]	Suggested Reading: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks Additional Reading: Scaling Laws for Neural Language Models REALM: Retrieval-Augmented Language Model Pre-Training Unsupervised Dense Information Retrieval with Contrastive Learning Improving language models by retrieving from trillions of tokens Atlas: Few-shot Learning with Retrieval Augmented Language Models An Efficient Memory-Augmented Transformer for Knowledge-Intensive NLP Tasks
#20 - Thu Mar 30	Social concerns about LMs: Bias, fairness and toxic language Hallucination, truthfulness, and veracity [slides: pdf, pptx]	Suggested Reading: Bias Out-of-the-Box: An Empirical Analysis of Intersectional Occupational Biases in Popular Generative Language Models Additional Reading: UnQovering Stereotyping Biases via Underspecified Questions Robots Enact Malignant Stereotypes Fewer Errors, but More Stereotypes? The Effect of Model Size on Gender Bias Red Teaming Language Models with Language Models RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models TruthfulQA: Measuring How Models Mimic Human Falsehoods Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus		HW7 due + Project proposals deadline
#21 - Tue Apr 4	Alignment via language instructions: existing solutions and challenges [slides: pdf, pptx]	Suggested Reading: Training language models to follow instructions with human feedback (GPT3 + RLHF paper) Additional Reading: Illustrating Reinforcement Learning from Human Feedback Learning to summarize from human feedback Deep reinforcement learning from human preferences (an early RLHF paper) Training language models to follow instructions with human feedback (GPT3 + RLHF paper) Scaling Instruction-Finetuned Language Models (FLAN paper) Generalization via Declarative Instructions on 1600+ NLP Tasks (Super-NaturalInstructions paper) Self-Instruct: Aligning Language Model with Self Generated Instructions (Self-Instruct paper)
#22 - Thu Apr 6	Vision-language models [slides: pdf, pptx]	Suggested Reading: Multimodal Few-Shot Learning with Frozen Language Models Additional Reading: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale Learning Transferable Visual Models From Natural Language Supervision Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision Hierarchical Text-Conditional Image Generation with CLIP Latents Emerging Properties in Self-Supervised Vision Transformers MERLOT: Multimodal Neural Script Knowledge Models Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding Image as a Foreign Language: BEiT Pretraining for All Vision and Vision-Language Tasks Discovering the Hidden Vocabulary of DALLE-2 High-Resolution Image Synthesis with Latent Diffusion Models
#23 - Tue Apr 11	Guest speaker: Jim Fan (NVIDIA)
#24 - Thu Apr 13	Guest speaker: Tim Dettmers (UW) [slides]	Title: 8-bit Methods for Efficient Deep Learning Abstract: Large language models are effective tools for many tasks but are difficult to train and inference due to their size. Moving from 32-bit models to 16-bit models resulted in considerable efficiency gains that made training and inference of large models easier. Can we train and inference in 8-bit to make further gains? In this talk, I will show that 8-bit inference and training can be used without degrading performance while improving efficiency. To make 8-bit methods work, it is essential to understand how quantization precision affects model performance and training stability as we scale the model size. I will talk about how these factors change with scale and how we need to adjust 8-bit methods to make them work. In particular, I will speak about 8-bit optimizers for training and Int8 inference for large language models with up to 175B parameters. These methods make training and inference more efficient and make large models more accessible to researchers.
#25 - Tue Apr 18	Guest speaker: Yizhong Wang (UW) [slides]
#26 - Thu Apr 20	Guest speaker: Ruiqi Zhong (UCBerkeley) [slides]	Title: Getting AI to Do Things I Can't: Scalable Oversight via Indirect Supervision Abstract: Can we tame powerful AI systems even when we struggle to determine the ground truth ourselves? In this talk, I will cover two example NLP tasks: 1) automatically searching for patterns in large text collections and explaining them to humans in natural language; 2) labeling complex SQL programs using non-programmers with the aid of our AI system and achieving accuracy on par with database experts. In both cases, we build tools that help humans to indirectly scrutinize the AI’s output with high effectiveness but low effort, bringing new insights that human experts have not anticipated.		Midway reports deadline
#27 - Tue Apr 25	Guest speaker: Max Ryabinin (Yandex) [slides]
#28 - Thu Apr 27	Guest speaker: Rogério Bonatti (Microsoft) [slides]
#29 - Tue May 2	No Class - Reading Days
#30 - Thu May 4	No Class - Reading Days
May 15 EOD	Final project reports
May 15	Final project poster session (9 am-12 pm - Hackerman hall)

Relevant Resources

Here are several resources available for free:

Compute resources:
- Google Colab provides free GPU usage for up to 12 hours/day for academic purposes. One can obtain more compute on Colab with relatively minimal pay.
- Google offers research TPU credits.
- AWS and Azure both offer welcome credits to students.
- If you need credits to use GPT3, discuss it with the instructor.
Demos:
Tutorials:

PyTorch documentation
These tutorials do a good job of introducing PyTorch.
A course on Huggingface's Transformers library.
Tutorials on Learn with Torch Lightning

Besides these resources, we will try our best to satisfy individual needs through discussion.

Conduct

The strength of the university depends on academic and personal integrity. In this course, you must be honest and truthful, abiding by the Computer Science Academic Integrity Policy:

Cheating is wrong. Cheating hurts our community by undermining academic integrity, creating mistrust, and fostering unfair competition. The university will punish cheaters with failure on an assignment, failure in a course, permanent transcript notation, suspension, and/or expulsion. Offenses may be reported to medical, law or other professional or graduate schools when a cheater applies. Violations can include cheating on exams, plagiarism, reuse of assignments without permission, improper use of the Internet and electronic devices, unauthorized collaboration, alteration of graded assignments, forgery and falsification, lying, facilitating academic dishonesty, and unfair competition. Ignorance of these rules is not an excuse.

Academic honesty is required in all work you submit to be graded. Except where the instructor specifies group work, you must solve all homework and programming assignments without the help of others. For example, you must not look at anyone else’s solutions (including program code) to your homework problems. However, you may discuss assignment specifications (not solutions) with others to be sure you understand what is required by the assignment. If your instructor permits using fragments of source code from outside sources, such as your textbook or on-line resources, you must properly cite the source. Not citing it constitutes plagiarism. Similarly, your group projects must list everyone who participated.

In the above paragraph "outside sources" also include content that was produced by an AI assistant like ChatGPT. This follows either by treating the AI assistant as a person for the purposes of this policy (controversial) or acknowledging that the AI assistant was trained directly on people's original work. Thus, while you are not forbidden from using these tools, you should consider the above policy carefully and quote where appropriate. Assignments that are in large part quoted from an AI assistant are very unlikely to be evaluated positively. In addition, if a student's work is substantially identical to another student's work, that will be grounds for an investigation of plagiarism regardless of whether the prose was produced by an AI assistant.

Falsifying program output or results is prohibited. Your instructor is free to override parts of this policy for particular assignments. To protect yourself: (1) Ask the instructor if you are not sure what is permissible. (2) Seek help from the instructor, TA or CAs, as you are always encouraged to do, rather than from other students. (3) Cite any questionable sources of help you may have received.

Report any violations you witness to the instructor. You can find more information about university misconduct policies on the web for undergraduates and undergraduates students.

Johns Hopkins University is committed to equal opportunity for its faculty, staff, and students. To that end, the university does not discriminate on the basis of sex, gender, marital status, pregnancy, race, color, ethnicity, national origin, age, disability, religion, sexual orientation, gender identity or expression, veteran status, military status, immigration status or other legally protected characteristic. The University's Discrimination and Harassment Policy and Procedures provides information on how to report or file a complaint of discrimination or harassment based on any of the protected statuses listed in the earlier sentence, and the University’s prompt and equitable response to such complaints.

CS 601.471/671 NLP: Self-supervised Models

Johns Hopkins University - Spring 2023