Yuchen Li

I am a Ph.D. student in the Machine Learning Department at Carnegie Mellon University, where I am fortunate to be advised by Professor Andrej Risteski.

Previously, I spent four wonderful years at the University of Illinois at Urbana-Champaign, completing my bachelor's degree in Statistics & Computer Science, as well as Mathematics. I am grateful for the mentorship of Professor Jiawei Han, Professor AJ Hildebrand, and Professor Pramod Viswanath.

In the industry, I spent some exciting time at Google (summer 2023), ByteDance (TikTok) (summer 2022), Quora (2019-2020), and Facebook (summer 2018).

Email  /  CV (resume)  /  LinkedIn  /  Google Scholar  /  Twitter  /  GitHub  /  Quora

profile photo


Research

I study machine learning, natural language processing, and data mining. I currently work on (1) improving mathematical understanding of transformer-based language models (training dynamics, efficient sampling, mechanistic interpretability), and (2) developing principled approaches to self-supervised learning.



Publications

In the following list, the asterisk symbol (*) in the author list means equal contribution or alphabetical order.

Typical attention patterns do not exactly match the intuitively interpretable stack-like pattern Transformers are uninterpretable with myopic methods: a case study with bounded Dyck grammars
Kaiyue Wen, Yuchen Li, Bingbin Liu, Andrej Risteski

Conference on Neural Information Processing Systems (NeurIPS), 2023
bibtex / slides / Twitter summary / video recording of 15-minute talk (starts at 5:39:20)

We show that interpretability claims based on isolated attention patterns or weight components can be (provably) misleading.

Our results are based on next-token prediction on Dyck (i.e. balanced brackets). We show that Transformers (exactly or approximately) solving this task satisfy a structural characterization, “balance condition“, derived from ideas in formal languages (the pumping lemma).

Embedding weight dot product of models trained on synthetic topic modeling data How Do Transformers Learn Topic Structure: Towards a Mechanistic Understanding
Yuchen Li, Yuanzhi Li, Andrej Risteski

International Conference on Machine Learning (ICML), 2023
bibtex / slides / Twitter summary / project page / codes

We analyze the optimization process of transformers trained on data involving "semantic structure", e.g. topic modeling. Through theoretical analysis and experiments, we show that between same-topic words, the embeddings should be more similar, and the average pairwise attention should be larger.

In particular, we observe that with carefully chosen initialization and learning rate, the optimization process of self-attention can be approximately broken down into two stages: in stage 1, only the value matrix changes significantly; in stage 2, the key and query matrices catch up much later, even though all components are jointly optimized through standard SGD or Adam. This observation might be of independent interest, for future works on understanding the learning dynamics of transformers as well.

Minimum Max-Cosine for noncontrastive loss (NCL) and contrastive loss (CL) on an architecture with a randomly initialized linear encoder. Contrasting the landscape of contrastive and non-contrastive learning
Ashwini Pokle*, Jinjin Tian*, Yuchen Li*, Andrej Risteski

Conference on Artificial Intelligence and Statistics (AISTATS), 2022
bibtex / Twitter summary

We show through theoretical results and controlled experiments that even on simple data models, non-contrastive losses have a preponderance of non-collapsed bad minima (which do not exist for their contrastive loss counterpart). Moreover, we show that the training process does not avoid these minima.

pcfg_context The Limitations of Limited Context for Constituency Parsing
Yuchen Li, Andrej Risteski

Association for Computational Linguistics (ACL), 2021
bibtex / slides / Twitter summary / Andrej's Tweet / blog post

We prove that, if the context for the parsing decision at each word is unbounded & bidirectional, then the parsing model has full representational power.

On the other hand, even if the context is bounded in one direction (which is the case for various leading approaches to constituency parsing), the parsing power is quite limited for hard grammars.

complexity Complexity of Leading Digit Sequences
Xinwei He*, AJ Hildebrand*, Yuchen Li*, Yunyi Zhang*

Journal of Discrete Mathematics & Theoretical Computer Science (DMTCS), vol. 22 no. 1, Automata, Logic and Semantics, 2020
bibtex / poster

Consider the sequence of leading digits of 2^n: 1, 2, 4, 8, 1, 3, 6, ...... What is the complexity of such sequences? They are neither periodic nor random. How to quantify and distinguish the middle ground?

We prove that, the block complexity is linear for the sequence of leading digits of a^n in base b (except for some special cases). In fact, we give explicit formula for the linear coefficients in terms of a and b.

HyperMine Discovering Hypernymy in Text-Rich Heterogeneous Information Network by Exploiting Context Granularity
Yu Shi*, Jiaming Shen*, Yuchen Li, Naijing Zhang, Xinwei He, Zhengzhi Lou, Qi Zhu, Matthew Walker, Myunghwan Kim, Jiawei Han

Conference on Information and Knowledge Management (CIKM), 2019
bibtex

This work discovers hypernym-hyponym ("is-a" relation) pairs by an innovative combination of information from natural language and graph structure. Our system outperforms existing methods that utilize signals from text, graph, or both.

spell_correction Context-Sensitive Malicious Spelling Error Correction
Hongyu Gong, Yuchen Li, Suma Bhat, Pramod Viswanath

The Web Conference (WWW), 2019
bibtex

Given a misspelled word, our algorithm generates candidates based on the surface form, represents the context as a linear space, and selects the candidate whose embedding is closest to the context subspace. The spelling correction accuracy is better than the state-of-the-art methods, improving the performance in downstream applications.

hin_motif Temporal Motifs in Heterogeneous Information Networks
Yuchen Li*, Zhengzhi Lou*, Yu Shi, Jiawei Han

Workshop on Mining and Learning with Graphs (MLG), in conjunction with KDD, 2018
bibtex

This work shows the effectiveness of temporal motifs (frequent substructures) in quantifying the relationship between nodes in heterogeneous information networks, and proposes efficient algorithms for counting a family of useful motifs.




Talks


Typical attention patterns do not exactly match the intuitively interpretable stack-like pattern

Transformers are uninterpretable with myopic methods: a case study with bounded Dyck grammars
slides / video recording of 15-minute talk (starts at 5:39:20)

2023-11 DeepMath Conference on the Mathematical Theory of Deep Neural Networks

transformers can learn the topic structure in the data distribution How Do Transformers Learn Topic Structure: Towards a Mechanistic Understanding
slides

2023-05 NeuLab, Language Technologies Institute (LTI), CMU

2023-04 University of Illinois at Urbana-Champaign (UIUC)’s Machine Learning Reading Group

pcfg_context The Limitations of Limited Context for Constituency Parsing
slides

2021-08 Association for Computational Linguistics (ACL) Conference

2021-07 NEC Laboratories Europe

2021-06 Approximately Correct Machine Intelligence (ACMI) Lab, CMU


Teaching



a neural network model Advanced Introduction to Machine Learning (CMU 10-715, Fall 2022)

Role: Teaching Assistant

Course Webpage: This URL

Instructor: Nihar Shah
a directed graphical model Probabilistic Graphical Models (CMU 10-708, Spring 2022)

Role: Teaching Assistant

Course Webpage: This URL

Instructors: Andrej Risteski and Hoda Heidari



Template of this page: credit to: link.