Yi (Joshua) Ren

Intro. of My Research

I am exploring how to train models that generalize well systematically and why effective models naturally adhere to Occam's Razor. This idea stems from a talk by Professor Simon Kirby, which discussed how the pressures of compressibility and expressivity drive human language to evolve in a more compositional direction. We observe that not only humans but also neural networks tend to favor highly compositional mappings when trained on various tasks. Such a simplicity bias may be progressively amplified if an intelligent agent continuously learns from the data and experiences of its predecessors, which is a key concept in Bayesian-iterated learning (Bayesian-IL) in cognitive science. Investigating this intriguing framework has inspired two lines of my previous research.

First, I worked on extending the iterated learning framework to more general deep learning systems, starting with an emergent communication setting (a two-agent cooperative RL game, in Ren et al., ICLR 2020) and progressing to broader representation learning problems, including vision, language, and even molecular graphs (in Ren et al., NeurIPS 2023). The latter was achieved during my visit to Professor Aaron’s group at Mila. We discovered that a bottleneck in the network structure plays a critical role in introducing implicit bias, which is further amplified through multi-generation self-play. Additionally, our recent work demonstrates that Bayesian-IL partially explains the evolution of large language models (LLMs) in pervasive self-play (in Ren et al., NeurIPS 2024). This not only sheds light on why specific phenomena, such as diversity reduction and hidden bias amplification, occur in many self-improvement methods but also offers insights into mitigation -- namely, designing effective interaction phases to constrain unwanted biases.

Another line of work, which I am currently focusing on, explores the origins of the simplicity bias. One theoretical tool we use is learning dynamics, which examines how a model’s prediction on one example changes when it learns from another. This tool allows us to quantify simplicity bias through measurable properties such as learning speed or compression rate. Since a model cannot learn all possible data at once, it acquires new knowledge sample by sample. If learning from each example enables the model to make more accurate predictions on other samples, the training curve will decay rapidly, indicating a higher compression ratio (as highlighted in this talk on Compression for AGI). We have also applied learning dynamics to explain various intriguing behaviors in deep learning, such as identifying better supervisory signals (Ren et al., ICLR 2022), designing fine-tuning heads (Ren et al., ICLR 2023), and fine-tuning large language models (Ren et al., ICLR 2025). Interestingly, we found that self-preference amplification and simplicity bias are pervasive in gradient descent-based learning systems (Ren et al., CompLearn@NeurIPS 2024).

In addition to this microcosmic inspection of the learning process, I have recently realized that compression theory and Kolmogorov complexity-related theories offer a more macroscopic perspective on simplicity bias. I believe the mechanisms underlying these concepts -- such as Occam's Razor, the Platonic Representation Hypothesis , learning speed advantage, and systematic generalization -- may reflect fundamental principles of learning theory. I am eager to further explore this fascinating direction and uncover deeper insights.

Talks

Notes and TA