Publications

Feature learning

• Iteratively reweighted kernel machines efficiently learn sparse functions
  Zhu  L.,  Davis  D.,  Drusvyatskiy  D. & Fazel  M. Pre‑printed.

• Emergence in non‑neural models: grokking modular arithmetic via average gradient outer product
  Mallinar  N.,  Beaglehole  D.,  Zhu  L.,  Radhakrishnan  A.,  Pandit  P. & Belkin  M. ICML 2025.

• Catapults in SGD: spikes in the training loss and their impact on generalization through feature learning
  Zhu  L.,  Liu  C.,  Radhakrishnan  A. & Belkin  M. ICML 2024.

• Quadratic models for understanding catapult dynamics of neural networks
  Zhu  L.,  Liu  C.,  Radhakrishnan  A. & Belkin  M. ICLR 2024.

• A note on Linear Bottleneck networks and their Transition to Multilinearity
  Zhu  L.,  Pandit  P. & Belkin  M. Pre‑printed.

Kernel regime

• Neural tangent kernel at initialization: linear width suffices
  Banerjee  A.,  Cisneros‑Velarde P.,  Zhu  L. & Belkin  M. UAI 2023.

• Restricted Strong Convexity of Deep Learning Models with Smooth Activations
  Banerjee  A.,  Cisneros‑Velarde  P.,  Zhu  L. & Belkin  M. ICLR 2023.

• Transition to Linearity of General Neural Networks with Directed Acyclic Graph Architecture
  Zhu  L.,  Liu  C. & Belkin  M. NeurIPS 2022.

• Transition to Linearity of Wide Neural Networks is an Emerging Property of Assembling Weak Models
  Liu  C.,  Zhu  L. & Belkin  M. ICLR 2021 (Spotlight).

• On the linearity of large non‑linear models: when and why the tangent kernel is constant
  Liu  C.,  Zhu  L. & Belkin  M. NeurIPS 2020 (Spotlight).

• Loss landscapes and optimization in over‑parameterized non‑linear systems and neural networks
  Liu  C.,  Zhu  L. & Belkin  M. Applied and Computational Harmonic Analysis (ACHA) 2022.

Random matrix theory

• Spectral norm bound for the product of random Fourier-Walsh matrices
  Zhu  L.,  Davis  D.,  Drusvyatskiy  D. & Fazel  M. Pre‑printed.