William Kong William Kong
William Kong
Research Software Engineer · Google Research (NYC)

Weiwei "William" Kong

Hi! I'm William Kong, a Research Software Engineer at Google Research (NYC) in Athena. Here you'll find my publications, open-source software, mathematical blog posts, and course notes.

Previously, I was a postdoc in the CSMD at Oak Ridge National Labs under Miroslav Stoyanov, a Research Intern at Google Research (MTV), and a Senior Risk Analyst at TD Bank.

I received my Ph.D. in Operations Research from Georgia Tech under Renato D. C. Monteiro, along with an M.Sc. in Computational Science & Engineering (Georgia Tech) and a B.Math. from the University of Waterloo.

Research Pillars

27 papers →

Differential Privacy & Scalable DP-ML

Fast per-example gradient clipping, privacy auditing (DP-Auditorium), non-decomposable DP optimization, and production DP libraries in JAX (JAX-Privacy).

Synthetic Data & LLM Training

Escaping model collapse with weak data, limitations of ERM on synthetic corpora, private prediction for synthetic text, and ContinuousBench.

Nonconvex Composite Optimization

Accelerated proximal, penalty, spectral, and augmented Lagrangian solvers with certified stationarity bounds—implemented in Proxax (JAX) and NC-OPT.

Interactive Optimization Sandbox Inspired by Proxax

Nonconvex Banana Valley: Why Adaptive Acceleration Matters

Y-Axis Metric
Fixed-Step GD Backtracking GD Nesterov AGD (Fixed) Adaptive Nesterov Adaptive Accelerated (Proxax APD)
Rosenbrock Banana Valley: f(x1,x2)=(1−x1)2+b(x2−x12)2f(x_1, x_2) = (1 - x_1)^2 + b(x_2 - x_1^2)^2
log⁡10(f(xk)−f∗)\log_{10}(f(x_k) - f^*) vs. Iteration kk
Solve this nonconvex problem in Python / JAX with Proxax (px.solvers.APD) — or read the Acceleration blog series
import jax.numpy as jnp
from proxax import px

b = 100.0  # Rosenbrock valley steepness (nonconvex, kappa ≈ 2500 at x*)
def rosenbrock(x):
  return (1.0 - x[0]) ** 2 + b * (x[1] - x[0] ** 2) ** 2

problem = px.CompositeProblem(smooth=rosenbrock)
sol = px.solve(
    problem=problem,
    x0=jnp.array([-1.2, 1.0]),
    solver=px.solvers.APD(opt_tol=1e-6),
)
print(sol.status_enum.name, sol.x, sol.stats.opt_residual)

The live plot above runs directly in your browser in JavaScript (no backend server). Because the Rosenbrock valley is nonconvex and severely ill-conditioned (local condition number κ≈2500\kappa \approx 2500 at x∗=(1,1)x^* = (1, 1) when b=100b = 100), fixed-step GD crawls along the curved trough. Adaptive curvature estimation combined with accelerated extrapolation and restart—as implemented in Proxax (px.solvers.APD)—reaches machine precision orders of magnitude faster without tuning LL or μ\mu.