Meet the Fellows Welcome Event Fall 2024
2024-09-03..2024-09-04
Berkeley, United States
[Meet the Fellows Welcome Event, Meet the Fellows Welcome Event Fall 2024]
who it's for
machine learning researchers, theoretical computer scientists, quantum computing researchers, statistical learning theorists, postdoctoral researchers, graduate students, applied deep learning researchers
topics
Quantum Hamiltonian Descent; differentially private mean estimation; outlier-robust estimation algorithms; sub-quadratic sparse mean estimation; testable agnostic halfspace learning; computational limits of continual learning; finite-automaton analysis of Transformers; linear representation hypothesis model extraction; edge-of-stability training dynamics; SDEs for distributed optimizers; square-root batch scaling rule; feature learning sample complexity
| title | author | length |
|---|---|---|
| Quantum approach to classical optimization: why bother and what to do? | Jiaqi Leng | 14m |
Introduces Quantum Hamiltonian Descent (QHD) as a path‑integral quantization of gradient descent, implements it with Hamiltonian simulation and analog… | ||
| Data privacy and reliable learning | Lydia Zakynthinou | 17m |
Shows that private learners can achieve dimension-free error for some mean estimation tasks by using instance-adaptive bounds and exploiting low effective… | ||
| Towards Practical Algorithms for Outlier-Robust Estimation | Ankit Pensia | 17m |
Claims the first algorithm for sparse mean estimation with provable sub-quadratic runtime (for any Q>3) while maintaining near‑optimal error and sample… | ||
| Efficient Testable Agnostic Learning for Halfspaces | Arsen Vasilyan | 17m |
First efficient algorithm that is testable agnostic for learning halfspaces under Gaussian (or strongly log-concave) inputs, running in polynomial time and… | ||
| Theoretical understanding of learning through the computational lens | Binghui Peng | 12m |
Proves that worst-case continual learning requires memory scales linearly with number of tasks but shows allowing multiple passes over the task stream removes… | ||
| Guiding machine design using insights from simple sandboxes | Bingbin Liu | 18m |
Research uses minimal finite-automaton sandboxes to map which sequential tasks Transformers can implement exactly, proving solvable automata decompose into two… | ||
| Distilling models using the linear representation hypothesis | Enric Boix | 12m |
Claim: neural nets that implicitly encode decision trees via the linear representation hypothesis admit extraction. Method: run many linear probes (ordinary… | ||
| The dynamical view of machine learning: implicit biases of large learning rates | Yuqing Wang | 13m |
Large learning rates cause gradient descent to escape sharp minima toward flatter ones and produce edge of stability behavior where limiting sharpness… | ||
| SDE Approximations and Scaling Rules for Distributed Deep Learning | Kaifeng Lyu | 16m |
Derives stochastic differential equation approximations for large‑batch adaptive optimizers (RMSprop and Adam) and proves a Square Root Scaling Rule that… | ||
| Feature Learning: From Optimization to Modern Generalization | Kaifeng Lyu, Margalit Glasgow | 20m |
Shows that a two-layer neural network trained with online minibatch SGD and logistic loss on Boolean hypercube data learns an XOR-style multi-index target… | ||
| nothing matches | ||