what's new in machine learning foundations

Olivier Jeunen

this paper proposes using counterfactual estimation and policy overlap to reduce variance in a/b tests, thereby accelerating them. practitioners can leverage this to run online experiments more efficiently, achieving statistically significant results faster and reducing the cost of experimentation.

David P. Woodruff, Samson Zhou

this paper investigates the adversarial robustness of sketching and streaming algorithms, which are vital for massive datasets. practitioners deploying these algorithms in sensitive or adversarial environments can understand their vulnerabilities and design more robust systems against data poisoning or manipulation.

Federico L. Perlino, Oliver Hamelijnck, Adam M. Johansen +1

this paper introduces deep gaussian processes (dgps) adapted for directed acyclic graphs (dags), which are common in real-world processes. practitioners working with structured data, like causal graphs or biological networks, can use this to model complex relationships with uncertainty quantification.

Eitan Levin, Venkat Chandrasekaran

this paper proposes a learning framework for models that handle inputs of varying dimensions, like point clouds or sequences. practitioners working with variable-sized data can leverage this approach to develop more generalizable models without needing fixed-size input transformations.

Eli N. Weinstein, David M. Blei

this paper introduces geometric causal models for drawing causal inferences from structured, non-i.i.d. data like spatial data. practitioners dealing with complex, structured datasets can use this framework to uncover causal relationships more accurately, moving beyond traditional i.i.d. assumptions.

Patrik Wolf, Thomas Kleine Buening, Andreas Krause +1

this work proposes "partition, prompt, aggregate" to achieve statistical self-consistency in large language models, improving their reliability. practitioners can apply this technique to enhance the robustness and accuracy of llm outputs, especially in tasks requiring consistent reasoning.

Jianing Liu, Dong H. Zhang

this paper enhances the levenberg-marquardt optimization algorithm by incorporating higher-order geometric updates using riemann normal coordinates. this modification aims to improve the convergence speed and robustness of the method. practitioners can benefit from this for more efficient and reliable solutions to non-linear least-squares problems common in fields like robotics and computer vision.

Xiaochun Niu, Tselil Schramm, Jiaming Xu

this paper introduces a method for high-dimensional procrustes matching using tree counts, which helps align two sets of high-dimensional data points. it provides a robust way to find optimal transformations between datasets, even in the presence of noise or permutations. practitioners can use this for tasks like shape analysis, comparing complex data structures, or aligning point clouds.

Dan Yamins, Aran Nayebi

this paper proposes contravariance theory, a framework for finding minimal and strongly aligned solutions to complex machine learning tasks. it aims to simplify models while maintaining high performance. for practitioners, this could lead to more efficient, interpretable, and resource-friendly models, particularly for challenging problems.

Yao Fu, Chunxia Zhang, Junmin Liu +3

this paper proposes an improved sharpness-aware minimization (sam) technique for deep learning by incorporating extragradient methods. this enhancement helps models achieve better generalization and robustness. practitioners can use this optimized training approach to develop more stable and high-performing deep neural networks.

Yuqing Wang

this paper investigates the convergence of gradient descent for neural networks, moving beyond the well-understood neural tangent kernel (ntk) regime. it offers theoretical insights into how gradient descent behaves in more general, practical network settings. practitioners can use this understanding to better grasp why deep learning models converge and to potentially design more stable training procedures.

Linda Lu, Karthik Sridharan

this paper introduces predictability as a fine-grained privacy measure, offering an alternative to the worst-case guarantees of differential privacy. for practitioners, this new metric could enable more nuanced privacy-accuracy tradeoffs, potentially leading to more performant models while still providing meaningful privacy assurances.