A quick guide to Amazon's papers on ICML 2024

A quick guide to Amazon’s papers on ICML 2024

by

Amazon’s papers on International Conference on Machine Learning (ICML) Lean – as the conference as a whole – against the theoretical. Although some papers deal with important applications for Amazon, such as anomaly detection and automatic speech recognition, they are most concerned with more-general items related to machine learning, such as responsible AI and transfer learning. Learning algorithms, reinforcement learning and privacy as areas of special interest.

Active learning

Understanding the training speed of sampling with approximate losses
Rudrajit Das, Xi Chen, Bertram Ieong, Bansal Parikshit, Sujay Sanghavi

Anomali -Detection

Online adaptive anomali threshold with self -confidence sequences
Sophia Sun, Abishek Sankararaman, Balakrishnan (Murali) Narayanaswamy

“Online adaptive anomaly -threshold with confidence sequences” suggested for method of adapting anomaly detection limits to distribution operation. In the experiment, the researchers model a signal as a number of hand -drawn figures from the Mnist Data set and distribution operation as a change from the number 0 To numbers 1. Anomalies were modeled like other numbers than 0 gold 1.

Automatic speech recognition

An effective self -learning framework for interactive spoken dialog systems
Hitesh Tulsiani, David M. Chan, Shalini Ghosh, Garima Lalwani, Prabhat Pandey, Ankish Bansal, Sri Garimella, Ariya Rastrow, Björn Hoffmeister

Cause inference

Multiplicer-Robust causal change task
Victor Quintas, Taha Bahadori, Eduardo Santiago, Jeff Mu, Dominik Janzing, David E. Heckerman

Completion code

Repofform: Selective retrieval for completion code on Refund Levels
Di wu, wasi ahmad, dejiao zhang, murali krishna ramanathan, xiaofei ma

Continuous learning

MEMORYLM: Self -stay against large language models
Yu Wang, Yifan Gao, Xusi Chen, Haoming Jiang, Shiyang Li, Jingfeng Yang, Qingyu Yin, Zheng Li, Xian Li, Bing Yin, Jingbo Shang, Julian McAuley

Contrastive learning

EMC2: Effective MCMC -Negative sampling for contrastive learning with global convergence
Chung Yiu Yau, Hoi-to Wai, Parameswaran Raman, Soumajyoti Sarkar, Mingyi Hong

Data preparation

Fewer trunkings enhance language modeling
Hantian Ding, Zijian Wang, Giovanni Paolini, Varun Kumar, Anoo Deoras, Dan Roth, Stefano Soatto

Explainable AI

Explanation of probable models with distribution values
Luca Franceschi, Michele Donini, Cérric Archambeau, Matthias Seeger

Game-theoretical approaches to explainable AI, such as Shaply-Value Anyse, compare output from a black-box model with and without a particular input function (Irepresent as a blue square). Such methods typically work on scalar values (top)Discarding information trapped by probabilistic models. “Explanation of probable models with distribution values” Games the theoretical approach to models with distribution of outputs (Bottom).

Mitigated hallucination

Multialiali-
Gianluca Detommaso, Martin Bertran Lopez, Riccardo Fogliato, Aaron Roth

Learning algorithms

Mada: Meta-adaptive optimizers through hypergradient descent
Kaan Ozkara, Can Karakus, Parameswaran Raman, Mingyi Hong, Shoham Sabach, Branislav Kveton, Volkan Cevher

Variance Reduced Zero-Order Methods for Finely-Ticking Language Models
Tanmay Gautam, Youngsuk Park, Hao Zhou, Parameswaran Raman, Wooseok Ha

LLM decoding

Bifurcated attention for single content stainer
Ben Athiwaratkun, Sujan Gonugondla, Sanjay Krishna Gouda, Hantian Ding, Qing Sun, Jun Wang, Jiacheng Guo, Liangfu Chen, Haifeng Qian, Among Bhatia, Ramesh Nallapati, Sudipta Bedup, Bing Xiang Xiang Qian,

Model compression

Collage: Lightweight Low-precision strategy for LLM training
Tao Yu, Gaurav Gupta, Karthick Gopalswamy, Friend Mamidala, Hao Zhou, Jeffrey Huynh, Youngsuk Park, Ron Diamond, Anoo Deoras, Luke Huan

Privacy

Differentially Private Bias-Term FINKING FUNDER MODELS
Zhiqi bu, yu-xiang wang, sheng zha, george karypis

Graphing calculation for backloping on plain (black) and different privately (Black and red) Algorithms. In “Differentially Private Bias-Term FINKING FUNDER MODELS” Amazon scientists show that only fine-tuning bias-expression (Bottom right) Is much simpler than fine tuning both bias expressions and weights (Left and top right)Preserve accuracy while doing training 2 – 30 times as quickly.

Membership Inferens Ventacks on Diffusion Models Via Quantile Regression
Shuai Tang, Zhiwei Steven Wu, Sergul Aydore, Michael Kearns, Aaron Roth

Reinforcement learning

End Time Convergence and Test Complexity of Actor Criticism Reinforcement Learning Learning Learning Learning
TIANCHEN ZHOU, FNU HAIRI, HAIBO YANG, JIA (KEVIN) LIU, TIAN TONG, FAN YANG, Michinari Mimage, Yan Gao

To learn the target network in the function space
Kavosh Asadi, Yao Liu, Shoham Sabach, Ming Yin, Rasool Fakoor

Most reinforcement learning (RL) involves a functional algorithms (V) It predicts the value of taking a particular action when an agent is in a particular state. Often, the value function is approximated by two neural networks (θ and w)One that models the current value estimate and one that is updated in the light of new interactions. Using the RL TAB function is designed to unite the parameter values ​​of the two models. But in “Learning the Target Network in Function Space” Amazon scientists show that uniting the models in the function space (Left) Don’t understand the barrel of them in the parameter space (right). Their experience shows that dropping the requirement of parameter value equivalence can improve the performance of RL tasks.

Almost optimal regret in linear MDPs with overall bandit feedback
Asaf Cassel, Haipeng Luo, Dmitry Sotnikov, Aviv Rosenberg

AI Responsible

Discover Bias in Latent Rum: An Unnsupervised Debiiasing -Authorization
Dyah Adila, Shuai Zhang, Boran Han, Bernie Wang

Performed Generated Generation

Automated evaluation of retrieval-augmented language models with assignment specific examination generation
Gauthier Guinet, Behrooz Omidvar-Tehrani, Anoo Deoras, Laurent Callot

Robust learning

Robust learning with multiple tasks with excess risks
Yifei he, shiji zhou, guojun zhang, hyokun yun, yi xu, belinda zeng, trishul chilimbi, he zhao

Scientific machine learning

Ucretaintainin quantification to characterize and improve domain learning for PDEs
S. Chandra Mouli, Danielle Maddix Robinson, Shima Alizadeh, Gaurav Gupta, Andrew Stuart, Michael Mahoney, Bernie Wang

Transfer of Learning

Transfer of knowledge from large foundation models to small downnstream models
Shikai Qiu, Boran Han, Danielle Maddix Robinson, Shuai Zhang, Bernie Wang, Andrew Wilson

Leave a Comment