Date: August 15, 2021 (Sunday), 11AM-5:05PM ET

NOTE: Due to the uncertainty regarding the COVID-19 outbreak, the KDD conference has transitioned to a virtual event (visit the KDD 2021 website for details). Thus, this workshop has transitioned to a virtual workshop accordingly. Connection details will be added once they become available.

To access the virtual conference, go to https://virtual.2021.kdd.org/index.html

Scroll to the bottom for detailed agenda

Keynote Speakers

AutoML for Tiny Machine Learning

Today’s AI is too big, it's not only computationally intensive but also labor intensive to tune the model architectures, making it difficult to be deployed on different hardware devices at scale. In this talk I will introduce AutoML techniques to automatically synthesize efficient and tiny neural architectures given hardware constraints, outperforming the human design while requiring less engineer efforts. I will introduce how to automatically compress the model and co-design the inference framework to fit tiny devices. Besides, deep learning models need to be deployed on diverse hardware platforms, from cloud to edge. It requires a vast amount of computational resources to train models of different sizes. I will introduce once-for-all (OFA) network that trains one large neural network comprising many subnetworks of different sizes, so we can easily select specialized subnetworks for different hardware constraints without retraining. An extension of this work, MCUNet will be introduced that can perform ImageNet classification on microcontrollers that has only 1MB of Flash .

AutoML Systems in Action

Automated Machine Learning (AutoML) has become a very important research topic with wide applications of machine learning techniques. While many computational algorithms have been developed, this talk will focus on a complementary direction to introduce how to design an effective AutoML system in practice based on our existing open-sourced softwares. First, we will present an open-source AutoML system design, namely AutoKeras, based upon a novel method for efficient neural architecture search with network morphism. Second, we will discuss a pilot study of automatically designing architectures for the CTR prediction task, as well as AutoRec, an automated Recommender System. At last, an automated anomaly detection system, called AutoOD, via curiosity-guided search and self-imitation learning will be introduced. Through three real-world examples, we demonstrate how to bridge the gap between AutoML algorithms to systems in real production environments.

Grey Box Bayesian Optimization for AutoML

Bayesian optimization (BayesOpt) has long been an important search strategy for autoML, for both hyperparameter tuning and increasingly for neural architecture search. The first applications of BayesOpt to autoML assumed that validation error was a black box function of the hyperparameters, ignoring other information available during model training and also ignoring the ability to modify model training to more quickly access this information. Dramatic advances in the computational efficiency of BayesOpt for autoML have been achieved by moving beyond this assumption, leveraging the ability to "look inside" this black box. These approaches optimize ML models more quickly by observing how validation error improves with training iterations, observing validation error on individual folds, and by performing approximate model evaluations using less training data or (in vision applications) lower-resolution images. We describe these powerful grey-box BayesOpt approaches to autoML and promising new advances in general-purpose grey-box BayesOpt whose power have not yet been fully explored for autoML.

Fairness-Aware AutoML

Despite recent awareness, tackling algorithmic bias is still perceived as expensive by most industries. Model unfairness in production has several sources, from training data sampling to label definitions, pre-processing transformations to feature engineering, model selection criteria or simply concept drift. The current ML landscape lacks practical methodologies and tools to seamlessly integrate fairness objectives and bias reduction techniques in existing real-world ML pipelines. As a consequence, treating fairness as a primary objective when developing ML systems is not yet standard practice. Existing bias reduction techniques only target specific stages of the ML pipeline (e.g., data sampling, model training), and often only apply to a single fairness definition or family of ML models, limiting their adoption in practice. Moreover, the absence of major breakthroughs in algorithmic fairness suggests that an exhaustive search over all possible ML models and bias reduction techniques may be necessary to find optimal trade-offs. To overcome these limitations, we discuss a simple and easily deployed intervention: Fairness-Aware AutoML. By making the hyperparameter search fairness-aware while treating every decision on the ML pipeline as an hyperparameter, we are enabling ML practitioners to adapt pre-existing business operations to accommodate fairness with controllable extra cost, and without significant implementation friction.

Paper Presentations

• To tune or not to tune? An Approach for Recommending Important Hyperparameters - Mohamadjavad Bahmani, Radwa El Shawi, Nshan Potikyan and Sherif Sakr.

• Defining General-Purpose Machine Learning Problems for AutoML - Suilan Estevez-Velarde, Alejandro Piad, Yudivián Almeida, Ernesto Luis Estevanell-Valladares, Yoan Gutiérrez Vázquez, Andres Montoyo and Rafael Muñoz.

• Sequential Automated Machine Learning: Enhancing the Collaborative Filtering approach with a LinUCB exploration policy - Maxime Heuillet.

• Pooling Architecture Search for Graph Classification - Lanning Wei, Huan Zhao, Quanming Yao and Zhiqiang He.

Agenda

11:00 AM – 11:10 AM ET Opening remarks and workshop introduction

11:10 AM – 11:55 PM ET Keynote 1: Peter I. Frazier, Grey Box Bayesian Optimization for AutoML

12:00 PM – 12:30 PM ET Paper 1: To tune or not to tune? An Approach for Recommending Important Hyperparameters, Mohamadjavad Bahmani, Radwa El Shawi, Nshan Potikyan and Sherif Sakr.

12:30 PM – 12:40 PM ET Break 1

12:40 PM – 1:25 PM ET Keynote 2: Ben Hu, AutoML Systems in Action

1:30 PM – 2:00 PM ET Paper 2: Defining General-Purpose Machine Learning Problems for AutoML

Suilan Estevez-Velarde, Alejandro Piad, Yudivián Almeida, Ernesto Luis Estevanell-Valladares, Yoan Gutiérrez Vázquez, Andres Montoyo and Rafael Muñoz.

2:00 PM – 2:10 PM ET Break 2

2:10 PM – 2:55 PM ET Keynote 3: Pedro Saleiro, Fairness-Aware AutoML

3:00 PM – 3:30 PM ET Paper 3: Sequential Automated Machine Learning: Enhancing the Collaborative Filtering approach with a LinUCB exploration policy, Maxime Heuillet.

3:30 PM – 3:40 PM ET Break 3

3:40 PM – 4:25 PM ET Keynote 4: Song Han, AutoML for Tiny Machine Learning

4:30 PM – 5:00 PM ET Paper 4: Pooling Architecture Search for Graph Classification

Lanning Wei, Huan Zhao, Quanming Yao and Zhiqiang He.

5:00 PM - 5:05 PM ET Closing remarks