Google AI Research: Looking Back at 2020, and Forward to 2021
In a recently published blog post, Google Research say they published more than 800 research publications in the past year.
In 2020, as the world has been reshaped by COVID-19, Google says they saw the ways research-developed technologies could help billions of people better communicate, understand the world, and get things done. Proud of what they’ve accomplished, and excited about new possibilities on the horizon, they have published long posts covering numerous categories including research work done in robotics, COVID-19, AutoML, Open Datasets, Reinforcement Learning, Natural Language Understanding, and many more.
The goal of Google Research is to work on long-term, ambitious problems across a wide range of important topics — from predicting the spread of COVID-19 to designing algorithms, to learning to translate more and more languages automatically, to mitigating bias in ML models.
Google is also dedicated to ensuring that research is done responsibly and has a positive impact, using their AI Principles as a guiding framework and applying particular scrutiny to topics that can have broad societal impacts.
Check out their post for a more comprehensive look, there’s something for everyone!
A Python Library Providing Fast Interpretable Matching for Causal Inference
As human beings, we think in terms of cause and effect. When we understand why something happens, we can change our behavior and improve future outcomes. Causal inference enables AI and machine learning algorithms to reason in a similar way.
Understanding cause and effect in AI and ML systems can make them smarter and more efficient. The ability to understand causality can help us, ML engineers, to create models that better understand their data.
In this paper, researchers introduce dame-flame, a Python package providing two algorithms, DAME and FLAME, for fast and interpretable treatment-control matches of categorical data.
The dame-flame package offers efficient, easy-to-use implementations of the DAME and FLAME algorithms, allowing users to perform fast, and interpretable matching for causal inference for observational data with discrete covariates. The package is easily accessible and accompanied by detailed documentation, with concrete examples. The package is written in a highly modular manner, facilitating the introduction of new features and variations of the DAME and FLAME algorithms.
The package provides several adjustable parameters to adapt the algorithms to specific applications, and can calculate treatment effects after matching. Descriptions of these parameters, details on estimating treatment effects, and further examples are available here.
Access on GitHub
A Proven Systems Engineering Approach for Machine Learning Development and Deployment
The development and deployment of machine learning (ML) systems can be executed easily with modern tools. However, but the process can get rushed and a lack of diligence can lead to technical debt, scope creep and misaligned objectives, model misuse and failures, and expensive consequences.
To address the challenge, this research unveils Machine Learning Technology Readiness Levels (MLTRL), an industry-hardened systems engineering framework for robust, reliable, and responsible machine learning. MLTRL is derived from the processes and testing standards of spacecraft development, yet lean and efficient for ML and software workflows.
Examples from several organizations across industries demonstrate the effectiveness of MLTRL for AI and ML technologies, from research and development through productization and deployment, in important domains such as healthcare and physics.
Moreover, MLTRL establishes a much-needed jargon for the AI ecosystem. In the future, researchers aim to provide open-source tools for more teams and organizations to adopt MLTRL. Their hope is that the framework is adopted broadly in AI and ML organizations, and that “technology readiness levels” become common nomenclature across AI stakeholders – from researchers and engineers to sales-people and executive decision-makers.
Researchers Show That Progress Can Be Made On Machine Ethics
The demand for ethical machine learning has led researchers to study and propose various ethical principles for applications.
Recently, a collaboration of researchers with Microsoft, UC Berkeley, UChicago, and Columbia University has introduced an ETHICS dataset to assess a machine learning system’s ability to predict necessary human ethical judgments in open-world settings.
The dataset is based on natural language scenarios, enabling engineers to construct diverse situations involving interpersonal relationships, everyday events, and thousands of objects. This means models must connect various facts about the world to their ethical consequences. For instance, taking a penny on the street is usually acceptable, whereas taking cash from a wallet lying on the street is not.
The new benchmark spans concepts in justice, well-being, duties, virtues, and commonsense morality.
Researchers found that with the ETHICS dataset, current language models have a promising but incomplete ability to predict basic human ethical judgments. Their work shows that progress can be made on machine ethics today, and it provides a steppingstone toward AI that is aligned with human values.
A New Reinforcement Learning Approach for Providing Quality of Service
Providing guarantees on the end-to-end delay of a service system is of great interest to both customers and service providers.
Current state of the art methods do not directly consider the delay of the system in their problem formulation and therefore, cannot provide any delay guarantees.
This paper introduces an RL-based framework that takes a QoS constraint as input and provides a dynamic service-rate control algorithm that satisfies the constraint without overuse of the service resources. This makes our method distinct from the existing service-rate control algorithms that quantify their performance by the achieved overall reward, which is highly dependent on the reward definition and might have no practical interpretations.
The proposed method is capable of guaranteeing probabilistic upper-bounds on the end-to-end delay of the system, only using the queue length information of the network and without any knowledge of the system model. Since a general queueing model has been used in this study, our method can provide insights into various applications, such as VNF chain auto-scaling in the cloud computing context.