ReuseML: Towards a More Sustainably Focused AI Pipeline with Focus on Tactile Robotics 

Sustainability is a motivating factor for change in many aspects of society. There is a growing culture to efficiently utilise resources. However, this is an attitude that is not always shared when designing and implementing AI pipelines, where often all available power-hungry compute and large-scale data resources are often thrown at an AI task. 

Considering recent events like the energy crisis it is more relevant than ever to take sustainably in to account when researching and deploying AI systems. I think that the interaction between AI and sustainability is one of the more topical influences that AI is going to have on our society. 

For my PhD project I plan to create tools for a sustainably focused AI framework that helps AI and machine learning practitioners to interact with AI in not only a responsible but also a sustainable way. This is outlined at three main steps of the AI pipeline: data, models and evaluation. I plan to answer key questions at each stage of this pipeline:  

1) For data, how much data is sufficient to train AI algorithms and what methods are there to reduce data requirements? There is also a need to assess data quality and how similar different datasets and tasks are before even training models. 

2) For models, are there ways to reuse pre-existing models so that resources are not wasted on creating new models? This will utilise techniques such as transfer and multitask learning as well as exploring and developing others. 

3) For evaluation of models, can we trust models more through convenient and accurate evaluation? This links back to reusing and adapting models when they fail during evaluation, allowing models to not just be thrown away when their performance degrades. This increases model lifespan and reduces the necessity to train models from scratch. 

I will consider these key questions, methodologies and ideas to the area of sim-to-real transfer in tactile robotics. In this domain, touch sensors are used tasks such as object pushing or grasping. In order to reduce real life data collection as well as safety, the underlying models are often trained in simulation. The work to bring sustainably focused AI to simulation-to-real tactile robotics already started during my summer project as part of the CDT program. There, I investigated how to reduce the data requirements for transferring simulated models to real life by reusing the knowledge acquired in similar models. This is important because current methods do not make use of existing and available models. Instead, they train a new and costly model from scratch for each task.