Growing appetite for energy
Both science and industry are eager to use solutions using machine learning models, mainly deep neural networks. As we have more and more data, it is possible to create models with more and more parameters (in the case of some networks, such as GPT-3, even trillions); and this means that these models are complex and their operation requires a lot of computing power.
We therefore have complex networks that meet the growing expectations of users, while at the same time requiring complex infrastructure that is expensive to maintain. In addition, for these models to work, huge amounts of energy are needed, which is also becoming more and more expensive.
Due to its specificity, research in the area of sustainable development of AI is particularly important for machine vision; because, for example, in the medical industry, where algorithms support medical staff during operations using robots, the computational efficiency of the algorithms translates directly into the reaction time during the procedure, and thus a reduced risk of complications or shortened recovery time. The issue of machine learning performance is extremely important because models using it are widely used and are an inherent element of many digital economy solutions. However, it is worth being aware that their further development in an unchanged form may soon be slowed down by limited resources.
Recycling resources is the key to efficiency
Currently, the solution to this problem is seen in reducing computation by narrowing available resources or compressing models. However, few researchers pay attention to the fact that models use available resources through so-called recycling.
A sustainable economy has therefore become an inspiration for research – instead of limiting machine learning models, we will try to increase their effectiveness by using the information, resources and calculations to which we already have access. We will focus on reusing what is available: computations performed in previous data processing steps, partial information available when models are used, or knowledge acquired by the model during previous training episodes in the case of continuously trained models. We will look at the problem of effective machine learning from a computational recycling perspective.
Our project focuses on creating models that learn to be efficient rather than just solving a given task. We hypothesize that recycling resources used by machine learning models can significantly increase their performance. We are exploring a new research path, which is machine learning consistent with the zero waste concept, focusing on reducing the resources used by recycling them.
Action – Reaction. What happens when models aren’t fast enough?
For machine learning models to be widely used, they must be efficient. Those that are to improve, for example, the production process, must analyze data and react on an ongoing basis. Models responsible for the operation of autonomous cars also have to work quickly, otherwise they simply become dangerous. Many robots have limited capabilities because the time needed to analyze data and respond is too long.
Our research will have an impact on shortening this time, and thus will open up opportunities for a wider use of models, creating the possibility of commercializing research and development work in various sectors of the economy. We will focus our research on the three pillars of efficient AI: the use of computational recycling, the use of available partial information and the accumulation of knowledge in continually trained models.
Recycling of performed calculations
Conditional computation methods speed up decision-making by adapting the network’s internal processing path based on the input signal. The latest approaches that reduce the inference time of machine learning models focus on shortening decision paths in time-critical applications. They discard information processed at earlier stages, while recycling it can lead to a reduction in the waste of computing resources, which we will address throughout our research.
Using available partial information
Methods based on partial information use additional information available to the network during inference to increase prediction accuracy without retraining the entire model. They can perform computations when partial information about the input data is available. For example, by classifying photos of objects, knowing that the photos were taken on the beach, we can significantly reduce the amount of resources used, for example, to classify the objects visible in them.
Instead of constraining the number of computations or memory used by the models, we focus on reusing what is available to them: computations done in the previous processing steps, partial information accessible at run-time, or knowledge gained by the model during previous training sessions in continually learned models. We look at the research problem of efficient machine learning from the computation recycling perspective and propose methods that build upon our previous works and preliminary results.
Knowledge accumulation in continually learned models
Continual learning methods accumulate knowledge acquired from data arriving as a stream, building on previously acquired skills without forgetting them. They efficiently gather knowledge about previously seen data and reuse it when new data is introduced during training. This efficient knowledge accumulation mechanism encourages us to look at the computational complexity of models from the perspective of continuous learning models and explore the performance gains achieved with this approach. Moreover, the zero-waste machine learning paradigm not only sets a framework for efficient model training, but also includes existing mechanisms, e.g., knowledge accumulation methods in continuously trained models that prevent catastrophic forgetting.
Machine learning can be eco-friendly
What’s new in our research is our focus on creating “green” machine learning models. Better models will require fewer resources to perform calculations and will therefore be environmentally friendly. This assumption is the starting point for our research work. It will allow us to work on models that not only solve specific tasks, but also learn how to work more efficiently. These are solutions that will help create a sustainable economy and will have high implementation potential.