Cloud Hosting

As machine learning progresses at breakneck speed, its intersection with sustainability is ... [+] increasingly crucial.

Irene Unceta is a professor and director of the Esade Double Degree in Business Administration & AI For Business

As machine learning progresses at breakneck speed, its intersection with sustainability is increasingly crucial. While it is clear that machine learning models will alter our lifestyles, work environments, and interactions with the world, the question of how they will impact sustainability cannot be ignored.

To understand how machine learning can contribute to creating a better, greener, more equitable world, it is crucial to assess its impact on the three pillars of sustainability: the social, the economic, and the environmental.

The social dimension

From a social standpoint, the sustainability of machine learning depends on its potential to have a positive impact on society.

Machine learning models have shown promise in this regard, for example, by helping healthcare organizations provide more accurate medical diagnoses, conduct high-precision surgeries, or design personalized treatment plans. Similarly, systems dedicated to analyzing and predicting patterns in data can potentially transform public policy, so long as they contribute to a fairer redistribution of wealth and increased social cohesion.

However, ensuring a sustainable deployment of this technology in the social dimension requires addressing challenges related to the emergence of bias and discrimination, as well as the effects of opacity.

Machine learning models trained on biased data can perpetuate and even amplify existing inequalities, leading to unfair and discriminatory outcomes. A controversial study conducted by researchers at MIT showed, for example, that commercial facial recognition software is less accurate for people with darker skin tones, especially darker women, reinforcing historical racial and gender biases.

Moreover, large, intricate models based on complex architectures, such as those of deep learning, can be opaque and difficult to understand. This lack of transparency can have a two-fold effect. On the one hand, it can lead to mistrust and lack of adoption. On the other, it conflicts with the principle of autonomy, which refers to the basic human right to be well-informed in order to make free decisions.

To promote machine learning sustainability in the social dimension, it is essential to prioritize the development of models that can be understood and that provide insights into their decision-making process. Knowing what these systems learn, however, is only the first step. To ensure fair outcomes for all members of society, regardless of background or socioeconomic status, diverse groups must be involved in these systems design and development and their ethical principles must be made explicit. Machine learning models today might not be capable of moral thinking, as Noam Chomsky recently highlighted, but their programmers should not be exempt from this obligation.

The economic dimension

Nor should the focus be solely on the social dimension. Machine learning will only be sustainable for as long as its benefits outweigh its costs from an economic perspective, too.

Machine learning models can help reduce costs, improve efficiency, and create new business opportunities. Among other things, they can be used to optimize supply chains, automate repetitive tasks in manufacturing, and provide insights into customer behavior and market trends.

Even so, the design and deployment of machine learning can be very expensive, requiring significant investments in data, hardware, and personnel. Models require extensive resources, in terms of both hardware and manpower, to develop and maintain. This makes them less accessible to small businesses and developing economies, limiting their potential impact and perpetuating economic inequality.

Addressing these issues will require evaluating the costs and benefits carefully, considering both short- and long-term costs, and balancing the trade-offs between accuracy, scalability, and cost.

But not only that. The proliferation of this technology will also have a substantial impact on the workforce. Increasing reliance on machine learning will lead to job loss in many sectors in the coming years. Efforts must be made to create new job opportunities and to ensure that workers have the necessary skills and training to transition to these new roles.

To achieve economic sustainability in machine learning, systems should be designed to augment, rather than replace, human capabilities.

The environmental dimension

Finally, machine learning has the potential to play a significant role in mitigating the impact of human activities on the environment. Unless properly designed, however, it may turn out to be a double-edged sword.

Training and running industrial machine learning models requires significant computing resources. These include large data centers and powerful GPUs, which consume a great deal of energy, as well as the production and disposal of hardware and electronic components that contribute to greenhouse gas emissions.

In 2018, DeepMind released AlphaStar, a multi-agent reinforcement-learning-based system that produced unprecedented results playing StarCraft II. While the model itself can be run on an average desktop PC, its training required the use of 16 TPUs for each of its 600 agents, running in parallel for more than 2 weeks. This raises the question of whether and to what extent these costs are justified.

To ensure environmental sustainability we should question the pertinence of training and deploying industrial machine learning applications. Decreasing their carbon footprint will require promoting more energy-efficient hardware, such as specialized chips and low-power processors, as well as dedicating efforts to developing greener algorithms that optimize energy consumption by using less data, fewer parameters, and more efficient training methods.

Machine learning may yet contribute to building a more sustainable world, but this will require a comprehensive approach that considers the complex trade-offs of developing inclusive, equitable, cost-effective, trustworthy models that have a low technical debt and do minimal environmental harm. Promoting social, economic, and environmental sustainability in machine learning models is essential to ensure that these systems support the needs of society, while minimizing any negative consequences in the long term.

Go here to see the original:
We Need To Make Machine Learning Sustainable. Here's How - Forbes