In the last COP25 Climate Summit held in Madrid. Many subjects were being discussed on the matter of a possible climate crisis, and how to face it.
Has Machine Learning (ML) and Natural Language Processing (NLP) something to say about it? Surprisingly, yes, it does!
It seems obvious, but computers need the energy to work. There are more and more computers every day, and their energy needs are also higher.
In the past, the computing power needed to train state-of-the-art AI systems nearly doubled every two years (as we learned from this article).
Yet, the trend has been skyrocketing since 2012: currently, this requirement doubles in just 3.4 months (not 2 years anymore!). This graph is self-explanatory.
What does this mean? Even if computers are now more efficient than ever, if the computing power needed doubles every 3 months, the energy required will also be higher and higher.
AI and ML are severely affecting power requirements in the world. Needless to say that this fact is not good for the climate —nor for the economy of the companies that want to use such tools, of course—.
Can something be done? Yes, not relying so much on algorithms, but rather on data. The goal of these new ML algorithms is to work even in absence of good training data.
How does this solution work?
Simply by having machines create correct and realistic quality training data by itself, so that your ML algorithms won’t need so much computing power to be effective. On top of it all, they will be even cheaper for you!
Why is synthetic data important?
Developers need large, carefully labeled data sets to train neural networks. More diverse training data generally makes AI models more accurate.
The problem is that collecting and labeling data sets that can contain anywhere from a few thousand to tens of millions of items is time consuming and often prohibitively expensive.
So for cost savings and Since synthetic datasets are self-labeled and may deliberately include rare but crucial corner cases, it's sometimes better than real-world data. What's more:
- Mostly AI claims that synthetic data can retain 99% of the information and value of the original dataset while protecting sensitive data from re-identification. (Mostly AI)
- "The trend is going towards automating data generation. As NLG (Natural Language Technology) develops, synthetic text is becoming a solid alternative for question/answer systems, for the generation and labeling of textual data". claims Antonio Valderrabanos, CEO of Bitext
- When training data is highly imbalanced (e.g. more than 99% of instances belong to one class) synthetic data generation is necessary to build accurate machine learning models. (Tensorflow)
- With Synthetic Data you are guaranteed to be 100% free of privacy issues. Since data is created from scratch, there is no need to worry about PII or GDPR issues.