Weekly Digest Week 36/2020 – dealing with AI uncertainties

Secure AI Weekly admin todaySeptember 9, 2020 23

Background
share close

Uncertainties’ advantage, and AI in the military — learn more from our weekly digest.


Uncertainties presence and influence is greater you might be thinking

Towards Data Science on August 22, 2020

Uncertainties and risks as a form of benefit for an AI trained agent’s performance is quite a realistic scenario, according to this new research.

Uncertainty can be used as implementation into a trained agent’s behavior or be used to optimize this behavior, as shown in the mentioned study, – with, without, or even through uncertainty. Therefore, the more risks are taken into account and implemented during the training – the better and more effective results would be received.

However, developing the technique for predetermined routes and aiming at a long-term perspective are the challenges for the researchers for now.

Is it listening well enough?

Towards Data Science on August 28, 2020

Natural Language Processing or NLP used for hearing voice and speech, interpreting and processing it, is commonly used, still not a perfect tool. And yet it is constantly improving via adversarial training. For example, one of the sources for adversarial examples is TextAttack which provides components for NLP tasks. 

The generated adversarial attack in this particular research was based on semantic and visual similarity of sentences and aimed to train NLP to make it more effective and reduce the number of failures. And consequently, NLP might be used as a tool for phrase replacement or even sentence paraphrasing in literature applied not only to English, but more languages coming soon! 

As a result, applying adversarial examples to NLP improves the outcome of its usage and gives a base for further research of NLP models.

Tricks on military objects are not so fun anymore

Flight Global on September 2, 2020

Adversarial attacks are not just for fun when it comes to military devices risks.

The US Air Force implements AI and puts a focus on its significant part – adversarial training. To make the process of manipulation more complicated for adversaries, AI is provided with many types of data and learns to be robust in different situations in case of attack.  

This is a great example of why AI security is vital and why machine learning and AI training under various conditions is obligatory.

Written by: admin

Rate it
Previous post