IBIS researchers offered a 'fact sheet' for AI transparency

About:

IBM Research published a paper that demands a supplier's announcement to suit the AI ​​services. This announcement will include information on performance, safety and security. Paradise life is using AI to increase life insurance policy for those who will not be traditionally eligible, such as chronic diseases and non-U.S. With the citizens.

IBIS RESEARCHERS OFFERED A 'FACT SHEET

Description:

Google Driving Car Spinoffs Wemo is tapping to provide mobility to the elderly and people with disabilities. But despite good AI, it is clearly capable, doubt is dependent on its security, transparency and bias. In other industries, these documents are present, and although they are voluntary in many cases, these attempts often become standard. Energy Star or the U.S. Think of Bond Rating in Consumer Product Safety Commission or Financial Industry.

AI Research and AI Foundation's key AI Mozambique today said that AI services should be "created, tested, trained, deployed and evaluated", there has not been any compromise, IB Research and AI Science for Social Good Program. blog post. In theory, these documents will also enable more liquid AI service markets and bridge information intervals between consumers and suppliers. IBM Research said that SDOC should be voluntary.

Mojsilovic and colleagues are formally declared as supporters of the declaration of voluntary facts, which will be completed and published by those companies who develop and provide AI with the aim of increasing their transparency and enhancing their services. . Mojsilovic thinks such factsheets can give competitive advantage to companies in the market, such as how to get energy equivalent energy products for energy companies for energy efficiency.

Many core pillars make the basis for faith in AI systems, Mojsilovic explained: fairness, robustness, and interpretation. The fair AI system can be reliably trusted that there is no biased algorithm or dataset, or contributes to inappropriate treatment of some groups. If an AI system is fair but can not oppose the attack, then it will not be believed. If it is safe but we can not understand its output, then it will not be believed. To make AI systems that are really trustworthy, we need to strengthen all the columns together.

AI services will address such questions as:

1.) What is the expected behavior when distracted by data distribution distribution distribution?

2.) Who is the target user of the explanation?

3.) What was investigated for dataset and model prejudices?

4.) Use data is maintained / kept / maintained with service operation?

5.) What was a bias on the dataset?

6.) Was the service checked for opposition against the adverse attacks?

7.) Is there a datasheet or data statement in the dataset used to train the service?

8.) Was the service tested on any additional datasets? Do they have datasheets or data statements? If so, describe the test method.

9.) Can the algorithm be interpreted / interpreted as output? If so, explain why interpretation is achieved?

10.) Was the service checked for opposition against the adverse attacks?

11.) What type of governance is employed to track the overall workflow of AI service data?

Understanding and evaluating AI systems is a very important issue for the AI ​​community, it is an issue that we believe industry, academic and AI practitioners should work together.