We are living in data-driven societies whereby analytics and learning are key tools for value creation. I am interested in knowing about the trust issues across all the data transformation stages in the AI and machine learning applications, algorithms, and frameworks.
Depends of your desired A.I applications. In general , I think these are the issues :
1-DL model is less explainable . It is good at predicting the outcomes but fails to explain how the outcomes are selected and what are real factors ( features ) behind the model selection
2- in the applications that require to understand inner process of black box of DL to prevent the occurrence of the incident, DL model may just good at telling whatever the incident will happen or not but can not be provide step by step prevention strategies.
3- unfortunately in several application of. A.I or DL models are measured by metric such as accuracy on the test data . However such metrics do not necessarily the model is generalized to real life applications. Even can not measure how we can trust this model for sensitive applications.
4-still most of DL and machine learning models are behaving on just simple statistical or parameters optimization. not yet to have models that logical reasoning like human brain
Depends of your desired A.I applications. In general , I think these are the issues :
1-DL model is less explainable . It is good at predicting the outcomes but fails to explain how the outcomes are selected and what are real factors ( features ) behind the model selection
2- in the applications that require to understand inner process of black box of DL to prevent the occurrence of the incident, DL model may just good at telling whatever the incident will happen or not but can not be provide step by step prevention strategies.
3- unfortunately in several application of. A.I or DL models are measured by metric such as accuracy on the test data . However such metrics do not necessarily the model is generalized to real life applications. Even can not measure how we can trust this model for sensitive applications.
4-still most of DL and machine learning models are behaving on just simple statistical or parameters optimization. not yet to have models that logical reasoning like human brain
Developing Continuous Trust in AI : Once trust is established, it must be nurtured and maintained. This happens through the following:
• Usability and reliability. Performance includes the competence of AI in completing tasks and finishing those tasks in a consistent and reliable manner. The AI application should be designed to operate easily and intuitively. There should be no unexpected downtime or crashes. Usability and reliability contribute to continuous trust.
• Collaboration and communication. Although most AI applications are developed to perform tasks independently, the most likely scenario in the short term is that people will work in partnership with intelligent machines. Whether collaboration and communication can be carried out smoothly and easily will affect continuous trust.
• and etc......
(PDF) Building Trust in Artificial Intelligence, Machine Learning, and Robotics. Available from: https://www.researchgate.net/publication/324006061_Building_Trust_in_Artificial_Intelligence_Machine_Learning_and_Robotics .
Trust on data is an external attribute to the data. It is imposed by someone's subjectiveness and not inherent to the ML, AI or data. Different people will assign different trust levels to AI and machine learning algorithms based on the criticality of the application (Risk will be different for an application dealing with diagnosis of tumors than predicting relationships among individuals, hence different levels of trust on the ML, AI and its data).
Trust can be established based on an assessment of more primitive attributes. The attributes for ML or AI range from performance data to hyperparameter settings. On the data, trust can be established based on criteria such as integrity or the systems confidentiality and availability.
Some examples of data trust parameters are:
- That the data collection process is free from unintentional error
- The data source has not been compromised
- The data source is free from malware (see my work on NN, unsupervised learning Trojans)
- The communications channel to transfer the data is free from tampering
- The machine learning algorithm is free from compromise(see my work on NN, unsupervised learning Trojans)
- That the data will be kept secure from unintentional or intentional dissemination
- The data does not include an intentional discriminatory bias (discriminatory bias here is to be interpreted as discrimination with ethical implications)
- The ML has been trained with optimal parameters (in the case of a user of an application having an ML engine)
- That the data has been properly pre processed
To properly analyse the trustworthiness of the AI or ML algorithm each element of the "supply chain" must be individually assessed if one wishes to "grade" the trustworthiness of the data.
Usually, machine learning based models are expert-based systems meaning selection of all model parameters and the results, are dependent on the expertise of the modeller. However, Data sources, data handling and benchmarking are key to establishing trust in any developed machine learning-based models
- Badges
- Science topic