Machine Learning in Production
-Ethical Issues in Machine Learning
Machine learning algorithms affect daily human lives. Ethical issues are experienced in probabilistic programs, adversarial, reinforcement and federated learning. Tech giants like Amazon, Google, Facebook, Microsoft, and IBM believe that it’s the appropriate time to talk about the boundless landscape of AI. Typical algorithmic evaluation approaches used in predicting human the results like bail decisions, recruitment, mortgage approvals and insurance premiums are presently being trailed and subsequently deployed into production. There’s need for implementation of ML models that adhere to ethical and legal requirements of the society. European Union’s General Data Protection Regulation (GDPR) and United Sates’ Fair Credit Reporting Act prescribe that data ought to be processed in a fair and unbiased manner. Further, GDPR alludes for individual right to receive explanation regarding decision made by an automated system. There’s need to build accountable, fair, transparent and trustworthy can be developed in future AI and ML systems to ensure continued confidence of the people in the automated systems.
Some of the ethical concerns include
1. Unemployment might come to end when automation comes
2. Inequality the unequal distribution of wealth created by AI systems
3. Humanity interaction and behavior of bots like that presented by Eugene Goostman won the Turing Challenge
4. Racist robots influence AI bias
5. Evil genies depicted in AI systems
6. Singularity when controlling the complex intelligent systems
7. Robot rights in defining human treatment with AI among neuroscientists
-Best practices in Machine Learning
As a high level of software testing in units, Unit Testing is the best practice in building AI systems. It aims at validating individual unit’s performance. Unit testing has few inputs and single output. There are industrial frameworks, drivers, mock and stubs used in unit testing.
Goals of Unit Testing
- To ensure that the model works well
2. To assess if the model performs well under known best hyper-parameters
Best practices of Unit Testing
1. Seed all test singly
2. Test all functions of the model
3. Use pure functions unlike objects
-General Data Protection Regulation (GDPR)
The European Union developed and enacted the GDPR on 25th May 2018 which covers many sections regulating the handling of personal data and the need to comply with the law. Data scientist need to understand the rules of data privacy to ensure proper data control and processing. Since data scientists handle data from different industries, it is encouraged that they all understand the Data Protection Act and GDPR legislations. GDPR includes multiple principles of Data Protection, which are ideal for data scientists. Such knowledge can help in getting comprehensive guide towards GDPR compliance for mid and large businesses. The regulation also equips data handlers with knowledge of personal data breach policy and understand the economic impact of Brexit. Data scientists ought to understand the consent, contract and legal obligation. Knowledgeable data scientists ought to understand the best practices to implement changes introduced by the GDPR legislation. Homomorphic encryption was recommended to enhance security in ML models. The encryption allows recipients of the encrypted data to encrypt the results of computations without knowing the inputs. More information can be read here.
Failure rate of ML Models
During the Meetup, machine learning models were discovered to fail to have a return on investment (ROI) after deployment. The argument was supported by audacious predictions by industry research firm Gartner whereby 85% of AI projects fail to deliver for CIOs. Such results, 85 percentage means that out of 20 AI projects in the industry 3 succeed in production and the rest 17 fall. Gartner encourages to concentrate on scalability to ensure business growth. Experts estimated that cognitive and AI systems’ revenue could reach $12.5 billion. The difference between revenue prediction and the actual use of AI. Tech giants like Apple, and Google are intensively researching and reinventing the entire slew of AI practices and technologies. The performance of AI models is defined by the nature of technology and the adoption pace in the industry.