2021 - The Machine Learning Manifesto - Jean Voigt

Texto

After months of insightful discussion, feedback and research I venture and propose a list of 42 principles for more holistic machine learning initiatives.

Data Engineers, Agile Practitioners, Machine Learning Operators and Engineers, Data Scientists, Privacy Lawyers and all people of the world affected by the radical economic and work place transformation that is revolutionising our thinking resist the rise of data myopia. Hence we declare:

1.Design and implement models in accordance to as well as stay informed on the latest thinking regarding the ethical use of artificial intelligence

2.Encourage and foster direct discussions between model designer and model user

3.Reduce intermediaries and hierarchies so that information can flow freely and with minimal distortions to the modeling teams

4.Explore rigorously and communicate the model evaluation metrics transparently and not rely on beneficial metrics to advertise success

5.Champion to avoid bias in data, opinion, and design through balanced staffing and data collection strategies

6.Exercise great care in the preparation and collection of both label and feature data

7.Communicate openly, transparently, and repeatedly the model mechanics so that anyone who wishes to understand the logic will be able to do so

8.For every model deployed, establish a model operations plan to monitor performance and data consistency, quality, completeness, and structural changes

9.Use documentation effectively to facilitate communication of model assumptions, input, key algorithm design aspects, and mechanics

10.Strive to collect sufficiently accurate, consistent, clean, and definitive labels and maintain a label update process

11.Establish a positive error culture in order to fail early and often and lead by example by admitting to mistakes and staying open to critique

12.Discourage building extensive central ivory tower data science teams

13.Recognize that excellent data management and especially documentation and quality of data are essential for trusted data inputs driving models that people feel confident of using

14.Encourage and help to establish AI acceptance before deployment rather than relying on passive tolerance or rejection, especially for models that rely on feedback loops

15.Write data input quality checks and integrate this step in the model design process and as far as possible into model evaluation

16.Work towards reduction of bias against AI results and include AI as a dimension of diversity and inclusion for organizations

17.Establish cross-functional teams that involve as many people as possible using the model results in the engineering process to reduce fear and rejection

18.Provide tangible perspectives of model operations, maintenance, and parameter improvement to people whose present work will change because of AI deployments

19.Listen to advice from model and data engineers that have been employed irrespective of their social, ethical, professional, or corporate position

20.lways stay informed and attuned to technical aspects of the work

21.Support the industry in the creation of appropriate design, operations, control, and monitoring procedures and frameworks

22.Recognize that artificial intelligence is affecting the core of what defines human nature and thus the special responsibility for care to affected parties

23.Strive to avoid cross-functional team specialization and teach the essence of data and model engineering so that every team member can contribute meaningfully

24.Prevent that unrealistic expectations for AI/ML model speed and accuracy are established

25.Do not try to break the laws of physics or mathematics through the use of machine learning

26.Complement AI/ML with appropriate statistical analysis both pre- and post-modeling; it’s not a competition but a symbiotic relationship

27.Allow models to learn and improve over time and recognize that the learning process may require potentially extensive human activity observation before models perform in the desired fashion, especially for such models with a feedback loop/human-in-the-loop/collaborative filtering

28.Evaluate models against data issues before deployment

29.Celebrate the engineers as the heroes of data, understanding their critical role and heed their recommendations and advice

30.Celebrate the non-technical professions in the data field as the heroes in their domains

31.Never play engineers and non-technical professions against each other but recognize that good models need collaboration between several disciplines

32.Document and keep updated the details of data elements such as attributes, code meaning, time and source as well as legal permissions always as close to the data record so that it is evident during data exploration

33.Create nimble machine learning structures with many happy AI-enabled teams across the organization

34.Establish recurring procedures to keep data, models, and the environment they operate in well-aligned

35.Design AI/ML models in such a way that they can be asked to unlearn and forget if this should become necessary

36.Keep track of model, data, and configuration versions used for training and deployment

37.Establish the legally permissible use of each data element at the point of data collection and, in the event of secondary data, get a legal opinion on how the input data can be used

38.Stay curious, learn and accept the fact that experiences are biased by the events that people have witnessed by chance during the course of their life

39.Leverage existing role definitions from agile and technical domains rather than creating firms specific niche roles that make it harder for everyone to navigate the many responsibilities in the field of AI

40..Help your leadership team to establish serving leader incentives

41.Build a business case and revenue/cost model for your initiative

42.Do not solve mathematical problems using machine learning models

Enlaces

URL: https://medium.com/unmanage/the-machine-learning-manifesto-b26804e921eb

Wayback Machine: http://web.archive.org/web/20220509003334/https://medium.com/unmanage/the-machine-learning-manifesto-b26804e921eb