Hello all,
From experience there are several different design roles in AI/ML design that are important to separate.
I thought it might be useful to document the differences - not least to help clarify the language and terminology which is in danger of following the everything is an algorithm. Or worse, that someone does something that however peripherally involves AI or ML and then claims to be an algorithm designer.
If I've made any errors or missed anything out please say:
1. Algorithm designer: This kind of designer focuses only on the design of mathematical algorithms (equations and functions) that are used in AI/ML processes. The work of algorithm designers is seen in mathematical papers such as:
https://www.semanticscholar.org/paper/Algorithm-837%3A-AMD%2C-an-approximate-minimum-degree-Amestoy-Enseeiht-Irit/00aaa6b18f94cbc9bf95e9df9313dab436834ea5
https://link.springer.com/article/10.1007/BF01587094
https://www.sciencedirect.com/science/article/pii/0098300484900207
https://www.nowpublishers.com/article/Details/MAL-016
If as a designer you are not in that mathematical game then you are not an Algorithm Designer. Ditto, design research in algorithm design is the design research only in relation to those kinds of mathematical developments in creating mathematical algorithms .
2. AI/ML process designer: This area of design is the design of generic processes and methods that USE algorithms as the basis for those processes. Examples are the design of generic processes that will use algorithms designed by others to e.g. semantically analyse images, create ontologies of texts and academic fields,. This includes the development of e.g. OWL, neural net processes, genetic processes. It involves AI/ML process designers in having enough mathematical knowledge and knowledge of bias The kinds of papers you would see for this area of design include:
https://link.springer.com/chapter/10.1007/978-3-540-28650-9_4
https://www.nowpublishers.com/article/Details/MAL-044
https://link.springer.com/article/10.1023/A:1007677805582
This work is almost always the province of those designers that sit at the joint boundary of mathematics and computer science - the back office boffins at Google, Facebook and the like. A classic small-scale example is Optika Solutions, a firm we are collaborating with who do wonderful AI/ML process design work and are developing a great AI/ML process tool called Acumen (http://optika.com.au/ and https://akumen.io/ )
3. AI/ML data management designer: The role of an AI/ML data management designer involves the defining, acquiring, structuring, cleaning of data to be used in AI and ML, both for creating training data sets to train AI/ML models and for processing using trained AI/ML models. In practice, the AI/ML data management designer role comprises a good understanding of data structures, objects and categories along with deep understanding of problems of bias and often involves a significant depth in statistics expertise. More recently, there have emerged platforms and software to support exploring and pre-processing data for use in AI/ML models. (A significant part of Facebook and Google's back end work is doing exactly this. T
The role of designing the best strategies for data management are typically crucial to the success of any AI/ML process. A key aspect of this role is understanding how choices in data, errors or biases in data can result in erroneous or biased outcomes. In some cases, this can lead to fatal outcomes. More information on this designer role is at:
https://www.cognilytica.com/2019/03/06/report-data-engineering-preparation-and-labeling-for-ai-2019/
https://www.altexsoft.com/blog/datascience/preparing-your-dataset-for-machine-learning-8-basic-techniques-that-make-your-data-better/
https://cloud.google.com/ml-engine/docs/tensorflow/data-prep
4. AI/ML product designer: This is an area of design allied to conventional design activities. It requires AI/ML product designers to have enough high-level mathematical skills to understand the overall AI/ML processes and data problems but the focus of the design work is on producing products. Three common classes of AI/ML-based products that AI/ML product designers create are:
1. Products that use AI/ML (e.g speech recognition, oil/gas exploration, chat bots, automated share trading, cyber-security SIEM products, automated advertising, political election influencing... )
2. AI/ML products that are tools for designers to use in creating other AI/ML products (e.g. Amazon's Greengrass, Microsoft's Machine Learning Studio, Sagemaker
3. Software products such as PyTorch, TensorFlow, i360 ...
4 AI/ML-based services. Sometimes these AI/ML-based services emerge as a product (e.g. Google, Facebook, Netflix...) other-times not, e.g. government anti-cheating services aimed at recovering money from social security recipients and tax-payers.
Key skills needed by AI/ML services designers (alongside the usual product design skills) include: a) sufficient mathematical skills to understand the essential aspects of the AI/ML processes (not algorithms); b) understanding of how bias and error and misdirection result from data and AI/ML processes and modes of software coding; c) skills in using the necessary AI/ML design software; d) systems knowledge especially of positive and negative feedback systems (AI/ML processes are themselves essentially feedback systems); e) sound understanding of counter-intuitive systems theories and that literature.
The latter, the understanding of counter-intuitive systems theories, is perhaps the most important. Any design for a service complex enough to benefit from AI/ML will almost ALWAYS result in counter-intuitive outcomes.
The issues of AI/ML design are not new. The same problems were being found and tackled in the 70s - its just been forgotten.
There is a large well-developed literature on counter-intuitive issues in complex systems design that applies to AI/ML services and product design.
John Flach's book and blog are good starting points as is the work of Jay Forrester. Tim Haslett gives an ordinary everyday example of easy counter--intuitive misunderstanding that givs a pointer to how much more sophisticatedly skilled in this area a designer of AI/ML-based services needs to be.
John Flach Book:
https://corescholar.libraries.wright.edu/books/127/
John Flach Blog
https://blogs.wright.edu/learn/johnflach/author/w001jmf/
Jay Forrester
http://static.clexchange.org/ftp/documents/system-dynamics/SD1993-01CounterintuitiveBe.pdf
Tim Haslett
https://timothyrhaslett.wordpress.com/2019/06/07/there-are-some-counterintuitive-effects-when-the-australian-federal-police-seize-1-3b-worth-of-drugs/
Feedback welcomed!
Best wishes,
Terry
-----------------------------------------------------------------
PhD-Design mailing list <[log in to unmask]>
Discussion of PhD studies and related research in Design
Subscribe or Unsubscribe at https://www.jiscmail.ac.uk/phd-design
-----------------------------------------------------------------
|
