Machine learning (ML) consulting, like AI consulting, is an emerging field where both traditional consultants and new startups compete. In this article, we will focus on ML specific challenges, for an overview of AI consulting and the players of the industry, please see our article on the topic.
What is the difference between ML and AI consulting?
Though machine learning (ML) is the subfield of AI with most commercial applications, it is best to distinguish between them
- AI: includes all applications where the computer mimics human intelligence
- ML: applications which use known data to create models that can be used to classify/process new data
Is ML consulting = deep learning consulting?
Not exactly. Deep learning is a subset of ML. However, deep learning is the most successful machine learning technique in terms of accuracy as of 2019 in most areas.
It is not uncommon to see alternative techniques such as decision forests being implemented in the industry instead of deep learning. This is because lack of explanation of outputs is a challenge for deep learning models. There are cases where deep learning models are not deployed in production
- since managers are not comfortable with models that they do not understand and that do not provide explanation for results.
- in cases where auditability is required. For example, labor law prohibits discrimination. Any algorithm that uses criteria which has been used for discrimination before (i.e. gender or race) can not legally make HR decisions without providing rationale that involves reasons other than those criteria. Excluding potentially discriminatory criteria from the model unfortunately does not solve the problem. For example, name, patterns in PTO, pay gap and numerous other data points could be used to indirectly include gender in decision making. Black-box models, no matter how accurate or useful their outputs, cannot be deployed in such situations.
Explaining deep learning is an active area of research called XAI (Explainable AI).
What are the barriers to ML adoption?
As Deloitte highlights, these are the most frequently mentioned barriers according to practitioners:
- Talent shortage: As of 2018 August, there were 150k unfilled data science jobs in the US which Linkedin described as an acute shortage in large US cities
- Immaturity of ML infrastructure and processes: ML is a new programming paradigm, one that derives rules from data rather than programmer input. It took us tens of years to come up with Scrum, the agile programming approach, currently widely used by most teams. Similarly, it will take time for ML processes and frameworks to reach maturity. TensorFlow, one of the most widely used ML frameworks was only published at the end of 2015.
- Most ML techniques are data hungry: Accurately labeled training data is time consuming and expensive to generate. ML practitioners need to be creative in leveraging public data or get the necessary data labelled. This is why numerous data labeling companies were founded since 2010s. Another solution to this is one shot learning and other less data hungry approaches, however this is an area of ongoing research.
- Deep learning is not explainable. As discussed, this is hindering progress and XAI attempts to address that.
What is the future of machine learning consulting?
ML consulting will grow by tackling the identified issues:
Expanding the talent pool: Most consultancies are analyzing their workforce in detail to identify those capable of data science. A background in programming, statistics or math tends to be sufficient for individuals to work as data scientists after a relatively quick training.
Improving ML infrastructure and processes: As ML matures as a programming paradigm, better processes, improved computing resources (i.e. GPUs and AI chips) and more automation will make ML faster and easier.
Getting creative with data: Advances in Natural Language Processing (NLP) were thanks to the wide availability of translated government documents in Canada and Europe. While finding data is a relatively straightforward solution, areas of AI research such as transfer learning or data synthesis could be more technical solutions.
There are also advances expected in explainable AI which would increase trust in ML systems and enable their more widespread adoption.
Finally local machine learning applications are likely to make IoT applications smarter and faster by pushing decision making to edge devices.
What are typical ML consulting activities?
Understanding the business needs
As in all consulting, everything starts with the business need. Whether it is predicting where to install telecom base stations or whom to show ads to, misunderstanding business requirements is still one of the major reasons for lack of success of consulting and software projects. ML consulting, at the intersection of consulting and software, is especially prone to this issue.
Setting up the team and process
Not all problems need machine learning. Machine learning and other heuristic approaches make sense in problems that can not be reduced to a set of rules. If the rules are well known and simple, rule based systems outperform machine learning and are simpler to maintain.
If ML is a good fit for a problem, project team, stakeholders and high level targets need to be outlined.
Data Collection & Exploration
If company has the data, this is a relatively easy step. Consultants need to work with business to validate that the data is correctly labeled and not self-contradictory.
If data is not available, the techniques outlined above such as leveraging online data, paying for data labeling as well as novel ML approaches such as one shot learning would need to be considered.
Thousands of experiments are necessary to develop a high performing machine learning model. This is an iterative process, taking into account latest research, understanding business dynamics and data exploration.
Ultimately, all models are evaluated against the same set of test data to assess their accuracy.
Full-stack application development
Taking a model to production requires additional software development and integration work.
Most of the time, ML models are encapsulated in APIs which are easy to integrate with any application. The development of the application which will operationalize the ML model and make it part of the decision making process can be harder than building the model. Application development may require integration to existing enterprise systems which requires working with external developers.
Scalability and data security issues also need to be addressed as part of operationalizing the model.
Feel free to read our AI consulting article to get a higher level view of consulting in the age of AI, including top companies in the field.
How useful was this post?
Click on a star to rate it!
Average rating / 5. Vote count:
What did you like about this post?
Appreciate it if you leave your name and surname so we can publish your review as a testimonial
Thank you for your feedback!