The 7 malignant myths of machine learning

Unfortunately, we are in the midst of a pandemic of ML misconceptions and illiteracy

gears orange large efficient automated machine learning automation
Thinkstock

Enterprises are wasting sweat and treasure on projects that are risky, expensive, and rarely make it into production

The biggest obstacle to unleashing the power of enterprise artificial intelligence (AI) and machine learning (ML) isn’t a lack of data scientistsit’s a lack of leaders who understand ML.

 Tech titans, industrial giants, and startups alike are falling over themselves to hire leaders with the business acumen and ML know-how to spot the most valuable opportunities and guide projects from inception to business value. 

Unfortunately, we are in the midst of a pandemic of ML misconceptions and illiteracy.

Executives regularly make statements that, at best, belie an ignorance of ML and, at worst, are categorically false. 

Faux pas like “ML automatically gets better over time”, or “it’s reinforcement learning because users provide feedback,” abound. 

Sadly, this ignorance has dire consequences. Enterprises are wasting sweat and treasure on projects that are risky, expensive, and rarely make it into production, because they are a poor fit for ML or because the projects are badly managed. 

All too frequently they turn into zombie projects destined to haunt all future ML initiatives. 

business zombie thinkstock Thinkstock

...All too frequently they turn into zombie projects destined to haunt all future ML initiatives

All the while they are missing the abundant opportunities to improve decision making, drive efficiency, and deliver new customer experiences with ML that litter every part of the enterprise. 

As they never look like something out of an AI commercial, they go unnoticed. 

Get ML literate by busting these malignant myths 

The good news is that you don’t need to become a data scientist to become a ML literate leader. Indeed, the skills you need - to identify the most valuable ML use cases, guide ML projects, avoid the many pitfalls, and manage diverse teams of data professionals – are rarely, if ever, taught in ML courses. 

These are skills that you get through experience, by initiating and participating in ML projects, and by spending quality time with your data scientists who will happily teach you about ML in exchange for nuggets of business wisdom. 

But first, you need to unlearn the destructive ML myths that most commonly lead business leaders astray. 

Myth 1: ML Is about computers learning to think like humans

Reality: ML is powerful because it does not learn to think like humans. ML is better than humans at analysing data, especially at scale and across multiple sources and at identifying complex patterns and outliers. In contrast, ML struggles with limited or noisy data, uncertainty, logical reasoning, and interacting with humans, which most humans have an easier time coping with. 

In short, use ML to tackle the data-rich problems that humans struggle with, and be wary of using ML to tackle anything that people do really well. 

Use machine learning to tackle the data-rich problems that humans struggle with, and be wary of using it to tackle anything that people do really well

Myth 2: ML is all about predicting the future

Reality: Use ML for more than predictions and avoid predicting a changing future. You should use ML for its full range of use cases: to generate business insights and add new application features in addition to predicting outcomes and forecasting. However, since ML will always be trained on historical data, it will struggle to predict the future when the future is expected to look very different from the past — e.g., predicting sales of a new product in a new country or the next financial crisis.

Myth 3: ML automatically gets better over time

Reality: Most ML models become less accurate. You can develop better ML models as you collect additional training data, but that accuracy will plateau. When your model is deployed, it is usually frozen in time and will inevitably degrade as the world around it changes. Mitigate this by monitoring the incoming data and the performance of your models, and develop processes for retraining them on a regular basis.

Myth 4: ML is about delivering higher accuracy

Reality: Forget accuracy, it’s about business impact. A 50.1 per cent accurate model that tells you when to take an additional card in blackjack could make you fabulously wealthy while a model that makes a 99 per cent accurate prediction about whether you will get sued is likely useless (unless you get sued extremely frequently). 

No amount of ML can turn missing, incomplete, or wrong data into valuable insights

Instead, benchmark your model against the business impact it has relative to your current decision-making processes and other candidate ML models.

Myth 5: ML Is objective

Reality: ML learns whatever biases you and your data have. It is really easy to create an unfair ML model — just train one on past human decisions, or include sensitive data (or data that is correlated with that sensitive data) in your model. Or train it with data that is not representative. However, by avoiding these obvious missteps, validating and testing your model, you can dramatically reduce your risk of being discriminatory — and certainly make your models far less “biased” than most humans. 

forrester graph on bias Forrester

Myth 6: ML is a black box

Reality: ML is far more transparent than people, but that’s not the real problem. Most ML models are fully transparent, but many are hard to explain. However, this is really about trust. Yes, you can use ML methods that are easier to explain or model-agnostic techniques that provide explanations for each prediction but you should use them in tandem with traditional change management. That is, build trust by involving and communicating with stakeholders throughout the process and translate results into the outcomes that make sense to them. 

Myth 7: ML can do anything with massive amounts of data

Reality: No data is inherently valuable — it is only “good” data with respect to a particular problem and the ML methods you are using to solve it. No amount of data will help if it’s irrelevant to the phenomenon you are looking to analyse and no amount of ML can turn missing, incomplete, or wrong data into valuable insights. The key is to iterate your business use cases, ML methods and data in tandem.

kjell carlsson Forrester

Kjell Carlsson, Ph.D., is a senior analyst at Forrester

Copyright © 2020 IDG Communications, Inc.

7 secrets of successful remote IT teams