Practical Machine Learning Tips and Tricks to Achieve Success Quicker

September 28, 2018

Raise your hand if you’re tired of reading and hearing how AI will solve the world’s problems. Give us a shout out if you’ve been led into believing that AI is going to displace most people from their jobs soon enough. We’re exhausted too, particularly considering how AI essentially is a 60+-year-old tech concept! Machine […]

Raise your hand if you’re tired of reading and hearing how AI will solve the world’s problems. Give us a shout out if you’ve been led into believing that AI is going to displace most people from their jobs soon enough. We’re exhausted too, particularly considering how AI essentially is a 60+-year-old tech concept!

Machine Learning
Machine Learning

What’s the cause of the hype overdrive?

Experts in the ‘real no-nonsense AI’ space say that it’s the exponential progress in the effectiveness of machine learning algorithms that’s caused the AI buzz. This, however, is the phase where the buzz will deplete (because it’s already peaked) and focus will move to real-world applications of AI at a scale. If your organization’s AI strategy is not checked out yet, or if you’re looking for a course correction, here are some practical tips to make machine learning the superstar of your present and future.

‘Good’ Shouldn’t Be Beaten Down By the Pursuit of ‘Great’

We strongly recommend you make a mental note to go through machine learning success case studies. Many of them will clearly showcase how machine learning success often largely depends on data quality, data scientists’ ability to understand the business use case, and a bit of luck.

Often enough, 60-70% of the desired functionality can be served by a ‘good enough’ machine learning algorithm. The journey from ‘good’ to ‘great’ is long, expensive, ambiguous, and tough. Hence, the additional effort must be backed by a business justification (a 5% improvement in performance is invaluable for an algorithm that detects tumours from medical scans, but isn’t half as useful for an algorithm that predicts songs that a listener might like).

Business First, Algorithm Second. Or the other way around?

Considering a team of machine learning experts in charge of an email marketing personalization algorithm. Consider another team in charge of an algorithm that makes an unmanned areal vehicle navigate safely in rainy conditions. Now, assuming both teams are able to build a minimum viable algorithm that delivers the core expected utility, where should the incremental effort go.

In the first case, algorithmic fine-tuning might not return half the value that the engineers would achieve by optimizing the data and logistics related to the process. In the second case, the algorithm comes first, because the cost of failure is massive. So, ML experts need to be able to understand their development in the context of the business use case and the logistics, for channeling their resource in the right direction.

It’s The Data, Not the Algorithm

Got resounding success for an ML algorithm? Scored a memorable failure with another algorithm? Chances are, it’s the data that’s driving results, and not the algorithm.

Several promising ML developments never saw the light of a day in production, because the developers’ ego made the algorithm undergo dozens of iterations, each making it more complicated and twisted. If you anticipate such a situation within your organization’s machine learning teams, take control, and begin by questioning the data, and not the algorithm.

An ordinary ML algorithm can deliver good results if it’s learning data is robust. An extraordinary algorithm, however, will deliver garbage if it’s learning data is not good.

Machine Learning Human

Add the Human Element

Let’s face it; you’ll be enviably lucky to have machine learning algorithm developers and data scientists with domain and business experience. To make sure that your organization’s machine learning projects don’t end up being mere laboratory successes, you need people who are masters of business processes, understand data pipelines, and appreciate the basics of machine learning. This is how you build a team that knows whether it’s necessary to create new data capture processes, eliminate expensive and laborious features that business folks won’t use, and how to evaluate the progress of the development in the context of the use case.

Feature engineering experts are a key element of any ML team. Though the feature selection process can be automated to a great extent, it’s the availability of insightful human oversight that helps ML teams create groundbreaking algorithms. This is equally true for all decisions related to data selection and processing. In fact, ML experts believe that the ability of a team to prove or disprove the effectiveness of an idea is a key differentiator for machine learning success. You need an insightful human presence for that.

Be Adroit With Tools

The average machine learning project will require developers to use anything between 5 to 8 tools, invariably. That’s natural, considering the wide range of niche tools available for the machine learning community.

For starters, subscription-based ML tools such as Amazon Machine Learning and Microsoft Azure Machine Learning Studio will help you scale up quickly. BigML is another platform slowly gaining traction. Then, there are the Apache Software Foundation (ASF) projects like Singa and Mahout. Open-source frameworks for machine learning, such as H2O, TensorFlow, and Shogun, of course, are already prominent in the market.

The point is - the speed of innovation in the machine learning market will require ML teams to be up and awake, and comfortable with a wide range of ML tools. Niche ML tools are mastering their art quickly, which means developers need to show agility in moving from tool to tool to realize businesses use cases.

Measure Small Successes

Here’s the hard truth. Soon enough, the pomp and show around ML powered applications will die down, and stakeholders will clamour for proof of business value. A great method to prepare yourself for such times is to always measure the success of your ML projects, aim at quick delivery of several small improvements, and to make the small improvements work in tandem to deliver significant business advantages.

Concluding Remarks

Machine learning is already solving business problems, enabling people to do their jobs better, and improving business processes exponentially. A practical approach to translating data and algorithm into valuable insights is what will ensure your organization’s machine learning projects’ success.

 

 

Author : Rahul Sharma

By Team FileCloud