Veritone’s Aaron Edell explains how data scientists can best harness artificial intelligence and machine learning.
Data science, artificial intelligence (AI) and machine learning (ML) are all massive areas that are undergoing growth in the tech industry and attracting increasing amounts of attention. But what about the jobs of the future that will combine all three?
Siliconrepublic.com spoke to Aaron Edell, director of applied AI at US software company Veritone, to learn how data scientists can harness AI and machine learning technologies for better outcomes.
Although there are plenty of possibilities to consider in this area, his main advice is to always maintain sight of the problem that needs to be tackled and keep the customer in mind.
Keep sight of the problem
With machine learning, business process scalability has made leaps and bounds, but it’s important not to get side-tracked by that, according to Edell. Instead, focus on the things that are going wrong, rather than attempting to improve the things that are already working.
“The most common mistake really anyone can make when building an ML solution is to lose sight of the problem they are trying to solve,” he said.
“As such, we can spend a lot of time making the tech better, but forgetting why we’re using the tech in the first place.
“For example, we may spend a lot of time and money improving the accuracy of a face recognition engine from 92pc to 95pc, when we could have spent that time improving what happens when the face recognition is wrong – which might bring more value to the customer than an incremental accuracy improvement.”
Prioritise the customer
The potential that emerging technologies can have for overcoming challenges with data science, no matter the industry, is monumental. But for the sectors that are client and consumer-facing, the needs of customers should still come first.
Edell said there are always ‘pain points’, but relieving customers of those pains first and foremost presents a much smarter business approach than spending time perfecting solutions.
“Customers who have a pain will accept a less-than-perfect product as long as it solves their problem,” he said.
“But if you find yourself in a situation where you think customers aren’t buying your product because it isn’t accurate enough, it might be time to rethink what you’re solving for.”
Avoid inventing new things to solve
Given the opportunities that AI and machine learning platforms will bring to industries such as data science, it may be easy, and tempting, to become swept away in a sea of new inventions waiting to be nurtured.
However, Edell stressed the importance of avoiding this, emphasising that the most important problems to solve are the ones that already exist for people. Manual tasks that stunt innovation and creativity are at the top of that list.
“Look at things that people do manually today that are painful, such as sorting through images, categorising customer service e-mails or deciding what product image to show on your website.
“These are great opportunities to implement machine learning.”
Don’t strive for perfection
The majority of customer problems don’t require total accuracy. That’s good news for those in the data science industry because achieving perfection is unquestionably difficult, no matter the quality of a machine learning model on a validation dataset, Edell said.
He went on to describe a scenario he had recently experienced.
“The goal was to tell passengers when a particular restaurant was full and to reroute them somewhere else. To solve this, they built a model that used an existing camera to count the number of people in a room.
“But what they quickly discovered was that they didn’t need to be accurate about the number, there was no difference between 100 and 115 people. What mattered was that they could generalise about how full the restaurant was and reroute people accordingly.”
‘It is important to think long and hard about the use case you are trying to solve, and worry less about the technology that would be used to solve it’
– AARON EDELL
He pointed out that the first version of that model completed the task they wanted it to, so it would have been a waste of time and money to go back and try to improve upon it.
“If you really focus on the problem you are trying to solve, you may discover that recall matters more than precision,” he said.
“For example, if I am building an app that tries to determine if a mole on your skin is cancerous or not, I’d quickly find that telling people they don’t have cancer when they do is more problematic than telling people they have cancer when they don’t.
“I can optimise to make sure that I never miss a cancerous mole even though sometimes I’ll get some non-cancerous moles tagged as cancerous in my results. The bottom line is that it is very important to think long and hard about the use case you are trying to solve, and worry less about the technology that would be used to solve it.
“Why use a flamethrower when a match will do?”