AI-driven, powered by AI, transforming with AI/ML, etc., are some taglines we have heard far too often from the products we are being sold every day. Everyone is chasing after the promised land of machine learning but so few fully understand it. And that is indeed the origin of most of the challenges that data scientists face today in the execution of their ML projects.
Eliezer Shlomo Yudkowsky, an American artificial intelligence researcher, says —
“By far the greatest danger of Artificial Intelligence is that people conclude too early that they understand it.”
In this blog, we will look into some of the complaints and misunderstandings that turn into these challenges when not addressed head-on from the beginning of the project.
Given how fascinated businesses are with artificial intelligence and machine learning, they are often keen on pursuing a particular problem statement with an ML solution. ML can bring in a lot of value but not everything needs its application.
1. AI for CFD: Intro (part 1)
2. Using Artificial Intelligence to detect COVID-19
3. Real vs Fake Tweet Detection using a BERT Transformer Model in few lines of code
4. Machine Learning System Design
Many problems are straightforward and can be solved with a simple exploratory data analysis. It’s important to recognize the right use cases that require the heavy artillery of ML.
With so much hype around ML, the expectations are set high. While other technologies are required to be explained in terms of their capabilities, AI and ML need to be explained in terms of their limitations.
The media and marketing promise the moon but the reality is different. AI and ML are complex technologies that take time to implement and fully leverage. They consume a lot of resources in the process of delivering ROI. It’s important for data scientists to manage the expectations from the start.
An ML model isn’t a magic crystal ball and cannot predict accurately if there is not enough data. A couple of rows in a spreadsheet won’t drive actionable insights.
To develop a model that delivers the desired business outcomes, the data science team will have to ask for more, relevant data. With enhancements like augmented data management, they can then figure out how to best leverage that data.
There is a persisting expectation for highly accurate models. But in chasing after 100 percent accuracy, businesses tend to forget other factors like simplicity and engineering costs.
The most accurate model for Netflix that won the million-dollar prize didn’t end up getting implemented. Instead, another model that had a good balance of accuracy with simplicity, stability, and interpretation was adopted.
While early machine learning algorithms were simple to explain and understand, the deep learning algorithms are different — building layers with their own understanding.
AI supervisors definitely understand how a single prediction was made but cannot explain how the entire model works at large. Ali Rahimi, an AI researcher at Google, agrees that the entire field has become a black box. Due to this, it becomes increasingly difficult to explain the recommendations made by these models to the end-users.
Creating a good ML model is a lot of work and the data scientists aren’t always capable of foreseeing how much time it would take. The projects shouldn’t have harsh timelines and milestones.
The data science team can happen to achieve it in less or more time than predicted. Businesses should be patient while continuing to provide them with the resources they need. The data science team can also explain the long, iterative process to the stakeholders and what they can miss out when rushed.
After the data scientists have put so much hard work into building and testing ML models, it’s often asked if the models have learned all that they ever need to. An ML model needs to be continually trained to ensure that it’s future-ready. Businesses must incorporate the costs of doing so when they begin an AI/ML project.
Sometimes, when the model is almost done, there comes along a request to replace the output variable. It isn’t as easy as tweaking the title of a blog you are about to publish.
To fulfill a request like this, data scientists have to often go back to the drawing board. They have to pick the right influencers for a given outcome variable and map their relationship. This is why it’s important to ensure that the project requirements and goals are well-thought-out at the beginning.
Many people are drawn to the AI/ML industry with its promise of high salaries. Still, there are very few who have enough knowledge of both machine learning and software engineering and can build functional models.
Talent has always been scarce in the AI/ML industry. In the unavailability of the right expertise, projects can get frustrating and difficult to execute. As you begin, you must make sure that you have the right experts and consultants on your team to handle the magnitude of the project.
While a framework like Python-based Django is 13 years old, employing ML models is relatively new despite the hype around it for so many years. Google’s open-source framework for ML, TensorFlow was released in February 2017.
Deploying this new technology can come with unprecedented challenges and you should be prepared for them. Given how new it is, it’s also required to help all the stakeholders understand its realistic capabilities and the investment required.
Communication is key to deal with the challenges in machine learning projects. Data scientists should empathize with the stakeholders and understand the root cause of any disconnect. They can try to explain as best as possible what to expect in the execution of the project and hence, manage expectations.
Acuvate helps organizations implement custom big data and AI/ML solutions using various Microsoft technologies. If your organization needs help in leveraging these technologies, please feel free to get in touch with one of our experts for a personalized consultation. source