Artificial intelligence promises a great deal, but the road to success is often littered with unexpected pitfalls:
⚠️ 70% of companies experience little to no impact from A.I.
⚠️ 87% of data science projects never reach production
Until recently, the COVID-19 pandemic dominated our lives, and naturally artificial intelligence was also used in attempts to diagnose infections or predict risks. But researchers found that none of the hundreds of AI models developed were suitable for clinical use, mainly because of data quality problems:
📌 Some models used a dataset of healthy chest scans from children as examples of non-COVID cases for A.I. This caused the A.I. model to train itself to recognise scans of children's lungs rather than COVID cases.
📌 In other models, the text label printed alongside X-rays incorrectly became the trigger for recognising COVID cases.
📌 Still other models misclassified COVID cases because they incorrectly distinguished between lung scans of people lying down and people standing up. Because sick people were often scanned while lying down, these models made predictions based on whether a lung scan was taken in a lying or standing position.
💡 In other words, bad data leads to bad results, regardless of how advanced the algorithms are. Data quality is one of the biggest factors determining the success of an AI model. There are also several other important reasons why AI projects fail:
1️⃣ Unclear objectives: The success of A.I. starts with a clear, unambiguous definition of the business problem you are trying to solve. Without this foundation, projects are doomed to fail.
2️⃣ Lack of collaboration: A.I. requires seamless collaboration between multidisciplinary teams. Isolating an AI team from the rest of the company is a recipe for failure. For a data scientist, it is extremely important to understand the substance of the data, and business units generally know their own data best.
3️⃣ Insufficient expertise: A shortage of skilled staff is a significant barrier. Investing in training and possibly outsourcing are crucial considerations.
4️⃣ Ethical and privacy considerations: A.I. systems must be deployed ethically and responsibly. This requires continuous attention to data ownership and transparency.
✉️ Hopefully this post gives a little more insight into the challenges involved in an AI project. Feel free to connect if you are interested in AI!
🖼️ Image generated by ChatGPT 4 / Dall-E 3, based on the text above.
