Approaches to AI that Unlock Business Value
Artificial intelligence (AI) is already transforming business, driving down costs, maximizing revenue, and enhancing customer experience. And many organizations are taking notice: The AI market size is expected to grow to $390.9 billion by 2025, and industries within the space show a similar trend—automotive AI, for example, is expected to grow by 35% year over year, and manufacturing AI will likely increase by $7.22 billion by 2023. We see organizations accelerating their adoption of AI projects as well, with Gartner reporting that the average company adopted four AI projects in 2019 and is expected to adopt 35 in 2022. Even with this immense growth, challenges in deploying AI remain. According to top industry analysts, most (about 80%) of AI projects stall at the pilot phase or proof-of-concept phase, never reaching production. In many cases, this is due to a lack of high-quality data. Ethics and responsible AI continue to be obstacles for many companies, which often lack the resources or internal talent to build unbiased models in a time where AI is making increasingly impactful decisions. Companies also face an uphill battle with scaling and automation; while tech leaders are keen to apply DevOps principles to AI, they still struggle with architecting a solution for automating end-to-end machine learning (ML) pipelines. Developing the right tools and strategies upfront will help overcome these challenges, giving businesses the confidence to deploy and the potential to scale.Techniques and Tooling to Train, Deploy, and Tune ML Models
If there’s one key takeaway for deploying AI with confidence, it’s this: it’s all in the data. You know you need high-quality training data to launch effective models. So defining your data strategy upfront, including what your data pipeline will look like, will be crucial to success. To illustrate, let’s walk through a healthy ML pipeline:Collect and Annotate Data
Many data scientists and machine learning engineers say that about 80% of their time is spent wrangling data. That’s a heavy uplift, but a model can’t work without training data. The model build process, then, starts with collecting and labeling training data. You’ll want to start with a clear strategy for data collection. Think about the use cases you’re targeting and ensure your datasets represent each of them. Have a clear plan for collecting diverse datasets. For example, if you’re building AI for a self-driving car, you’ll likely want data representing different geographies, weather, and times of the day. Next, you’ll want to implement your data annotation process, which in most cases, requires a diverse crowd of human annotators. The more accurate your labels, the more precise your model’s predictions will ultimately be. Various perspectives will enable you to cover a broader selection of use and edge cases. At the data collection and annotation phase, it’s critical to have the right plan for tooling in place. Be sure to integrate quality assurance checks into your processes as well. Given that this step takes up most of the time spent on an AI project, it’s especially helpful to work with a data partner in this area.Train and Validate Model
When your training data is ready, train your model using that data. Most ML models leverage supervised learning, which means you’ll need humans to provide ground truth monitoring. They’ll check to make sure the model is making accurate predictions. This is often a critical phase, but is a lighter lift. If the model isn’t working in this phase, go back and ensure your training data is truly the right data you need. Optimize with a focus on the business value that this model is supposed to bring.Deploy with Confidence and Tune Model
Once your model reaches the desired accuracy levels, you’re ready to launch. Post-deployment, the model will start to encounter real-world data. Continue to evaluate the model’s output; if it fails to output the correct data, loop that data back through the validation phases. It’s helpful to keep a human-in-the-loop to manually check a model’s accuracy and provide corrected feedback in the case of low-confidence predictions or errors. Remember to tune your model regularly after deployment. According to McKinsey, 33% of live AI deployments require “critical” monthly data updates to maintain accuracy thresholds as market conditions change. In our State of AI 2020, we found that 75% of organizations said that they must update their AI models at least quarterly. Regardless, every model should be continuously monitored for data drift to ensure it doesn’t become less effective over time or even obsolete.Real-world Success Stories
