AI Capabilities
In the manufacturing industry, there is a general lack of AI computer vision knowledge and experience. Most people in the industry have learned about AI through the news media or some type of online course. It is one thing to study AI and it is another to apply AI into the real world.
As such, it is rare to come across people within the manufacturing industry who have had experience applying AI computer vision to real use cases. Because of this, there is a lack of understanding of the capabilities of applied AI computer vision. Building a simple AI computer vision model and getting it to work on a test set on a desktop is a different exercise than having the AI computer model run in a live production environment with variation in the process.
AI Transformation
Manufacturers who are serious about AI transformation should seek to acquire AI knowledge through both study materials and practical experience. These efforts can be done solely in-house or can be accelerated via cooperation with third parties that already have the expertise such as i-5O.
i-5O offers an AI-powered computer vision solution that helps manufacturers gain access to comprehensive production floor intelligence based on a holistic approach to enhancing manufacturing efficiency and performance.
AI Limitations
Unfortunately for all, AI is not magic and there are limitations to all the different AI algorithms that can be deployed into mass production use within a manufacturing facility.
Questions to consider regarding the limitations of AI include:
1. Use case limitations
2. Statistical limitations
How accurate will the AI be to create value and the level of effort needed to reach those benchmarks?
3. Performance limitations
How will the AI model be applied to the use case?
4. Cost limitations
Depending on the requirements placed upon the AI, how much will it cost to develop, train, and run?
Trade-Offs
Plenty of times, AI limitations can be thought of in terms of trade-offs instead of absolutes. Some examples of tradeoffs include:
1. Precision vs Recall
The tradeoff between precision vs recall for any AI model. Higher precision means that an algorithm returns more relevant results than irrelevant ones, and high recall means that an algorithm returns most of the relevant results. Both are not possible due to algorithmic limitations.
2. AI speed vs available compute power
The tradeoff between AI speed vs available compute power. High speed will mean low latency but higher cost, but if quick insights are not a high priority, the focus should be on higher compute power at the expense of quick insights.
AI Costs
An important factor that anyone looking to deploy AI computer vision in mass production needs to keep in mind, is that several levers can increase AI cost.
Cost levers in AI include:
1. AI implementation scale
The size of the implementation drives cost which can only be brought down once scale is achieved.
2. AI use case
Different AI use cases require the application of different AI models which can all have other costs to develop, train, run, and maintain.
3. AI use case complexity
Even within an applied AI use case there can be varying degrees of complexity which will impact development and training costs.
4. AI model latency
AI models are typically run on GPU and fast AI results mean that the overall system needs to be optimized to a higher level which drives costs.
5. AI model accuracy
The more accurate the AI model, the higher the associated costs. It is important to note as well that the relationship between accuracy and cost is not linear.
6. Manufacturing operation complexity
The nature of the manufacturing operation can be a factor in the costs. Dynamic operations with high variability can increase costs. Cycle time length can also be a cost lever.
7. Manufacturing process complexity
The nature of the manufacturing being targeted with AI can also affect costs. Highly complex processes are typically more costly to deploy AI on than simple processes.
When it comes to AI costs, a classic trap that manufacturers face is the allure of creating highly complex and accurate solutions to the point where the AI costs kill any potential to generate ROI. In many cases, an AI system that is ‘good enough’ is already much better than the status quo and will have a higher chance of being implemented into mass production use because of the enterprise value created.
ROI Calculation
For manufacturers looking to implement AI in their operations, they need to have a clear understanding of the ROI that they can gain to justify the investment. Because of the lack of understanding by most people in manufacturing on the capabilities and limitations of AI, ROI calculations are typically incorrect by a large margin. There is a tendency to overestimate the capabilities of AI and to underestimate the costs which leads to unrealistic expectations and scope creep. Scope creep or deviation from the original scope of the AI project can pose significant challenges to project success, impacting budgets, deadlines, deliverable quality and credibility.
Questions to ask when calculating ROI include:
• What is the financial upside of applying AI to a particular use case and the cost of inaction?
• Is AI the appropriate solution to solve this problem and what are the appropriate AI models?
• What level of AI performance will be needed to solve this problem at an acceptable level?
• Will the financial gain from applying AI to a use case be greater than the cost of AI?
Conclusion
While AI computer vision holds immense potential to revolutionize the manufacturing industry, successful implementation requires a nuanced understanding of both its capabilities and limitations. Manufacturers must approach AI not as a magical solution but as a tool that demands careful consideration of trade-offs, costs, and real-world applicability. The journey to AI integration should be guided by clear ROI calculations, practical knowledge, and collaboration with experienced experts. By striking a balance between ambition and feasibility, manufacturers can unlock significant value and drive long-term operational improvements without falling prey to overhyped expectations or unsustainable costs.