Implementing an AI project the right way: Here's how it worksposted in tech
Do you want to reduce costs and introduce more efficient workflows in your company? Then you may have thought about using artificial intelligence to automate processes. We want you to be optimally prepared. That’s why we show you how an AI project works.
We share our experiences and best practices from deep learning projects where we focussed on processes with visual data (images or videos).
Read on to learn more about the project steps, the sticking points, and how everyone involved can help make an AI project a success.
1. Initial consultation
In an initial discussion, we record your requirements.
- Inventory of your current processes and your change requests: How are your current processes structured? Which processes would you like to change?
- Goal setting: What end result do you want? How exactly should the new processes look like? The goal should be described as detailed as possible.
- Budget: What’s your budget? Together with the intended goal, the budget sets the possible paths we can take together in the project. In most cases, you want to save costs or achieve higher sales by introducing AI. This plays a decisive role for the budget.
- Data situation: Do you have data we can use for training? If so, which data and how much data is that? Is there ongoing data collection throughout the project, or does the foundation need to be laid first?
In this step, we evaluate and plan the implementation of the project together with you. In detail, this means the following.
Data inspection and further data planning
We review training data provided by you, e.g. labeled images, and get an idea of whether they can be used for the training.
Since you need a lot of training data for deep learning, this is a crucial point. The evaluation of the data also includes the assessment of quality and balance, because this influences how well an AI model learns and makes correct predictions.
If you cannot provide any data at the start of the project, a separate project is required first, which only serves to collect data. For you, this means the structured collection of data from the company. Alternatively, it is also possible to purchase data sets or use labeling services. We are at your side to advise you.
New data is constantly needed throughout the duration of the project in order to further improve the quality of the model. Therefore, we need to plan with you how you continuously collect this data, identify and correct false predictions of the model so that you can provide them to us. The correctly recognized data as well as the incorrectly recognized data with corrections flow into the next training session.
Definition of the Minimum Viable Product (MVP)
Together with you, we define what a minimally functional version of the AI can look like. The basic question here is: Which components or features should go live first so that you can derive added value as quickly as possible?
An advantage of this approach is that you can test the new AI-based process on a small scale. At the same time, we can identify improvements more quickly. You can then scale up and add more features at a later date.
However, the compelling arguments for starting with an MVP are cost reduction and risk minimization. Instead of implementing a huge project, a small project that creates added value is put together and tested in reality. This avoids bad planning and developments that can cost a lot of money.
Definition of Key Performance Indicators (KPI)
Key performance indicators are important for objectively measuring the quality of AI and the business impact. These targets define what the planned system should be able to achieve. Examples of KPIs can be:
- The average time saving of the process by partial automation
- Guaranteed response time with maximum requests per second
- Parallel possible requests to the AI
- Accuracy of the model
We plan the necessary integration into your production system with you. Important questions are: How should the AI be used in the existing software environment and in the workflow? What is necessary to access the AI?
With the initial discussion and evaluation, the foundation for the project has been laid. In the subsequent steps, we are pushing the development further and further. Steps 3 to 5 are repeated until we have progressed from MVP to the desired solution.
3. Training Iteration
We train the model with most of the available data. Then we check the performance of the model with unseen data.
Training a deep learning model on images or videos is more complex and time-consuming than for text-based or numerical machine learning tasks. This is because we use deep models (with many layers) and the datasets processed are usually very large.
Depending on the project, training the model is only a fragment of the entire development process. It is often necessary that we build a complete process in which the model can be embedded, such as a web service for example.
If a good quality level of the model is reached after training, we will deliver the first version to you (MVP). We usually provide you the version as a Docker image with API. You then start integrating it into your system and workflows. We will be happy to accompany you.
5. Collect Feedback
After integration into productive operation, it is very important that you collect data from the usage. This is the only way to assess whether the AI is working as you imagined. This way, we can see what the model can and cannot do in real operation.
You collect this production data and transmit it to us. We then feed this into the next training iteration. Please don’t be surprised, this data acquisition may take some time. Of course, it depends on the extent to which you collect data. But it may take weeks or even months before the next iteration begins.
The next iteration
In order to achieve a significant increase in result quality with the next iteration, it may be necessary for you to provide us with more data or other data that arise from real operation.
However, a next iteration can also be motivated by a change in the requirements, for example if new categories have to be recognized in a classification model. In this case the current model isn’t able to make good predictions, it has to be trained with new data first.
Tips for a successful AI project
A crucial sticking point is the iterative approach and gradual introduction of an AI-based process. This is the only way to increase the quality and range of functions of the development.
Furthermore, you should keep in mind that the provision of training data is not a static process. It is a cycle in which you, the customer, play a crucial role.
One last important point: the measurability of the project. You need to monitor and track it. Because only if the target values are measured during the project we can see regression or progress and you can finally achieve your goal.
This article was made possible through the great cooperation with DATANOMIQ, a consulting company for data science, business intelligence and process mining.