1. What is the project’s objective?
Carefully defining the project’s objective before undertaking it is a crucial phase to ensure its successful completion. This first step is an opportunity to sit down with the client and discuss their expectations. Depending on what the client hopes to predict and the nature of the data available, the project team can then determine the project’s feasibility and very carefully select the ideal type of learning (e.g., supervised, unsupervised, semi-supervised). Our qualified team will find the optimal approach to suit your needs.
2. Do you have enough data?
The crux of an AI project is the accessibility and availability of data for training the predictive models. Data accessibility is a key element that needs to be properly assessed, as it can influence the volume of data available. In other words, for certain learning models (e.g., deep neural networks or multilayer neural networks), a large quantity of data is required to obtain models that provide good predictions.
The volume of data available is an essential element that will dictate the training approach (classic machine learning or deep learning). Too little data can lead to biases during the learning phase and generate weak models that are therefore unable to make adequate predictions based on new observations. However, if the volume of data is small, it’s possible to bypass this constraint. Data augmentation is a method to artificially increase the size of the data set. Inexpensive and easy to implement, this approach makes it possible to improve model performance.
3. Is the available data of good quality?
Having a large data set is not sufficient for generating robust models. The models’ predictions will only be as good as the data set. Data quality must be validated in partnership with the client and the project team. Data weaknesses will be targeted, and measures can be implemented to remedy certain gaps (data engineering, imputation of missing values, data enrichment, etc.).
4. Do you have a good understanding of your data?
Having a “profile” of the data allows you to control it and helps you, among other things, understand it so you can make the right decisions in terms of data engineering. Various techniques and methods exist to achieve this, such as data visualization and correlation analysis, to name a few.
5. Is your model actually good?
Once the models have been trained, it is imperative to evaluate them. In other words, has the model correctly learned how to make predictions? A model may perform well in training, but how will it perform with a new data set? A very effective approach to validate a model’s performance and accuracy is to test it in a real-world context. It then becomes very easy to assess the validity of the predictions.
It is possible that certain issues may be observed, such as overfitting and underfitting, which are significant problems that need to be addressed. Once these issues are identified, there are techniques to boost model performance, such as cross-validation, regulation methods and increasing training data.
6. How can the model be maintained over time?
While training, testing and deploying an AI model are necessary steps to keep the model up to date, they are not enough. With the accumulation of data and possible business changes, the model will need to evolve over time to remain “alive” throughout the organization. So, it’s important to preserve the model’s robustness and durability by retraining it and making the necessary adjustments. To achieve this, a maintenance strategy must be defined and implemented.
7. Do you have a good data science platform?
There are a multitude of platforms offering advanced data analysis services. It’s essential to carefully choose the platform that best meets project needs. Some platforms are more specialized for specific domains while others are more generic. Carefully choosing the platform will ensure that you have the best tools adapted to project specifics and at a fair price. The use of non-adapted platforms can lead to poor time management; consequently, usage costs will be less than optimal. The final choice of a platform will be determined following an interview focused on the client’s needs and expectations, in both the short and long terms.
8. Is a multidisciplinary team in place to ensure the project’s success?
Properly selecting the experts who will be part of the team is also an important step in ensuring the success of an AI project. It’s well known that working in a multidisciplinary team has its advantages, making it possible to benefit from the expertise of each and every one. It’s not uncommon to need certain advanced knowledge in a particular area to fully understand what needs to be done and to do it properly. Therefore, the contribution of data science experts, engineers, geomaticians or geologists can prove to be very useful and give a 360-degree perspective of the project.
9. Does the client have full control of the AI solution?
At BBA, we take great care to continuously involve our clients in the project’s life cycle. Client involvement in project progress offers many benefits. First, by being aware of the progress, pitfalls or successes that the project team may encounter, the client can actively participate in developing innovative solutions to bring the project to completion. In fact, the client is constantly encouraged to contribute to finding solutions. In addition, throughout the project, the team can offer support to the client to remove any doubts. In the end, the client is in full control of the product. This way, the client can take full advantage of the benefits of the solution and even improve it over time.