Keys to Success: implementing AI

Part II: Implementation

Data, data, data

Data is a crucial part which can make a project a success or failure. Machine Learning works better with large amounts of accurate data.
Even if you are not planning to start with an AI project right now, it is a good idea to make sure that the data you are collecting is suitable for future use.

While lack of accurate data can be a pain, data can also be leveraged. For instance, by adding distroted pictures of traffic signs to the original dataset, Ciresan et al managed to more than halve the error rate and beat human recognition rates!

Data collection checklist

Is my data collected in a consistent way? (E.g. not multiple labels for the same things)
Am I collecting all attributes that might be relevant?
Is labeling precise enough?

It is advisable to check with an expert if data is correctly collected and can be used in the future. Even if you don't leverage your data now, that data may be valuable in the future or to someone else.

Proof-of-concept

Accuracy, but also scalability and budget are factors to consider when looking for tech solutions. Even though AI research needs large investments, very often proven technologies can be re-used.

Today, real-world AI applications are rarely build from scratch, rather a composition of building blocks. The AI expert may be able to propose simpler or less expensive technical solutions by re-framing the problem.

Heuristic vs Model vs Machine Learning

Machine Learning sounds appealing but is not always the best solution. Essentially, Machine Learning aims at creating a model from a dataset. So if there is a theoretical model which is perfect, it is better to use it instead.

For example, it is better to use a formula to calculate the area of a circle rather than Machine Learning. Sometimes, a perfect model does not exist but if a good enough heuristic is at hand, it may be a better alternative to Machine Learning as well.

Find the right expert: hallmarks of quality

Before you buy a pair of shoes, you will look for precise details that show the quality of craftmanship. Before you work with anyone, it is important to know what to look for so you do not end up spending efforts, time and money for a solution that predictably won't work. Price is not always an indicator of quality either.
Look for these:

Honesty
De-risking: They start by building a simple version before improving it
They consider all of your requirements: scalability, speed, accuracy, confidentiality
They are able to simplify the problem
They come up with a clear pipeline which you can understand
Each stage in the pipeline is upgradable independently
They offer to re-use available tools when possible
Qualifications: Check is they have worked with the same type of problem class and tools before. The AI problem type matters more than the applications.