How to evaluate the performance of AI models?

Evaluating the performance of AI models has never been easy. To be honest, every time I see various models claiming how accurate they are, I wonder, "Are these numbers really reliable?" Just as we can't test electronic products by just looking at the running score, AI models need to be evaluated by considering all aspects. Sometimes a fairy model that performs well in the lab becomes unbearable in the real world. This reminds me of a while ago a friend complained that their NLP model was as accurate as 95% on the test set, and the result was that the actual application almost pissed off the customer - because the 5% errors just happened to all appear in the keywords. So yeah, you really can't take this whole AI performance evaluation thing at face value.

Assessment indicators are not a panacea

If you think that metrics like accuracy and recall are a complete description of the strength of an AI model, you might be naive. The reality is that a model that is invincible in a Kaggle competition may not even pass the grade when put into a production environment. Take our common accuracy - in a dataset of 99% negative samples, even if the model does nothing but predict negatives, the accuracy is still as high as 99%. but with a model like this, do you think it's useful?

Don't overlook the cost of time

I've found that many teams are particularly prone to ignoring the time factor when evaluating models. They spend two weeks to train a model with 0.5% improvement in accuracy and then rejoice, without calculating how much computational resources are consumed by this 0.5% improvement. In reality, we often need to make a trade-off between "good enough" and "timely response". Like online customer service scenarios, users would rather have a model that gives an 80-point answer in 1 second than a system that gives a 95-point answer in 10 seconds.

Then again, it's the data bias issue that gives me the most headaches in evaluating AI performance. Sometimes, I've gone to great lengths to tweak hyper-parameters, only to realize that the real bottleneck is in the data quality. I remember once we made a facial emotion recognition system, which performed like a mind reader in the lab, but turned out to be frequently wrong when used on Asian users - because white samples accounted for 80% of the training data. At such times, the perfect algorithm is useless.

The real scenario is the final exam room

A/B testing is the ultimate test of a model, and I know it well. Lab environments are like greenhouses, while real user scenarios are like wilderness survival, sometimes even developers don't know how your AI will be used. Sometimes even developers can't imagine what kind of posture a user will use your AI in. we have a dialog system that performed perfectly in internal testing, but when it went live, we found that users were always asking questions in all kinds of dialects and slangs, which made the model confused. So, evaluating AI performance is not just a matter for engineers, it requires the participation of product and operation roles. After all, AI is ultimately to serve people, isn't it?