In production use cases, since we always need 100% accuracy. Eg. Like a chatbot cant ever be false in replies about something. How do we deal with accuracy of anything less than 100%?
The required accuracy depends on the criticality of the situation.
If you want 100% accuracy in all cases, you cannot even use human intelligence for that, let alone AI.
1 Like
I don’t remember now which course Andrew Ng mentions, the idea of achieving the best model based on accuracy isn’t about just statistics or computation as we try to achieve as much as model accuracy of 99.999(so even agree to have a realistic model), always comes with that 1% probability of model not being accurate and I think this statement he probably mentions based on how a model would work on unseen data or realtime data.
creating the most perfect data with 100% model accuracy would be limited to simple classification model, for example detect apple or oranges product in a bakery products. Such models comes with its own limitation of where and how one can use the model.
So expectation of chatbot with 100% model accuracy is not only unrealistic but also alarming for reviewing the data, model architecture as well as the limitating the complexity of a chatbot. Like how they imperfection also includes (I am perfection) probably that’s why in neural network training we add biases to the weight, for model to learn it’s feature significance and variability.
Regards
DP
1 Like
If you haven’t yet, I suggest take a peek under the cover of a chatbot algorithm and learn how sequences of output words are generated. Here, Attention in Transformers: Concepts and Code in PyTorch - DeepLearning.AI for example. It will clarify why chat bots can’t be relied on to generate 100% predictable output.
1 Like