1

As we know, classification problems are just a bunch of probabilities, commonly it comes from logits or softmax output. Performing $argmax$ to get the most favorable class by model, discarding us from some information. Such as the distribution of the probability distribution.

For deep lerning model that doesn't have reasoning capabilities, So I formulate new term based on model favorable and model asking for validation so I have four quadrants. It's based on purely softmax output only.

For example, here 8 softmax output with 3 classes:

Out0 [0.99  0.005 0.005]: Confidence: 0.98, Uncertainty: 0.00
Out1 [0.8 0.1 0.1]: Confidence: 0.70, Uncertainty: -0.00
Out2 [0.5  0.25 0.25]: Confidence: 0.25, Uncertainty: 0.00
Out3 [0.33 0.33 0.33]: Confidence: -0.00, Uncertainty: 0.01
Out4 [0.99  0.009 0.001]: Confidence: 0.98, Uncertainty: 0.80
Out5 [0.8  0.19 0.01]: Confidence: 0.70, Uncertainty: 0.90
Out6 [0.5  0.49 0.01]: Confidence: 0.25, Uncertainty: 0.96
Out7 [0.1  0.45 0.45]: Confidence: 0.18, Uncertainty: 0.64

Rescaling that range from $[0,1]$ to $[-1, 1]$ for visualization, draws: enter image description here

Then I tried with real data of MNIST handwritten digit of 10 classes. Focusing on label number 1, subjevtively model should be easy for him to classify it, but what if I ask him for something unusual that never seen before, it's 90° degree rotate version of number one.

enter image description here

As I expected, model should be wrong to classify something unusual which is rotated version of number one, and therefore leading to low confidence and high needed assistance.

If I visualize with the four quadrants, it's clear that the blue (which is rotated sample) is seeking for the validation or assistance compared to red (the original).

Therefore this is useful for critical inference such as medical diagnoses that doesn't need ground truth for explainable AI.

enter image description here

The question is, what is the better description to represents x-axis and y-axis?

As you see I have inconsistent to use the terms for y-axis:

  • Uncertainty (I think it doesn't fit with my definition)
  • Affirmation (Maybe there is another better)
  • Assertion
  • Assistance
  • Vagueness

So, what is the best name for this? (no, this is not opinion question, it's objective question that intersect with philosophy)

Muhammad Ikhwan Perwira
  • 800
  • 3
  • 10

1 Answers1

2

The terms you choose for the x-axis and y-axis should ideally align with the concepts of most people's intuitive understanding who will view these results. "Confidence Score" of your probability classifier as the predicted maximum class probability itself is a well-understood term in ML and statistics, and accurately describes what your x-axis captures, so no change is needed here. "Assistance Needed" could be the best term comparatively from your candidate list since it essentially measures the spread of all the predicted class probabilities like entropy. Therefore in ML a good term for your y-axis could be "Predictive Entropy" or "Confidence Calibration".

Calibration can mean procedures in statistical classification to determine class membership probabilities which assess the uncertainty of a given new observation belonging to each of the already established classes.

cinch
  • 11,000
  • 3
  • 8
  • 17