Authors:
(1) Anonymous authors ’paper under a dual -blind review Jarrod Haas, SARLAB, Department of Engineering Sciences Simon Fraser University; Canada digital group and [email protected];
(2) William Yoland, Metaptima and [email protected];
(3) Bernhard Rapos, SARL, Department of Engineering Sciences, Simon Fraser University and [email protected].
- Abstract and 1 introduction
- 2 background
- 2.1 Definition of the problem
- 2.2 Related work
- 2.3 Deep inevitable uncertainty
- 2.4 L2 normalizes the space of feature and nervous breakdown
- 3 methodology
- 3.1 Models and losses
- 3.2 Measurement of nervous breakdown
- 4 experiments
- 4.1 Faster and more powerful results
- 4.2 Linking nervous breakdown with OOD discovery
- 5 Conclusion, future work and references
- extension
- A.1 Training details
- A.2 L2 normalization effect on SoftMax to detect OOD
- A.3 GMMS installation on the registry space
- A.4 Excessive training with L2 normalization
- A.5 Nervous breakdown measurements for NC loss intervention
- A.6 additional numbers
3 methodology
3.1 Models and losses
For all experiments, we used either Resnet18 or Resnet50 models that were provided with DDU (Mukhoti Et al., 2021). All of them are trained from fifteen independent seeds (Training details in the appendix A.1 can be found. All foundation lines used an objective function via entropy (CE) during training. The NC intervention group described in section 4.2 did not use CE loss, but instead used the loss function that contains discriminatory scales shown below in Section 3.2:
Note that the scale for NC4 is not used, because it requires the Argmin function that cannot be dismantled. Although these scales do not have the same range, they all advance to scratch, and we did not find that it is necessary to use any weighting scheme within the loss function.
3.2 Measurement of nervous breakdown
NC has four properties:
NC1: The collapse of contrast: contrast within the category for each category in the feature of the feature, approaching zero.
NC2: Racing with the narrow frame of EquianGular Simplex (Simplex ETF): The corners are maximized between each pair of the category means to the maximum, equality and distances of each category of the global average of equal groups, that is, the category, is placed in the maximum location on excessive locations
NC3: Concentration with self -duplication: The ideal decision -making areas and the class are converging with symmetry, where each separation is occupied by the average center of its decision, and all areas of decision -making with an equal size.
NC4: Simplification to the nearest NCC Class Center: The work is appointed as the highest possibility of a specific point in the space space to the nearest category.
Papyan et al. (2020) Use seven different standards (EQ. 3 to eq. 8) to monitor these properties. All are discriminatory and are used in the NC loss function in Section 4, with the exception of the NC4 scale (EQ 9), which is not discriminatory
The contrast is measured inside the category, NC1, by comparing the contrast within the category with the contrast between the category,
NC2 is referred to through four measurements. The equivalent of the methods and means classified through the contrast factor are given,
Finally, the nearest class classification, NC4 is measured, as a percentage of the training group samples that are classified when using a simple decision base based on the distance of the feature of the feature to the nearest category,
Where z is a feature of the advantage of a specific entry.
