This blog post is a first of a serie of paper reviews. We will write in this serie notes and explanations of research papers that we liked.This blog post is based on the research paper "On Calibration of Modern Neural Networks" by (Guo et al, 2017 .).
How To Be Confident In Your Neural Network Confidence?
Very large and deep models, as ResNet, are far more accurate than their older counterparts, as LeNet, on computer vision datasets such as CIFAR100.
Most neural networks for classification uses as last activation a softmax: it produces a distribution of probabilities for each target (However while they are better at classifying images, we are less confident in their own confidence!cat, dog, boat, etc.). These probabilities sum to one. We may expect that if for a given image, our model associate a score of 0.8 to the target 'boat', our model is confident at 80% that this is the right target.
Over 100 images that were detected as boat, we can expect approximately that 80 images are indeed real boats, while the 20 remaining were false positives.
It was true for shallow model as LeNet but as newer models gained in accuracy their confidences became decorrelated from the "real confidence".
This does not work anymore for deep neural networks:
As you can see, older networks as LeNet had a low accuracy (55%) but their confidence was actually in line with the accuracy! Modern networks as Resnet have a higher accuracy (69%) but as showed in figure 1, they are over-confident.
This discrepancy between the model confidence and the actual accuracy is called miscalibration.
Why It Is Important?
Outside of toy datasets used in the academy, it can be useful to know how much confident our model is.
Imagine we have a model predicting frauds. We want to flag some transaction as suspicious based on the model confidence that it is a fraud. We could definitely compute an optimal threshold on the validation set, and then every confidence above this threshold would be flagged as a fraud. However this computed threshold could be 0.2 or 0.9 but would probably make much sense to a human.
A model without miscalibration would help the users to interpret better the predictions.
Why It Happens?
The authors explores empirically what are the causes of this miscalibration in modern networks.
They measure the miscalibration with theExpected Calibration Error (ECE): the average difference between the confidence and the accuracy. This metric should be minimized.
Higher Capacity & Cross-Entropy
The most interpretable cause of the miscalibration is the increase of capacity and the cross-entropy loss.
Model capacity can be seen as a measurement of how much a model can memorize. With an infinite capacity, the model could simply learn by heart the whole training dataset. A trade-off has to be made between a low and high capacity. If it is too low the model wouldn't be able to learn essential features of your data. If it is too high, the model will learn too much and overfit instead of generalize. Indeed comprehension is compression: by leaving few enough capacity the model has to pick up the most representative features (pretty much in the same way PCA works) and will then generalize better (but too few capacity & no learning will happen!).
The new architectures such as ResNet have way more capacity than the older LeNet (25M parameters for the former and 20k for the latter). This high capacity led to better accuracy: the training set can almost be learned by heart.
In addition the models optimizes the cross-entropy loss that force them to be right AND to be very confident. The higher capacity helped to lower the cross-entropy loss and thus encourages deep neural networks to be over-confident. As you've seen on figure 1, the new models are now over-confident.
The Mysterious Batch Normalization
Batch Normalization normalizes the tensors in a network. It greatly improves the training convergence & the final performance. Why exactly it works that well is still a bit undefined (see more).
The authors remark empirically that using Batch Normalization increased the miscalibration but could not find an exact reason why.
Could the help given by this method in training facilitate the over-confidence?
The weight decay is an additional loss that penalizes the L2 norm of the weights. The larger the weights, the bigger the norm and thus the loss. By constraining the weights magnitude, it avoid the model finding extreme weight values that could make it overfit.
The authors found that increasing the regularization decreases the model accuracy as expected. However it also decreased the miscalibration! The answer is then again because regularization avoid overfitting & thus over-confidence.
How To Fix Miscalibration
This article's title, "Similar to (How To Be Confident In Your Neural Network Confidence", led you to believe that you would discover how to reduce miscalibration.
You're not going to reduce the capacity, remove Batch Normalization, and increase the regularization: you'll hurt too much your precious accuracy.
Fortunately there are post-processing solutions. The authors describe several but the most effective one is also the simplest: Temperature Scaling.
Instead of computing the softmax like this:
All the logits (values just before the final activation, here softmax) are divided by the same value called temperature:
Similar to (Hinton et al, 2015.), this temperature softens the probabilities.
Extreme probabilities (high confidence) are more decreased than smaller probabilities (low confidence). The authors find the optimal temperature by minimizing the Expected Calibration Error on the validation set.
The miscalibration is almost entirely corrected:
Another cool feature of Temperature Scaling: because all logits are divided by the same value, and that softmax is amonotone function, the accuracy remains unchanged!
Originally published at https://lab.heuritech.com.