Summary
This submission proposes a post-hoc method for detecting out-of-distribution samples, by pruning the final classification layer, using a sensitivity metric based on the gradients of the energy scores. The proposed method is mainly validated on residual networks and visual transformers based on the ImageNet dataset.
Strengths
- The submission proposes an interesting application using pruning to out-of-distribution detection, the latter being an important research problem
- The proposed method is simple, and can be applied with a fairly small computational overhead
- The results suggest that the proposed method, also when used in conjunction with other methods, can achieve state-of-the-art results on several tasks.
Weaknesses
- The submission is not very well written, particularly the introduction, with quite a few typos and grammar issues (e.g. using “post-hot” instead of “post-hoc” in quite a few places). This makes reading a bit difficult.
- The proposed method is not very novel, as other OOD detection methods have used pruning (e.g. references 20, 35 seem quite similar)
- The performance of the method depends quite heavily on the pruning thresholds (see Figure 3) and can substantially decrease the accuracy on the original ID classification task (see Table 3). This makes the method less likely to be used “out of the box”.
- While the authors introduce their method as "post-hoc", in fact it has a similar cost to training the model for one epoch, since the gradients for all the samples have to be computed.
Questions
- Suggestion: I would advise the authors to go through the submission and correct the typos and grammar issues, for the next revision
- I believe Equation (4) is an approximation, rather than an identity. Have the authors considered what would be a theoretical explanation that would connect the approximation in Equation (4) and the removal of least and most sensitive weights?
- What is the norm in Equation (6)? Should it be an absolute value instead?
- Can the authors provide more explanations on the intuition that “parameters and neurons with exceptionally large sensitivity can easily lead to overconfidence” (51-52)?
- Lines 145-147: it is hard to see these differences from Figure 1, I would suggest adding numbers
- Lines 181-182: this is not really a Gaussian (e.g it is not symmetric)
- Can you provide more details on how OPNP is combined with ReAct?
- In Table 2, can you provide results for DICE + ReAct?
- Some of the plots and tables do not include information about the model and dataset used. For example, what is the model used in Table 3 or Table 4? What is the model and dataset from Figure 4?
- The authors show results when (independently) pruning other layers in a ResNet50 model (Appendix Table 3). What happens if the weights are pruned globally, instead of each layer at a time?
- Have the authors considered second order information for computing the sensitivity metric? For example, a second order Taylor approximation could be used instead of Equation (4), and since pruning is only done for the final layer, the Hessian could be more easily calculated. This is a similar approach to Optimal Brain Surgeon (Hassibi et al.)
- Can you clarify what is meant by "(subset of) iNaturalist, SUN, Place and Texture" (line 211)? Are the same corresponding datasets used for evaluation when comparing with the other methods from Tables 1 and 2?
- Can you please confirm that the ID test set used for validation (and determining the pruning hyperparameters and OOD threshold) was different than the one used for the final testing?
- Can you please clarify the differences between Algorithm 1 and the text which specifies that the parameter and neuron sensitivities are computed based on the entire training set? In line 5 from Algorithm 1, it sounds like neuron pruning is done individually for each sample, but from the text the neuron sensitivities are computed on the entire training set. In my understanding, the model is pruned only once, and the same sparse model is used for evaluation on different OOD tasks. Can you please confirm whether my understanding is correct?
Rating
5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.
Confidence
3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
Limitations
While the authors have addressed some of the limitations of their work in Section 5, I believe the proposed method, in its current form, has some technical limitations that would make it more difficult to use “plug-and-play” as the authors hope. For example, the method relies quite heavily on the optimal hyperparameters (thresholds) for pruning and it can have a substantial negative impact on the ID classification accuracy for the original task. I would advise the authors to consider how the search of the pruning hyperparameters could be automatized, or at least made more efficient.
--------------------------------
**Edited after rebuttals**
After reading the authors' answer, I have raised my score from 4 to 5.