CEIR: Concept-based Explainable Image Representation Learning

It seems to me that the main weakness of this paper lies in a form of misunderstanding. The authors pretend that their work is an unsupervised method (representation learning), and the whole motivation of the paper comes from this, , all the way down to the clustering task used to validate the method.

The reality is in fact quite different, in that the core of the method is based on CLIP, which has been learned from a gigantic database of annotated (captioned) images. CLIP's zero-shot classification performance is very good, as numerous papers have shown. The CLIP representation is likely to contain precise definitions of the database labels used to evaluate the proposed method. Not surprisingly, a representation derived from CLIP is good at doing clustering. The problem is that evaluating a method on a clustering task when the method uses labels is not fair at all.

(1) I cannot understand what the method is doing. For example, I don’t see what the superscript “3” in Eq. (1) means. Does it mean cubic? Why it’s necessary? $\mathcal{Q}$, $\mathcal{H}$, $\mathcal{A}(\mathbb{R}^\mathcal{Q})$, and $\mathbb{R}$ are not defined (at least the paper is not self-contained). All these missing details make it almost impossible to see what’s going on in the method.

(2) My major question is the reason why the backbone and the projection layer are necessary. If $P_{i,:}$ is the target when training the projection layer and $q_i$ is expected to be closer to $P_{i,:}$, one can use $P_{i,:}$ as a concept vector.

(3) If I understand correctly, VAE compresses the concept vectors to a latent vector. I don’t see why this is necessary. What is the typical number of concepts for each dataset (are they like 182 or 1401 as in Table 10)? As all the information provided in the latent vector is in the concept vector, I’m not sure applying dimensionality reduction really benefits some aspects of the method. An ablation study may help understand.

(4) I cannot get what is evaluated in the experiment section. I think the paper should evaluate how well the method can find the designated concepts in the concept vectors. The method has two concept representations (concept vectors and concept importance), so their consistency should be evaluated. I understand that this is not straightforward, as there are no annotations on the concepts, but I think evaluation over a small subset is helpful. Also, the reference set for each class should be evaluated in some way (I’m sorry, I didn’t come up with a good way except for human evaluation) to show the validity of the approach. I also think the representation should be evaluated for downstream tasks, like classification.

The writing is frequently unclear, which makes many parts of the paper hard to understand.

The evaluation is limited to everyday object/scene datasets (ImageNet, CIFAR, STL-10) which are the datasets where this approach should work best due to the high overlap with CLIP’s training set. It would be nice to see evaluation on a broader range of datasets.

There's no user experiment, so it's unclear if these explanations would be useful for humans or how they compare to other XAI approaches.

The explanations don't seem to be entirely correct, given the examples in Figure 3. It seems like the model just lists concepts that are related to the predicted class, regardless of whether they are present in the image (e.g., “lion-like mane” for a female lion, or “sun lotion” for an image mistaken for a swimsuit photo, or “rotational movement” for an image containing a ball).

I have several concerns :

• My first concern is related to the positioning of the paper compared to concept-based explainable image representation. In particular, in the XAI field, some criteria and properties have been proposed to define the concept of good explainable representation (see for instance the work of Ghorbani here) such as meaningfulness, coherency, and importance. How the proposed approach tackles these aspects is not clear and not evaluated at all. In particular, the proposed approach is not compared to the existing state-of-the-art on concept-based explainable image representation, supervised or unsupervised.
• In the same vein, since the claim of the paper is to reach to human understandable image representation, an experimental study that supports this claim should be provided. In particular, some metrics (correctness, stability, plausibility) have been proposed in the XAI field to evaluate explanations without human-level studies. See for instance this paper for faithfulness and understandability criteria.
• Some components of the proposed approach lack a clear justification and motivation for instance the GPT part compared to the use of existing explicit knowledge (e.g. Wordnet ontology or existing knowledge graphs).
• Some technical details are missing to evaluate the results correctly. For instance, in Table 2, what is the size of the corresponding representations ?
• Another concern is related to written level of the paper. The paper needs a complete proofreading since it contains a lot of mistakes (bad punctuation, strange ways to indicate the references).
• Sometimes some references given are not good ones. For instance, in the introduction, Yang et al and Crabbé are not seminal works for concept-based representations.