Listenable Maps for Zero-Shot Audio Classifiers

Summary

This paper describes an extension of the LMAC method for explaining decisions made by audio classifiers. The novelty in the present article is to extend from fixed-vocabulary settings to zero-shot / open text description settings. To accomplish this, the authors propose a training objective that aims to preserve proximity of (learned) masked audio embeddings to embeddings of corresponding text descriptions. The method is evaluated on the CLAP model over two standard sound event detection datasets, and compares quite favorably to prior work along several evaluation criteria.

Strengths

Overall, I found this paper to be well written and easy to follow. The topic of the paper is timely and relevant, as much audio analysis research does seem to be trending toward open vocabulary settings. The empirical evaluations are appropriate and generally convincing. The proposed method makes a lot of intuitive sense, and appears to be a natural extension of the fixed vocabulary setting (section 2.2).

Weaknesses

The main weakness I see in this work is a somewhat shallow investigation of the proposed method itself, as opposed to high-level comparisons to prior work. The core technical contribution is described in equations 5-7, and consists of 3 terms: one to approximately preserve the embedding of audio (relative to text embedding) after masking, one to promote sparsity on the masks, and one to promote diversity of masks when conditioning on different text embeddings. As in CLAP, contrast is obtained by comparing amongst other samples within the training batch. This raises a handful of questions about the various components of the training objective which I will detail below; however, the critical issue here is that no ablation study is conducted to measure the importance of the various terms. This leads the reader at a bit of a loss to understand how these components interact with each other, whether they're all necessary, and so on.

Questions

My questions primarily center around equation 6: - Notational quirk: $C_{i,j}$ is a scalar value (equation 5), right? Why are norm bars ($\|$) used in eq 6 when computing the loss on approximating $C_{i,j}$? - Reasoning through the first term of eq6, a few things jump out at me. First, expanding out the definition in eq5, the summand should be equivalent to $\left|t_i^\mathsf{T} \left(f_\text{audio}(X_{\text{audio},j}) - f_\text{audio}(M_\theta(t_i, h_j) \odot X_{\text{audio},j}) \right)\right|$. That is, the gap between the original and masked audio embedding should be small (ie the mask doesn't do much) or approximately orthogonal to $t_i$, and this should hold for all $t_i$ in batch when the outer sum is computed. - What is the effect of the batch size / batch diversity on this loss? Presumably, the more (and more diverse) text embeddings we compute against, the harder it will be to achieve this orthogonality while having nontrivial masks. Similarly, a too-small batch could make it too easy to achieve this orthogonality in a high-dimensional embedding space while not producing useful masks. Is there a sweet spot or range sizes where this seems to work (on CLAP)? - What happens if the third term is left out of the loss? Similarly, how sensitive is this term to size and diversity of the batch? A couple of minor questions: - Line 155 describes the audio reconstruction by ISTFT. It's not clear how phase is treated here, since $X$ appears to be magnitude spectra. Presumably phase is retained from the STFT and propagated through, but if some other method is used here (eg griffin-lim or the like) it should be made explicit. - In general, the spectrogram visualizations (eg figure 4) are quite difficult to read in this paper. It would be helpful to A) label the axes with appropriate units, B) use a log-scale frequency axis for display (even if the underlying method uses linear spectra, the results will be more visible), and C) use a log amplitude scaling on the magnitudes to enhance visual contrast.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The limitations and impacts are well and appropriately described by the authors.

Summary

The paper introduces a post-hoc interpretation method for zero-shot audio classifiers, named LMAC-ZS (Listenable Maps for Audio Classifiers in the Zero-Shot context). It addresses the challenge of interpreting predictions from zero-shot audio classifiers that define audio classes based on textual prompts, where labels are not predefined but generated dynamically. LMAC-ZS outputs saliency maps that highlight important regions within input audio, correlating these with corresponding text prompts. The method involves a novel loss function that maintains the original audio-text similarity, enhancing interpretability. Experiments were conducted using the CLAP model on standard datasets like ESC50 and UrbanSound8K.

Strengths

1. Motivation is Clearly articulated. It stresses the importance of interpretability in AI, particularly for models used in critical decision-making areas, focusing on the less-explored zero-shot audio classification. 2. The problem of providing interpretable explanations for zero-shot audio classifiers is clearly formulated and identified as a novel challenge in the field. 3. The paper compares LMAC-ZS with several baseline methods like GradCAM++, SmoothGrad, and Integrated Gradients, providing a thorough comparative analysis and justifying the selection based on their relevance and common use in related tasks.

Weaknesses

1. It is observed that in some scenarios LMAC-ZS (CT) performs better than LMAC-ZS (FULL) in metrics such as AI, AD, and AG. This is counterintuitive, as one would expect the model trained on the full CLAP dataset to perform better. An explanation for this discrepancy is needed. 2. The paper mentions exploring the training of LMAC-ZS only on the Clotho dataset to simulate a limited computational budget scenario. It raises the question of whether training on other individual datasets or comparing the performance across different single datasets versus the full dataset would yield different insights. This aspect needs further exploration and clarification. 3. In Table 2, the use of ESC50 contamination for ZS classification on the ESC50 dataset seems confusing. The rationale behind using the same dataset for contamination needs to be explained in detail. Similarly, in Table 3, the reason for using US8K contamination for Mel-Masking and ESC50 contamination for STFT-Masking in ZS classification on US8K requires clarification.

Questions

1. In Figure 7, the visual explanations for 'Toilet flushing' and 'Water drops' look very similar despite different similarity scores (Sim=0.69 and Sim=0.14). How does the model ensure distinguishable and meaningful explanations, and can more distinct examples be provided to demonstrate the method's effectiveness? 2. I have a question about a set of textual prompts. What happens if there are sounds that are not in the text?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The author has addressed the limitations.

Summary

In this paper, the authors focus on interpreting the decisions of zero-shot audio classifiers, particularly ones based on the Contrastive Language-Audio Pretraining (CLAP) model. To achieve this, the authors propose to learn a decoder that predicts a "listenable" audio saliency map (a mask on the input spectrogram) from an audio-text pair. The decoder is trained with a novel loss function that remains faithful to the characteristics of CLAP features. The authors demonstrate the effectiveness of their approach with extensive experiments across various datasets.

Strengths

### Strengths 1. Although there is a significant body of research focusing on interpretability of audio classifiers, interpretability of *zero-shot* audio classifiers (and zero-shot classifiers in general) is an under-explored area, which certainly merits more investigation. 2. From a technical perspective, the proposed approach is sound and has no glaring weaknesses. 3. The qualitative and quantitative results show that the method indeed works as claimed. - Particularly the anonymized qualitative samples provided are highly appreciated.

Weaknesses

### Weaknesses 1. Even after reading two out of the three subsections in Section 2: Methodology, a reader learns nothing new from the paper. To be more specific: - Section 2.1 provides a high-level overview of "Contrastive Language Audio Pretraining" (CLAP) [10], while Section 2.2 gives a brief summary of the approach of the paper "Listenable Maps for Audio Classification" (L-MAC) [8]. Both these sections discuss the previous papers as is, without any new perspective. - It is solely Section 2.3 that pertains to the method proposed by the paper. Even in Section 2.3, the sub-subsection @line-152, **"Producing Listenable Interpretations"** is not something new achieved by the proposed method, but rather a feature of the previous L-MAC paper [8]. - Notably, [8] also has an explicit sub-section with the *exact same title*; **Section 2.2: Producing Listenable Interpretations**. The authors make no effort in lines 152-156 to clarify that the "listenable" feature comes from [8]; thus this is not only redundant, but also very close to plagiarism. - **Suggestion:** I strongly suggest that Sections 2.1, 2.2, and the sub-subsection (@lines 152-156) be discussed separately as a "Background" or "Preliminaries" section, instead of Methodology. Specifically, since the proposed approach builds significantly upon L-MAC [8], it is necessary to disentangle your contributions from that of the authors of [8]. 2. The technical novelty is limited by the L-MAC paper. - From my understanding, the only novelty in the method is the loss function in Equation 6. And even that is mostly changing the cross entropy-based objectives in L-MAC (Equation (2) in [8]) to contrastive losses characteristic of CLAP. Overall, the same min-max objective is retained, and the same regularization term is added for mask sparsity. The same general framework from L-MAC is followed: a decoder is trained on the same dataset as the encoder to predict a mask on the input audio spectrogram. 3. (minor weakness) Comparison against baselines. - The authors are comparing against the baselines of GradCAM, GradCAM++, SmoothGrad, Integrated Gradients. While all these methods were also used for comparison with the original L-MAC in [8], that was in a regular classification setting; these methods are, by nature, poorly suited to zero-shot settings. - For instance, for both "cat" and "glass-breaking", the CLAP model needs to "look" at the important regions of the audio to give an encoded audio feature---which can be observed in Figure 4 (for GradGAM) in the paper. - **Suggestion**: A better baseline would perhaps be from the paper "gScoreCAM: What objects is CLIP looking at?" [a]. The paper empirically establishes that for CLIP, in zero-shot settings, ScoreCAM performs better than the other forms of CAM. - *Note:* I understand that in general, there is a lack of baselines to compare to (for instance, listen-to-interpret and L-MAC are not applicable). I do not expect the authors to conduct any additional experiments. --- In general, the paper has a significant amount of similarity/redundancy with the L-MAC paper [8], from the naming of sections and content (noted in Weakness 1), figures (Fig 1 in [8] and fig 2 in current paper), and even the exact set of metrics posed in the identical order in Section 3.1, which makes it difficult to discern the novelty and contributions. --- [a] Chen, Peijie, et al. "gscorecam: What objects is clip looking at?." Proceedings of the Asian Conference on Computer Vision. 2022.

Questions

### Questions 1. **line 147** *"The intuition is that the similarity between two text prompts should be reflected in the similarity of the audio embeddings from the corresponding masked spectrograms"*. Equation (7) is the only part of the loss function that is distinctly novel, as it adds a third term to the original loss in [8]. Have you tried training the overall decoder without the third term? It would be helpful to the paper if you can show that adding the third term yields a noticeable improvement over just using the first two terms; otherwise, the hypothesis remains unsupported by evidence. 2. It is common to perform zero-shot classification with foundation models trained on a large dataset. Suppose we have one such hypothetical model, trained on a dataset with millions of samples. - Is it necessary to train the decoder on the same large dataset? If not, can an estimation be made as to what percentage of the training data the decoder needs to see to achieve reliable performance? (For instance, it may be the case that after seeing 20% of the data, the model achieves 80% of its full performance). - If it is necessary to train the decoder on the pre-training dataset, using LMAC-ZS becomes expensive in many cases. Can LMAC-ZS be jointly learned during the pre-training process of the base model (e.g. CLAP)? - Note that this may not be so straightforward, as the first term in the loss objective will try to match the decoder predictions to faulty entries of $C$ at the start of the training. ### General Suggestions - **line-114**; **line-122**: The authors note that they omitted a part of Equation (4) for brevity. I strongly suggest against doing this; *clarity* is more important than brevity when posing a loss function or optimization objective. Equation (4) is supposed to be a min-max objective; the goal is to maximize the classification confidence of the masked-in (salient) part of the audio, while minimizing the confidence of the masked-out part. Thus the whole equation should be written together, and parts of it should not be omitted. If brevity is desired, it may be achieved by abbreviating $\text{CrossEntropy}$ to $CE$ or $\mathcal{L}_{CE}$. - **line 120:** The abbreviation L-MAC is used, but it has not been defined previously. The authors should define the abbreviation on line 70. - **line 62-63:** The citation style is inconsistent. For example, in lines 59-60, it has been written "Key approaches in this category include [19, 20, 21, 22],". The same citation style can be followed in line 64: "Notable attempts in this vein include [23, 24, 25]." --- [b] Shimada, Kazuki, et al. "Zero-and Few-Shot Sound Event Localization and Detection." ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024.

Rating

6

Confidence

4

Soundness

3

Presentation

1

Contribution

2

Limitations

### Limitations - One limitation that I feel is not addressed is that the approach needs to be specifically trained on CLAP's data; it is not plug-and-play like Grad-CAM or similar approaches. - In **line 216** It is mentioned that the decoder uses CNN14 layers, presumably because the audio part of CLAP is based on CNN14. - Now if we have another zero-shot audio classifier, LAION-CLAP, that (suppose) has the same dataset but a different transformer-like architecture, then the decoder may not be transferable. Another decoder architecture needs to be designed to suit the alternative zero-shot classifier. So for different zero-shot foundation models, it may become necessary to have different architectures, which can be a hurdle. - As noted earlier, for same architecture but different datasets, it is still needed to train the decoder again. Unless of course, the decoder from one pre-training dataset transfers to another dataset, as a general purpose audio method. Apart from these scalability concerns, I believe the remaining limitations are adequately addressed. --- Update: The rebuttal addresses most of my concerns, so I will raise my score to 6.