[论文解读] Label-Free Concept Bottleneck Models
论文提出 Label-free CBM,将任何神经网络转变为可解释的概念瓶颈模型,而无需标记的概念,同时保持高准确性,具备对 ImageNet 的可扩展性,并实现事后可解释性与编辑能力。
Concept bottleneck models (CBM) are a popular way of creating more interpretable neural networks by having hidden layer neurons correspond to human-understandable concepts. However, existing CBMs and their variants have two crucial limitations: first, they need to collect labeled data for each of the predefined concepts, which is time consuming and labor intensive; second, the accuracy of a CBM is often significantly lower than that of a standard neural network, especially on more complex datasets. This poor performance creates a barrier for adopting CBMs in practical real world applications. Motivated by these challenges, we propose Label-free CBM which is a novel framework to transform any neural network into an interpretable CBM without labeled concept data, while retaining a high accuracy. Our Label-free CBM has many advantages, it is: scalable - we present the first CBM scaled to ImageNet, efficient - creating a CBM takes only a few hours even for very large datasets, and automated - training it for a new dataset requires minimal human effort. Our code is available at https://github.com/Trustworthy-ML-Lab/Label-free-CBM. Finally, in Appendix B we conduct a large scale user evaluation of the interpretability of our method.
研究动机与目标
- 自动创建可解释的概念瓶颈模型,而无需标记的概念数据。
- 实现对大规模数据集(包括 ImageNet)的可扩展性与高效训练。
- 保持或尽量接近原骨干网络的准确性。
- 提供全局和局部的决策解释能力,并展示模型编辑的潜力。
提出的方法
- 将任何神经网络骨干网转换为 Concept Bottleneck Model (CBM),无需标记概念。
- 步骤1:使用 GPT-3 以及多项过滤标准生成并筛选初始概念集。
- 步骤2-3:通过优化投影 Wc 来最大化 activations 与目标概念之间基于 CLIP-Dissect 的相似度(使用 cos cubed 相似度度量)来学习 Concept Bottleneck Layer (CBL)。
- 步骤4:用弹性网正则化训练稀疏的最终层 WF 以产生预测。
- 训练设计为自动化、可扩展且高效,最终模型呈现可解释性的稀疏性。
- 该方法支持大规模数据集,并在 CIFAR-10/100、CUB200、Places365 和 ImageNet 上进行评估。
实验结果
研究问题
- RQ1完全自动化、无标签过程是否能够在不使用概念注释的情况下产生可解释的概念瓶颈?
- RQ2标记无关的 CBM 是否能保持与原骨干网络相当的准确性,包括在 ImageNet 规模数据上的表现?
- RQ3相较于传统 CBM,最终决策规则的可解释性和每个样本的解释有多大程度的可读性?
- RQ4是否可以对最终层进行稀疏化以提高可解释性,同时不牺牲性能?
- RQ5是否存在手动编辑最终层权重以纠正错误并提高整体准确性的潜力?
主要发现
| 数据集 | 模型 | 稀疏最终层 | CIFAR-10 | CIFAR-100 | CUB-200 | Places365 | ImageNet | ||
|---|---|---|---|---|---|---|---|---|---|
| CIFAR-10 | Standard | No | 88.80%* | 2 datasets provided in table show multiple entries; the label-free CBM row is presented separately below. | |||||
| CIFAR-10 | Standard (sparse) | Yes | 82.96% | ||||||
| CIFAR-10 | P-CBM | Yes | 70.50%* | ||||||
| CIFAR-10 | P-CBM (CLIP) | Yes | 84.50%* | ||||||
| CIFAR-10 | Label-free CBM (Ours) | Yes | 86.40% | ±0.06% | ? | ? | ? | 71.95% | ±0.05% |
| CIFAR-100 | Standard | No | 70.10%* | ||||||
| CIFAR-100 | Standard (sparse) | Yes | 58.34% | ||||||
| CIFAR-100 | P-CBM | Yes | 43.20%* | ||||||
| CIFAR-100 | P-CBM (CLIP) | Yes | 56.00%* | ||||||
| CIFAR-100 | Label-free CBM (Ours) | Yes | 65.13% | ±0.12% | ? | ? | ? | ||
| CUB-200 | Standard | No | 76.70% | 74.31% | ±0.29% | ||||
| CUB-200 | Standard (sparse) | Yes | 75.96% | ||||||
| CUB-200 | P-CBM | Yes | 59.60%* | ||||||
| CUB-200 | P-CBM (CLIP) | Yes | N/A | ||||||
| CUB-200 | Label-free CBM (Ours) | Yes | 74.31% | ±0.29% | ? | ? | ? | ||
| Places365 | Standard | No | 48.56% | 43.68% | ±0.10% | ||||
| Places365 | Standard (sparse) | Yes | 38.46% | ||||||
| Places365 | P-CBM | Yes | N/A | ||||||
| Places365 | P-CBM (CLIP) | Yes | N/A | ||||||
| Places365 | Label-free CBM (Ours) | Yes | 43.68% | ±0.10% | ? | ? | ? | ||
| ImageNet | Standard | No | 76.13% | 71.95% | ±0.05% | ||||
| ImageNet | Standard (sparse) | Yes | 74.35% | ||||||
| ImageNet | P-CBM | Yes | N/A | ||||||
| ImageNet | P-CBM (CLIP) | Yes | N/A | ||||||
| ImageNet | Label-free CBM (Ours) | Yes | 71.95% | ±0.05% | ? | ? | ? |
- 标记无关的 CBM 在五个数据集上取得具有竞争力的准确性,其中 ImageNet 上的模型在最终层稀疏的情况下达到 72% 的 top-1 精度。
- 该方法在评估数据集上优于后验 CBM,同时保持完全可解释(没有不可解释的残留项)。
- 最终层权重可视化以揭示可解释的决策规则,全球模式与直观概念(如柑橘类水果与橙子和柠檬相关)一致。
- 通过简单线性地将概念贡献可用于解释单个预测,便于直观可视化特征贡献。
- 对最终层权重进行人工编辑,在数据子集上可带来准确性提升,对整个网络影响较小,展示了实际调试潜力。
更好的研究,从现在开始
从论文设计到论文写作,大幅缩短您的研究时间。
无需绑定信用卡
本解读由 AI 生成,并经人工编辑审核。