PaperlayerView exact revision on arXiv

Source-linked AI summary

Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification

Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, Mark Yatskar

arXiv:2211.11158v2cs.CVcs.CL

TL;DR

Concept Bottleneck Models require costly manually specified concepts and often underperform black-box models. LaBo uses GPT-3 to generate and select diverse discriminative concepts, aligns them to images with CLIP, and achieves stronger few-shot accuracy while retaining the CBM abstraction. Its limitations include dependence on GPT-3 knowledge, CLIP recognition, and unreproducible large-model training.

  • Problem

    Concept Bottleneck Models require costly manually specified concepts and often underperform black-box counterparts, limiting their broader adoption.

  • Method

    LaBo prompts GPT-3 for candidate concepts, selects diverse and discriminative bottlenecks with submodular optimization, and aligns them to images using CLIP.

  • Results

    LaBo outperforms linear probes by up to 11.7% at 1-shot, averages 1.5% higher accuracy across dataset sizes, and remains competitive with more data on 11 datasets.

  • Takeaways & Limitations

    Automatically designed CBMs can approach or exceed black-box performance while maintaining factors contributing to interpretability.

  • Takeaways & Limitations

    LaBo depends on GPT-3 generating useful facts and CLIP recognizing them, an alignment that may not hold universally; its large-model training is not currently reproducible.

Abstract

from arXiv · show

Concept Bottleneck Models (CBM) are inherently interpretable models that factor model decisions into human-readable concepts. They allow people to easily understand why a model is failing, a critical feature for high-stakes applications. CBMs require manually specified concepts and often under-perform their black box counterparts, preventing their broad adoption. We address these shortcomings and are first to show how to construct high-performance CBMs without manual specification of similar accuracy to black box models. Our approach, Language Guided Bottlenecks (LaBo), leverages a language model, GPT-3, to define a large space of possible bottlenecks. Given a problem domain, LaBo uses GPT-3 to produce factual sentences about categories to form candidate concepts. LaBo efficiently searches possible bottlenecks through a novel submodular utility that promotes the selection of discriminative and diverse information. Ultimately, GPT-3's sentential concepts can be aligned to images using CLIP, to form a bottleneck layer. Experiments demonstrate that LaBo is a highly effective prior for concepts important to visual recognition. In the evaluation with 11 diverse datasets, LaBo bottlenecks excel at few-shot classification: they are 11.7% more accurate than black box linear probes at 1 shot and comparable with more data. Overall, LaBo demonstrates that inherently interpretable models can be widely applied at similar, or better, performance than black box approaches.

1. Introduction

LaBo automatically constructs interpretable Concept Bottleneck Models by combining GPT-3-generated concepts with CLIP-based image alignment. Across diverse image-classification tasks, it improves few-shot performance while preserving the CBM abstraction and producing understandable, factual bottlenecks.

  • Motivation: CBMs improve transparency by exposing human-understandable concepts, but costly expert annotations and lower performance limit their application.These limitations motivate automatically constructing CBMs without sacrificing interpretability.
  • Approach: LaBo prompts GPT-3 to generate textual candidate concepts, then uses CLIP to score their presence in images and form the bottleneck layer.The concepts are factual sentences elicited from language-model knowledge and aligned with images through CLIP.
  • Approach: LaBo selects concepts with a submodular criterion that favors discriminative, diverse, recognizable, short, and class-name-free attributes.The criterion provides efficient control over bottleneck composition and targets both classification accuracy and interpretability.
  • Results: 11.7% higher accuracy at 1-shot and 1.5% higher average accuracy than linear probes demonstrate LaBo’s strongest gains in low-data regimes.With larger data settings, LaBo marginally underperforms linear probes while remaining competitive; the evaluation spans 11 diverse datasets.
  • Results: Human evaluations find LaBo bottlenecks largely understandable, visual, and factual, with GPT-3-derived concepts more factual and groundable than WordNet- or Wikipedia-based alternatives.The findings support automatically designed CBMs as effective models that retain important interpretability properties.

2. Related Work

Prior interpretable-by-design methods include post-hoc explanations, prototypes, and Concept Bottleneck Models. LaBo replaces limited external knowledge bases with language-model-generated concepts and achieves strong few-shot performance without breaking the CBM abstraction.

  • Interpretability methods: Interpretability methods are broadly divided into post-hoc approaches and models designed to align explanations with their computations.Post-hoc methods do not impose model constraints, but their explanations are not guaranteed to faithfully represent reasoning.
  • Concept Bottleneck Models: Concept Bottleneck Models predict targets from human-understandable intermediate attributes but require costly annotations and often perform worse than end-to-end models.Recent variants address these gaps partly through external knowledge bases or residual predictors, while retaining limitations.
  • Concept Bottleneck Models: Post-hoc Concept Bottleneck Models use static knowledge bases and residual connections, but limited knowledge-base coverage restricts larger-scale and domain-specific applications.These modifications can also weaken the bottleneck abstraction by adding a residual predictor.
  • LaBo: LaBo queries LLMs instead of external knowledge bases and automates extraction and filtering of concepts for interpretable classification.It surpasses end-to-end models in few-shot scenarios and achieves comparable performance with larger data settings.
  • LaBo: LaBo can be viewed as interpretability-focused prompt tuning of CLIP, unlike prompting methods primarily designed to improve classification performance.This framing connects LaBo to broader work on prompting vision-language models.

3. Method

LaBo constructs interpretable concept bottlenecks by generating textual concepts with GPT-3, selecting discriminative and diverse subsets, and aligning them to images with CLIP. A linear predictor then maps concept scores to class predictions, with language priors optionally initializing concept weights.

  • Concept Bottleneck Construction: GPT-3 generates candidate textual concepts, which CLIP aligns with images to form the bottleneck embedding.Each selected concept is represented by its CLIP text feature.
  • Submodular Concept Selection: Submodular optimization selects a fixed number of class-specific concepts by jointly rewarding discriminability and coverage.The objective combines discriminability, which favors class-specific concepts, with coverage, which promotes diversity across candidate concepts.
  • Submodular Concept Selection: Discriminability favors concepts aligned with many images from the target class but few images from other classes.The similarity score is computed from mean image-text dot products, then normalized across classes.
  • Submodular Concept Selection: Coverage uses concept-feature similarity to select a diverse subset that represents varied appearances of a target class.The coverage term is implemented as a minimax facility-location objective over candidate concepts.
  • Class-concept Association: The concept predictor uses image-concept dot products, while a learned linear weight matrix maps concept scores to final class predictions.Softmax normalization is applied along the concepts axis of the weight matrix.
  • Class-concept Association: Language priors initialize weights higher for class-concept pairs selected in the bottleneck before optimization.This initialization is intended for extending concept bottleneck models to few-shot settings.

4. Experimental Setup

The evaluation uses 11 diverse image-classification datasets and few-shot protocols with 1, 2, 4, 8, and 16 labeled images per class, plus fully supervised training. LaBo is compared with linear probing and interpretable CBM baselines under test-accuracy evaluation.

  • Datasets and Protocol: The benchmark spans 11 datasets covering common objects, fine-grained objects, actions, textures, skin tumors, and satellite images.All datasets use train, development, and test splits.
  • Datasets and Protocol: Few-shot evaluation samples 1, 2, 4, 8, or 16 training images per class, with an additional fully supervised setting.Test accuracy is reported for all experiments.
  • Baselines: LaBo is compared with black-box linear probing and two interpretable methods, PCBM and CompDL.Linear probing uses L-BFGS logistic regression with an L2-regularization sweep.
  • Implementation: The default CLIP backbone is ViT-L/14, with CLIP-RN50 and ViT-B/32 used for fair comparisons with PCBM and CompDL, respectively.GPT-3-text-davinci-002 generates the concepts, and hyperparameters are tuned on the development set.

5. Evaluation

Across 11 datasets, LaBo remains competitive with black-box linear probes and improves interpretable CBMs through language-prior initialization and submodular concept selection. Human evaluations find its bottlenecks largely factual, groundable, and understandable.

  • Main Results: LaBo surpasses the linear probe by 1.5% on average across 11 datasets and is especially strong when little data is available.It significantly outperforms the end-to-end model in low-data settings and remains competitive as data increases.
  • Main Results: LaBo outperforms PCBM by 3.4% on CIFAR-10 and 13.1% on CIFAR-100 while matching PCBM-h without circumventing the bottleneck.It is also more accurate than CompDL without manually constructed concepts.
  • Ablation Study: The submodular function achieves the best performance across shot settings by jointly optimizing concept coverage and discriminability.Coverage or discriminability alone still outperform similarity-based and random selection, while selection matters less with more supervision.
  • Ablation Study: Language-model initialization is more important in low-shot settings because less data provides signal for guiding concept importance.The ablation compares language-prior and random initialization.
  • Ablation Study: Larger bottlenecks are usually better, whereas smaller bottlenecks can achieve similar performance with more data.The bottleneck-size ablation ranges from 1 to 50 concepts and averages performance over 11 datasets.
  • Human Evaluation: Human evaluations show that LaBo has higher factuality and groundability than human-written text and fewer invalid concepts than other baselines.Removing the submodular function or language prior hurts both human-evaluation metrics, while the full bottlenecks are largely coherent, factual, and groundable.

6. Conclusion and Limitation

The paper concludes that accuracy and interpretability may be less opposed than previously believed, while identifying limits imposed by GPT-3 knowledge, CLIP alignment, and reliance on large unreproducible models.

  • Conclusion: LaBo suggests that vision systems can achieve accuracy and interpretability together by using language models to provide visual knowledge.The authors propose enriching bottleneck construction with additional priors in future work.
  • Limitation: Low-resolution CIFAR images partially affect human-evaluation metrics because annotators have greater difficulty completing the task.This caveat concerns the reported human-evaluation results rather than classification accuracy.
  • Limitation: LaBo depends on GPT-3 generating useful, visually relevant knowledge, which may fail for some domains such as fine-grained datasets.The paper suggests specialized language models could improve performance in such cases.
  • Limitation: The method also depends on CLIP recognizing the aspects described by GPT-3, and that alignment may not hold in all cases.The authors identify dynamically prompting GPT-3 as a possible way to make the coupling more robust.
  • Limitation: The approach relies on large models trained at scales that are not currently reproducible, and undisclosed training details could affect the claims.This is presented as a possible reason for reevaluating the results.

A. Dataset Statistics

The benchmark covers 11 image-classification datasets spanning common and fine-grained objects, actions, textures, skin tumors, and satellite images, using dataset-specific train/dev/test splits.

  • Splits: The evaluation uses train/dev/test splits for all datasets, with splits adopted or sampled according to each dataset’s setup.ImageNet is evaluated only on its development set.

B. Implementation Details

The implementation evaluates LaBo against linear probes across 11 datasets, using GPT-3 prompts, concept extraction and filtering, and tuned training procedures. Frozen CLIP features and a single linear layer keep training efficient.

  • Prompting: GPT-3 prompts are shared across datasets with dataset-specific modifications, including super-class names for finer-grained categories.Super-class prompts reduce ambiguity but do not eliminate errors such as confusing the ImageNet animal mouse with a computer device.
  • Training and tuning: The method uses the CLIP-RN50 linear-probe comparison protocol and tunes submodular-function weights with five runs across datasets and shot settings.Hyperparameters are selected on development data, with the complete settings listed in Table 16.
  • Concept processing: T5 extracts shorter concepts from GPT-3 sentences, after which heuristics replace detected class names with dataset-specific super-class names.The extraction training uses T5large for five epochs with batch size 8 and learning rate 1e−5.
  • Evaluation setup: LaBo is compared with linear probes on development and test sets across 11 datasets.The full results report both validation and test accuracy.
  • Cost: Generating 500 GPT-3 sentences for one class takes approximately 5 minutes and costs about $0.2 per class.The reported GPT-3-Davinci price is $0.02 per 1k tokens.
  • Runtime: Frozen CLIP image features are extracted once, and training a single linear layer for full ImageNet takes less than 1 minute per epoch on an NVIDIA RTX A6000.Feature reuse and the single-layer classifier reduce training time.

C.1. Activation Function

The activation-function ablation finds that softmax is important for LaBo performance, while larger GPT-3 models help most when data are scarce. With enough data, model-size differences become marginal.

  • Embedding analysis: The concept and class-name embeddings are visualized with t-SNE using each bottleneck’s top-1 concept per ImageNet class.GPT-3, WordNet, and Wikipedia bottlenecks are compared.
  • Activation function: Removing the activation function significantly hurts LaBo performance, while ReLU and sigmoid perform poorly compared with softmax.The comparison reports mean accuracy across the 11 datasets.
  • Language model size: Larger GPT-3 models improve performance, especially in few-shot settings, but yield only marginal gains when enough data are available.The comparison uses Curie, Babbage, and Ada on ImageNet.

C.3. Performance of Human-Written Text

GPT-3-generated concepts outperform WordNet- and Wikipedia-sourced concepts in one-shot settings while retaining nearly comparable performance with more data. Their qualitative advantage is linked to describing classes rather than repeating class-name features.

  • Human-written text comparison: GPT-3 concepts outperform WordNet- and Wikipedia-based concepts in 1-shot experiments, with less than 1% average performance drop in larger-data settings.Human evaluation on ImageNet also judges GPT-3 concepts higher quality than human-designed concepts.
  • Embedding analysis: WordNet and Wikipedia concepts overlap more with ImageNet class-name embeddings than GPT-3 concepts do.The analysis uses t-SNE over class names and each bottleneck’s top-1 concept.
  • Interpretability: The paper attributes the lower interpretability of human-written text to its tendency to replicate class-name text features rather than describe the class.This explains why human-written text can have higher accuracy while being less interpretable.
  • Prompt-tuning comparison: LaBo’s performance is similar to CoOp without class names and exceeds CoOp on Aircraft and UCF-101 when class names are added.The comparison covers five datasets at 16 shots using CLIP-ViT-B/32.

D. Human Evaluation

Human evaluation measures the factuality and groundability of LaBo concepts through controlled image-selection tasks. The results indicate that the bottlenecks are generally understandable, visual, and factual, with resolution affecting some judgments.

  • Evaluation dimensions: Factuality and Groundability evaluate bottleneck concepts along two dimensions using human annotations.The study introduces both metrics to identify possible improvement areas.
  • Annotator statistics: Each concept receives three annotations from a diverse pool of 477 Mechanical Turk annotators, with 69.83% average pairwise agreement before preprocessing.The annotation cost was $2,100, and annotators were paid an average of $14.5 per hour.
  • Annotation interface: Annotators select concept-matching images from 12 choices containing 10 target images and 2 randomly sampled control images.Groundability instead presents the dataset’s top-10 images ranked by CLIP similarity.
  • Validity rules: Annotations selecting control images are excluded from Factuality and Groundability unless all annotators fail the control, while invalidity requires agreement from at least two of three annotators.18% of annotations are disregarded for control-image selection.
  • Invalid concepts: CIFAR-10 and CIFAR-100 have a high percentage of non-visual concepts, which the authors hypothesize reflects low image resolution rather than absent visual content.Invalid-concept distributions are reported for all datasets except ImageNet.
Loading 2211.11158v2…