🤖 machine learning

Spatially Grounded Concept-Based Image Classification

The paper proposes SEG-MIL-CBM, a spatially grounded Concept Bottleneck Model that decomposes images into concept-guided regions and aggregates segment-level evidence via attention to simultaneously improve classification accuracy and provide intrinsic, human-interpretable explanations without requiring separate post-hoc attribution modules.

Original authors: Ran Eisenberg, Amit Rozner, Ethan Fetaya, Ofir Lindenbaum

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ran Eisenberg, Amit Rozner, Ethan Fetaya, Ofir Lindenbaum

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: What is in this picture?

Most modern AI detectives (Deep Neural Networks) are incredibly good at solving the case. They can tell you, "That's a bird!" with 99% accuracy. But they are terrible at explaining how they know. They might be looking at the background (like a lake) instead of the bird itself, or they might be using a secret code that no human can read. If you ask, "Why did you say it's a bird?", they just shrug and say, "Because the math says so."

Other AI detectives, called Concept Bottleneck Models (CBMs), try to be more honest. They say, "I see a 'beak' and 'feathers', so it must be a bird." This is better, but they still have a flaw: they treat the whole picture like a smoothie. They blend all the "beaks" and "feathers" together into one big cup of information. They can't tell you which specific part of the picture they are looking at. Did they see the beak on the left? Or the tail on the right? They just give you a global score.

Enter the New Detective: SEG-MIL-CBM

The authors of this paper built a new detective called SEG-MIL-CBM. Think of this detective as a forensic accountant who doesn't just guess; they keep a detailed, itemized receipt for every decision they make.

Here is how it works, using simple analogies:

1. The "Cut-and-Paste" Team (Preprocessing)

Before the detective even looks at the photo, they use a team of specialized robots (pre-trained AI tools like CLIP, GroundingDINO, and SAM) to cut the image into puzzle pieces.

The Robots: They look at the image and say, "I see a patch that looks like a 'bright orange breast' here," and "I see a patch that looks like 'black wings' there."
The Result: The image is no longer one big blob. It's a bag of distinct, labeled puzzle pieces (segments).

2. The "Committee Vote" (The Model)

Now, the main detective (the model) looks at this bag of puzzle pieces. Instead of blending them, it treats each piece like a witness in a courtroom.

The Witnesses: Each puzzle piece (segment) says, "I am a piece of the bird, and I contribute 40% to the 'Bird' verdict." Another piece says, "I am a piece of the background, and I contribute 0%."
The Vote: The detective weighs these testimonies. It uses a special "attention" mechanism to listen more closely to the pieces that seem most important and ignore the noise.

3. The "Itemized Receipt" (The Explanation)

This is the magic part. Because the final answer is just a sum of the witnesses' votes, the detective can print out an itemized receipt.

The Receipt: It doesn't just say "Bird." It says:
- Witness 1 (Orange Breast): +0.42 points to "Bird."
- Witness 2 (Black Wings): +0.32 points to "Bird."
- Witness 3 (Blue Sky): 0 points.
The Benefit: You can see exactly where the AI looked and what it saw. If the AI made a mistake, you can look at the receipt and say, "Ah, you focused on the wrong piece of the puzzle!"

Why Does This Matter? (The "Shortcut" Problem)

Sometimes, AI gets lazy. It learns "shortcuts."

The Scenario: Imagine an AI trained to spot birds. It notices that in most training photos, birds are on water. So, it learns: "If I see water, it's a bird."
The Failure: If you show it a bird on a tree, it might get confused because it's looking for water.
The Old Way: Global CBMs might still get tricked because they blend the "water" and the "bird" into one score.
The SEG-MIL-CBM Way: Because this model looks at pieces separately, it can see: "This piece is a bird on a tree (Good!), but this piece is water (Bad/Shortcut)." It can choose to ignore the water piece and focus on the bird piece.

What Did They Prove?

The authors tested this new detective on several challenges:

It's Honest: They ran tests where they removed the "most important" puzzle pieces one by one. The model's confidence dropped exactly as expected, proving the explanation matches the thinking process.
It's Smart: It didn't just get better at explaining; it actually got better at solving hard cases where other "honest" models failed (specifically on datasets where the AI usually gets tricked by backgrounds).
It's Fast Enough: It works well on standard image recognition tasks, not just the tricky ones.

The Bottom Line

This paper introduces a system that forces AI to show its work. Instead of a black box that gives an answer, it gives you a highlighted map of the image with a scorecard for every highlighted area.

Old AI: "It's a bird." (Trust me, I'm smart.)
Old "Honest" AI: "It's a bird because of concepts like 'feathers' and 'beak'." (But I can't tell you where I saw them.)
SEG-MIL-CBM: "It's a bird. Here is the 'feather' patch on the left (Score: 0.4), and here is the 'beak' patch on the right (Score: 0.3). I ignored the water in the background."

The authors claim this makes the AI more reliable, especially when it might be tempted to take a lazy shortcut, and it allows humans to audit the decision without needing a separate, confusing tool to explain it.

Technical Summary: SEG-MIL-CBM

Problem Statement
Deep neural networks often achieve high accuracy by relying on spurious correlations or evidence that is difficult to inspect, particularly in high-stakes settings. While Concept Bottleneck Models (CBMs) offer human-interpretable reasoning by mapping images to human-defined concepts, most existing CBMs treat concepts as global attributes. This global approach fails to identify which specific image regions support a concept, leaving a gap between the prediction and its explanation. Furthermore, spatially grounded CBMs (e.g., SALF-CBM) localize evidence but do not explicitly account for how region-level evidence is aggregated into the final decision logit, often resulting in explanations that are separate from the computation producing the prediction.

Methodology: SEG-MIL-CBM
The authors propose SEG-MIL-CBM (Spatially Grounded Concept-Based Image Classification), a model designed to decompose images into concept-guided regions and classify them via attention-based aggregation of segment-level evidence. The core innovation is that the same segment evidence terms used to form the prediction are directly read to generate the explanation, eliminating the need for a separate post-hoc attribution module.

The pipeline consists of four stages:

Offline Preprocessing: Using pre-trained foundation models (CLIP, GroundingDINO, and SAM), the system generates a "bag" of concept-guided segments for each image. GroundingDINO localizes concepts, and SAM produces binary masks. These segments are cached along with frozen CLIP concept-similarity targets.
Segment Encoding: A trainable model encodes each segment using a shared backbone ( $\phi$ ) to produce features ( $h_i$ ). A linear concept head projects these features into a concept space to yield learned concept activations ( $z_i$ ).
Accountable Aggregation (MIL): The model employs an attention-based Multiple Instance Learning (MIL) framework. It computes class-conditioned attention weights ( $\alpha_{i,y}$ ) for each segment and aggregates segment-class scores ( $g_{i,y}$ ) into an image-level logit ( $\ell_y = \sum \alpha_{i,y} g_{i,y}$ ).
Explanation Generation: The final logit is a sum of explicit segment-level contributions. The explanation is derived by reading the non-bias evidence terms ( $C^{sp}_{i,y}$ ) from this decomposition. This reveals which regions contributed positively to the prediction and which specific concepts within those regions supplied that support.

The training objective combines image-level cross-entropy loss with a cosine alignment loss between the learned concept activations and the frozen CLIP targets, ensuring semantic consistency without requiring explicit concept or group labels during training.

Key Contributions

Accountable Aggregation: The authors formulate spatial concept prediction such that every displayed region score is read directly from the non-bias segment evidence term in the logit decomposition, creating a unified prediction and explanation.
Faithfulness Validation: Through segment-level deletion and insertion experiments on the CUB dataset, the paper quantitatively validates that the learned segment ranking is faithful to the model's own prediction pathway, outperforming controls based on raw CLIP salience, segment area, and random ranking.
Interpretability-Robustness Trade-off: The model demonstrates that an interpretability-first design can preserve competitive recognition accuracy while improving worst-group accuracy over evaluated CBM-family baselines. It explicitly exposes the Pareto trade-off between interpretability and the highest possible worst-group accuracy achievable by non-interpretable group-robust methods.

Results

Group Robustness: On the Waterbirds dataset (a benchmark for spurious background correlations), SEG-MIL-CBM improves worst-group accuracy from 65.1% (SALF-CBM) to 72.0%. On the synthetic Pawrious dataset, it achieves 87.4% worst-group accuracy, significantly outperforming global CBM baselines which fail to mitigate background shortcuts.
Standard Recognition: The model remains competitive on standard recognition tasks, achieving the best CBM accuracy on CIFAR-100 (85.3%) and improving over prior spatial CBMs on Places365 and CUB.
Faithfulness: On CUB, the model achieves the best Deletion/Insertion AUC (DAUC/IAUC: 0.444/0.670) among evaluated controls. Top-1 insertion tests show that keeping only the highest-ranked segments preserves the original prediction in 90.5% of images.
Ablation: Experiments confirm that the performance gains are driven by the explicit aggregation of localized evidence rather than hidden feature bypasses, and that the concept-alignment loss is crucial for maintaining the semantic interface used in explanations.

Significance and Claims
The paper claims that SEG-MIL-CBM provides a stronger form of explanation by serving as an "audit trail" of the prediction itself. Unlike post-hoc methods that attach heatmaps to separate classifiers, SEG-MIL-CBM ensures that the displayed region evidence is intrinsic to the forward computation.

The authors position SEG-MIL-CBM not as a replacement for non-interpretable group-robust methods (which may achieve higher worst-group accuracy on specific datasets like Waterbirds), but as a distinct point on the Pareto frontier. It offers a trade-off where practitioners gain a spatially grounded, concept-level accounting of why a decision was made—specifically identifying which object parts or background regions supported the prediction—while maintaining competitive accuracy and improved robustness over other interpretable baselines. The work emphasizes that this approach makes failures (e.g., reliance on background shortcuts) more inspectable than in global bottleneck models, even if it does not guarantee the removal of all shortcuts.