Enterprise AI Analysis
OCCAM: Open-set Causal Concept explanation and Ontology induction for black-box vision Models
OCCAM introduces a black-box framework for open-set causal concept explanation and ontology induction in vision models. It leverages Multimodal Large Language Models (MLLMs) to propose and ground open-set concepts spatially. By performing object-level deletions and measuring prediction shifts, OCCAM estimates each concept's causal contribution. This interventional evidence is aggregated across datasets to induce a structured conceptual ontology, capturing how classifiers organize visual concepts globally. Reasoning over this ontology reveals stable conceptual dependencies, compositional patterns, and systematic model biases. Empirical results on Broden and ImageNet-S show improved explanation quality in black-box settings and richer global insights compared to per-image attribution methods.
Executive Impact & Key Findings
OCCAM's unique approach delivers measurable improvements in AI interpretability and transparency, crucial for enterprise adoption.
Accuracy Improvement
Explanation Quality
Concept Coverage
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
OCCAM reframes explanation as structured interventional pruning. It uses MLLMs to propose open-set concepts, grounds them spatially via text-guided segmentation, and performs object-level deletions. By measuring prediction shifts, it estimates each concept's causal contribution under query-only access.
99.6%
of images had valid grounded concepts
| Metric | CE-FAM [24] | OCCAM |
|---|---|---|
| Average Drop Percentage (ADP) | 44.18% | 47.19% |
| Maximum Drop Percentage (MDP) | 84.14% | 93.89% |
| Maximum Absolute Logit Drop (MAD) | 3.37 | 4.57 |
Note: OCCAM consistently induces stronger confidence reductions, demonstrating effective identification of decision-critical evidence.
OCCAM aggregates interventional measurements across images to induce a structured conceptual ontology. This ontology reflects how classifiers globally compose, prioritize, and relate concepts, and is not externally imposed but emerges from interventional evidence. Reasoning over this ontology reveals stable conceptual dependencies, compositional patterns, and systematic model biases.
Enterprise Process Flow
Input Image
→
Open-set Concept Proposal (MLLM)
→
Spatial Grounding
→
Object-level Deletion
→
Measure Prediction Shift
→
Aggregate Evidence
→
Induce Ontology
| Judge | LLM | LLM + JSON | LLM + Ontology |
|---|---|---|---|
| ChatGPT 5.2 | 4.60 | 4.20 | 4.80 |
| Gemini 3 | 4.00 | 4.20 | 4.80 |
| Human | 4.55 | 4.33 | 4.70 |
Note: Ontology-grounded reasoning consistently achieves the highest scores across all judge groups, indicating enhanced coherence and faithfulness.
The framework extends naturally to vision-language classifiers like CLIP and SigLIP. Input-level interventions provide a unified explanation mechanism across unimodal and multimodal architectures, confirming its broad applicability without access to internal representations.
SigLIP
more robust than CLIP variants under intervention
| k=0 (Original) | k=1 Removed | k=2 Removed | k=3 Removed |
|---|---|---|---|
| 0.69 | 0.23 | 0.11 | 0.07 |
Note: Accuracy decreases monotonically as concepts are removed, indicating successful identification of causal factors.
Advanced ROI Calculator
Estimate the potential efficiency gains and cost savings by integrating OCCAM's advanced AI explanation capabilities into your enterprise.
Annual Cost Savings
$0
Annual Hours Reclaimed
0
Your Enterprise AI Implementation Roadmap
Our structured approach ensures a smooth integration of OCCAM's capabilities into your existing systems, maximizing impact with minimal disruption.
Phase 1: Concept Extraction & Grounding Setup
Configure MLLM and SAM3 models for open-set concept proposal and spatial grounding on your specific vision models. Initial calibration and data pipeline integration.
Phase 2: Causal Intervention & Data Collection
Execute object-level interventions across representative datasets. Collect and store causal contribution scores and related artifacts. This phase focuses on generating the raw interventional evidence.
Phase 3: Ontology Induction & Reasoning Engine
Aggregate interventional data to construct the conceptual ontology. Implement SPARQL queries and LLM reasoning prompts for global insights and bias detection. Integrate with existing knowledge bases if applicable.
Phase 4: Integration & Continuous Monitoring
Integrate OCCAM's explanations into your existing AI workflows and dashboards. Establish continuous monitoring for model drift and evolving biases, ensuring ongoing interpretability and trust.
Ready to Transform Your AI Interpretability?
Book a free consultation to explore how OCCAM can revolutionize your black-box model understanding and build trust in your AI systems.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat