I Covered AI’s Eyes to See If It Could Still Recognize Faces

2026-07-19 · Source: Artificial Intelligence on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Intermediate, medium

Summary

An experiment investigated a deep learning face recognition model, InceptionV3, trained on the Labeled Faces in the Wild (LFW) dataset, achieving high metrics (96.4% accuracy, 96.8% precision, 96.4% recall, 96.2% F1-score) to assess its explainability. Using LIME and saliency maps, the study identified the right eye and glabella as consistently critical regions for the model's predictions. Subsequent occlusion tests confirmed that blacking out both eyes and the glabella significantly degraded performance, while single-eye occlusion had minimal impact. A crucial finding was the uneven dependency across individuals; for example, Colin Powell remained recognizable despite occlusion, while Donald Rumsfeld's accuracy varied wildly (0-100%). This highlights person-specific blind spots, indicating that high average accuracy can mask inconsistent failure modes.

Key takeaway

For Machine Learning Engineers deploying facial recognition systems, understand that high average accuracy (e.g., 96%) can conceal critical, person-specific failure modes. You must move beyond explainability tool outputs as final answers; instead, use them to form hypotheses and then physically test those hypotheses through occlusion experiments. This iterative validation helps identify uneven dependencies and blind spots, ensuring your systems perform reliably and equitably across all users, especially under real-world partial occlusion conditions.

Key insights

Explainable AI tools provide hypotheses, not final answers; validation requires intervention.

Principles

• Explainability tools offer initial hypotheses.
• Validate explanations by testing interventions.
• Average accuracy can mask person-specific blind spots.

Method

The method involved training InceptionV3, applying LIME and saliency maps to identify critical regions (right eye, glabella), then physically occluding these regions to validate the explanations.

In practice

• Occlude critical features to validate XAI.
• Continuously audit feature dependency.
• Train models with occluded face examples.

Topics

• Facial Recognition
• Explainable AI
• LIME
• Saliency Maps
• Model Interpretability
• Occlusion Testing

Best for: Computer Vision Engineer, Research Scientist, CTO, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Improving AI models’ ability to explain their predictions — Extracting inherent model concepts yields more accurate and interpretable AI explanations than predefined human concepts.
• LLM-Based Visual Explanation Evaluation Framework for Assessing the Explainability of Facial Skin Disease Classification Models — A framework uses LLMs to evaluate clinically relevant visual explanations for facial skin disease diagnosis.
• A Multi-Dimensional Evaluation of Explainability in Media Bias Detection — Explanation plausibility and mechanistic faithfulness are distinct, complementary XAI dimensions for media bias detection, requiring separate evaluation.
• The Sequence Opinion #810: The Black Box Behind the Mask: The Imperative of Post-Training Interpretability — Post-training interpretability focuses on understanding how AI models modulate knowledge to align with human values and safety.
• 😺 🎙️ Watch: Opening AI’s black box — AI interpretability reveals models' internal "shape-like" reasoning, enabling safer, more reliable, and intentionally designed systems.
• Explainable Reinforcement Learning for assisting Air Traffic Controllers — Applying explainability techniques to Reinforcement Learning in safety-critical domains like Air Traffic Control is crucial for building trust.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence on Medium.