Model Metamers: Exploring the Differences Between Computational Models and Human Perception

In the world of neuroscience and artificial intelligence (AI), researchers face a common challenge: the differences between computational models and human perception. Recent studies have shown that artificial neural networks, designed to imitate the functions of human sensory systems, often exhibit different invariances compared to those found in humans. This raises questions about the principles behind these models and their practicality in real-world situations.

Addressing the Discrepancies

Historically, efforts to tackle the issue of invariance discrepancies have involved studying areas such as model vulnerability to adversarial perturbations and the impact of noise and translations on model judgments.

Introducing Model Metamers

Model metamers draw inspiration from human perceptual metamers, which are stimuli that produce indistinguishable responses in certain stages of the sensory system, despite being physically distinct. In computational models, model metamers are synthetic stimuli that generate nearly identical activations as specific natural images or sounds in a model. The key question is whether humans can recognize these model metamers as belonging to the same category as the corresponding biological signals.

A Study’s Findings

The research team behind a recent study created model metamers from various deep neural network models, including supervised and unsupervised learning models, for vision and audition. Surprisingly, the late-stage model metamers were consistently unrecognizable to human observers. This indicates that many invariances in these models do not align with the human sensory system.

Effective and Predictable

Model metamers prove their efficacy in highlighting the differences between models and humans. Interestingly, the human recognizability of model metamers strongly correlated with their recognition by other models, suggesting that the gap between humans and models lies in the idiosyncratic invariances unique to each model.

Conclusion

The concept of model metamers is a significant step toward understanding and addressing the disparities between computational models of sensory systems and human sensory perception. These synthetic stimuli offer a fresh perspective for researchers striving to create more biologically faithful models. Though more work is needed, model metamers provide a promising benchmark for future model evaluation and the development of artificial systems that better align with the complexities of human sensory perception.

If you are interested in reading the full research paper, you can find it here.

Niharika Singh is a Technical consulting intern at Marktechpost. She is currently pursuing her undergraduate degree in B.Tech from the Indian Institute of Technology (IIT), Kharagpur. Niharika has a strong interest in machine learning, data science, and AI, and keeps up-to-date with the latest developments in these fields.

Source link