Embodied Object Representation Learning and Recognition.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Robotics

Background:

• Current AI struggles with object recognition beyond trained datasets, unlike humans who learn through interaction.
• Neuroscience suggests cortical columns build predictive models for object recognition within their reference frame.

Purpose of the Study:

• To develop an enactive embodied agent capable of learning and recognizing novel objects through active interaction.
• To implement a generative model inspired by cortical columns for object understanding.

Main Methods:

• Utilized Cortical Column Networks (CCNs), a deep neural network for each object category, learning a generative model.
• Employed active inference for model parameter optimization via variational free energy minimization.
• Developed an ensemble of CCNs for object classification and instantiation of new CCNs for unknown objects.

Main Results:

• Demonstrated improved classification accuracy with increased evidence gathering in a simulated environment.
• Showcased the agent's ability to learn and recognize previously unseen objects.
• Validated that active inference can guide agent actions towards preferred observations.

Conclusions:

• The proposed enactive embodied agent effectively learns and recognizes objects, including novel ones, through active interaction and generative modeling.
• The system's performance improves with experience, highlighting the benefits of active learning in AI.
• Active inference provides a viable framework for intelligent agents to explore and understand their environment.