Learning Then, Learning Now, and Every Second in Between: Lifelong Learning With a Simulated Humanoid Robot

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Long-term human-robot interaction necessitates continuous knowledge acquisition, known as lifelong learning (LL).
• Catastrophic forgetting is a major challenge in LL, where new learning degrades previously acquired knowledge.
• Existing LL approaches like Growing Dual-Memory (GDM) and Self-organizing Incremental Neural Network+ (SOINN+) utilize growing neural networks but SOINN+ lacks classification investigation.

Purpose of the Study:

• To introduce an extended SOINN+ model, associative SOINN+ (A-SOINN+), for enhanced classification capabilities in lifelong learning.
• To present a novel lifelong learning object recognition dataset, v-NICO-World-LL, recorded in a virtual environment.
• To evaluate the performance of A-SOINN+ against existing methods on object recognition tasks.

Main Methods:

• An associative SOINN+ (A-SOINN+) model was developed, incorporating properties from the GDM model to enable classification.
• A new dataset, v-NICO-World-LL, was created featuring a virtual humanoid robot interacting with 100 objects across 10 classes with varied background complexities.
• The A-SOINN+ model was evaluated on both v-NICO-World-LL and the CORe50 dataset.

Main Results:

• A-SOINN+ achieved classification accuracy comparable to state-of-the-art GDM architectures.
• The A-SOINN+ model utilized 30 to 350 times fewer neurons compared to GDM.
• Training time for A-SOINN+ was approximately 268 times lower than GDM, indicating significant efficiency gains.

Conclusions:

• The A-SOINN+ model offers a highly efficient solution for lifelong learning object recognition.
• Reduced neuron count and training time make A-SOINN+ suitable for robots with limited computational resources.
• This research facilitates more efficient lifelong learning in autonomous social robots for long-term human-robot interactions.