Expanding continual few-shot learning benchmarks to include recognition of specific instances

Area of Science:

• Machine Learning
• Artificial Intelligence
• Cognitive Science

Background:

• Continual learning and few-shot learning are key ML frontiers.
• Combining these approaches is crucial for advanced AI.
• Existing frameworks like CFSL offer a starting point.

Purpose of the Study:

• Extend the Continual Few-Shot Learning (CFSL) framework.
• Introduce more realistic challenges for intelligent agents.
• Evaluate ML model performance in complex learning scenarios.

Main Methods:

• Increased the number of classes by tenfold compared to prior CFSL.
• Introduced an 'instance test' for recognizing specific class examples.
• Evaluated baseline models and a variant with replay.

Main Results:

• Learning more classes proved more challenging.
• Presentation of instances and classes impacted classification.
• Instance test accuracy was comparable to standard tasks but degraded with occlusion/noise.
• Replay substantially improved performance on both tasks.

Conclusions:

• Expanded CFSL addresses more real-world AI challenges.
• Instance recognition is a critical, often overlooked, capability.
• Replay mechanisms are vital for robust continual few-shot learning.