TL;DR
This paper introduces a novel "Growing Dual-Memory" (GDM) architecture for lifelong learning in robots and agents. GDM uses two interconnected, self-organizing recurrent networks to learn spatiotemporal information from sequential data, effectively addressing catastrophic forgetting and outperforming existing methods on a challenging object recognition benchmark.
What Problem Does This Paper Address?
Artificial agents operating in dynamic environments need to continuously learn and adapt without forgetting previously acquired knowledge. This is known as lifelong learning, and a major hurdle is "catastrophic forgetting," where learning new information drastically degrades performance on older information. Existing models often simplify this problem by using static image datasets and controlled learning scenarios. However, real-world agents encounter sequential data (like video) where information arrives incrementally, and past experiences might not be readily accessible. This paper tackles the challenge of building robust lifelong learning systems that can handle continuous streams of spatiotemporal data, mimicking more natural learning conditions.
Methods
The core of the proposed solution is a Growing Dual-Memory (GDM) architecture. It consists of two main components:
- Growing Episodic Memory (G-EM): This is an unsupervised, recurrent self-organizing network (an extension of the Gamma-GWR model) that learns fine-grained spatiotemporal representations of individual object instances from sequential input (like video frames). It dynamically grows by adding neurons and connections as needed to capture novel sensory experiences.
- Growing Semantic Memory (G-SM): This network also grows and self-organizes but uses task-relevant signals (like object labels) to regulate its learning. It receives information from G-EM and learns more compact, generalized representations of object categories. This top-down regulation helps prevent interference and consolidate knowledge.
A key innovation is memory replay. When external sensory input is absent, G-EM periodically replays learned neural activity trajectories. These "pseudo-patterns" are fed back into both G-EM and G-SM to reinforce existing knowledge and mitigate forgetting, especially when dealing with imbalanced datasets or infrequent encounters.
Datasets
The primary dataset used for evaluation is CORe50. This dataset features video sequences of 50 objects across 10 categories, captured under various conditions (backgrounds, poses, occlusions). It's designed for continuous object recognition and incremental learning scenarios, making it suitable for testing lifelong learning capabilities.
Key Insights and Ideas
- Dual-Memory for Stability and Plasticity: The separation into episodic (instance-level, fine-grained) and semantic (category-level, generalized) memories provides a biological inspiration for balancing the need to learn new things (plasticity) with retaining old knowledge (stability).
- Recurrent Networks for Spatiotemporal Data: Directly processing sequential data with recurrent networks, rather than treating frames independently, is crucial for capturing temporal dynamics and improving recognition accuracy in video.
- Unsupervised Learning with Regulated Plasticity: G-EM learns unsupervisedly, adapting to new inputs, while G-SM uses labels to guide its learning, preventing disruptive interference and promoting generalization.
- Memory Replay as a Consolidation Mechanism: The periodic replay of learned neural trajectories acts as a form of "internal practice," reinforcing memories and significantly reducing catastrophic forgetting, especially in incremental learning settings.
- Dynamic Network Growth: The ability of both G-EM and G-SM to grow dynamically allows them to adapt to the complexity of incoming data without a fixed capacity, a common limitation in traditional neural networks.
Key Results
The GDM model demonstrated significant improvements over state-of-the-art lifelong learning approaches on the CORe50 benchmark across three incremental learning scenarios:
- Batch Learning: GDM achieved an accuracy of 79.43% (instance) and 93.92% (category), outperforming a fine-tuned VGG network by over 10% and 13% respectively. This highlights the benefit of temporal context.
- Incremental Learning (without replay): Achieved 75.93% (instance) and 85.53% (category) accuracy.
- Incremental Learning (with replay): Significantly boosted performance to 82.14% (instance) and 91.18% (category), demonstrating the effectiveness of memory replay in mitigating forgetting.
- CORe50 Scenarios (NI, NC, NIC): GDM with replay achieved state-of-the-art results, with accuracies of 87.94%, 86.14%, and 87.06% respectively, showing substantial improvements (over 20%) compared to previous cumulative approaches.
Limitations and Future Work
- Unseen Classes: The current model predicts a label for any input sequence, even if it's an unseen class. Future work could involve designing a novelty detection mechanism to identify and handle truly unknown classes.
- Multisensory Integration: The authors suggest extending the model to multisensory domains (e.g., audio-visual streams) for more comprehensive learning.
- Connection/Neuron Deletion: While the model grows, it doesn't explicitly discuss mechanisms for deleting old connections or isolated neurons, which could be a future area of exploration for memory management.
- Comparison to ART: The paper notes conceptual similarities to Adaptive Resonance Theory (ART) but points out that ART's performance can be order-dependent, and extensive benchmark evaluations are lacking. Further comparison and integration could be beneficial.
Summary updated