Towards Improved Imbalance Robustness in Continual Multi-Label Learning with Dual Output Spiking Architecture (DOSA)

TL;DR

This work tackles task-agnostic continual multi-label learning by introducing DOSA, a dual-output spiking neural network, and an imbalance-aware loss with per-class margins to address label scarcity and changing label spaces. The proposed loss, $\ ext{L}_{fmm}$, combines a maximum-margin objective with a per-class trainable margin $\\mathbf{b}$ and an importance weighting $\\exp(-(\\mathbf{\\zeta_k}-\\mathbf{b}))$, formalized as $\\mathcal{L}_{fmm}=\\sum_{k=1}^N\\exp(-(\\mathbf{\\zeta_k}-\\mathbf{b}))\\odot\\|\\mathbf{\\zeta_k}-\\mathbf{b}\\|^2$, where $\\zeta_k=\\mathbf{y_k}\\odot(\\mathbf{y_{k+}}-\\mathbf{y_{k-}})$. DOSA employs a SEA scheme to continually expand the model across tasks while maintaining prior knowledge. Experimental results on multiple MLL and CMLL benchmarks show that $\\mathcal{L}_{fmm}$ yields robust improvements over the baseline $\\mathcal{L}_{mm}$ and outperforms CIFDM in several datasets, highlighting the value of task-agnostic, imbalance-aware learning in neuromorphic architectures. The work also demonstrates that a careful ablation of network depth and gradient flow through the importance factor is important for maximizing performance in CMLL with SNNs.

Abstract

Algorithms designed for addressing typical supervised classification problems can only learn from a fixed set of samples and labels, making them unsuitable for the real world, where data arrives as a stream of samples often associated with multiple labels over time. This motivates the study of task-agnostic continual multi-label learning problems. While algorithms using deep learning approaches for continual multi-label learning have been proposed in the recent literature, they tend to be computationally heavy. Although spiking neural networks (SNNs) offer a computationally efficient alternative to artificial neural networks, existing literature has not used SNNs for continual multi-label learning. Also, accurately determining multiple labels with SNNs is still an open research problem. This work proposes a dual output spiking architecture (DOSA) to bridge these research gaps. A novel imbalance-aware loss function is also proposed, improving the multi-label classification performance of the model by making it more robust to data imbalance. A modified F1 score is presented to evaluate the effectiveness of the proposed loss function in handling imbalance. Experiments on several benchmark multi-label datasets show that DOSA trained with the proposed loss function shows improved robustness to data imbalance and obtains better continual multi-label learning performance than CIFDM, a previous state-of-the-art algorithm.

Figures (4)

• Figure 1: DOSA and inference process. The Sequential Learning with Model Adaptation Setup for Continual Learning with DOSA is shown in the bottom half.
• Figure 2: Ablation Study on the Birds Dataset: Effect of number of hidden layers on overall multi-label classification performance.
• Figure 3: Proportion of Samples and Normalized margin values. PLIF denotes parametric leaky integrate and fire neuron (used at the hidden layers) PLIF.
• Figure 4: CMLL Performance Comparison: DOSA with maximum margin loss ($\mathcal{L}_{mm}$), with focal maximum margin loss ($\mathcal{L}_{fmm}$), and CIFDM.