RAM-CIL: Reliability-Aware Multimodal Class-Incremental Learning under Modality Reliability Shift
Withdrawn from ICLR 2027, so left out of the paper lists and the map.
Abstract
Experience replay preserves historical training examples, but their observed modalities need not remain reliable learning signals. We study audio-text class-incremental intent recognition under controlled changes to current and replay observations. Holding current inputs clean, replacing historical audio and transcripts with paired shifted versions reduces old-task retention by 15.90 percentage points. We introduce Reliability-Aware Multimodal Class-Incremental Learning (RAM-CIL), which combines a base retrieval score with rank-normalized, label-conditioned predictive support from independent modality classifiers, without changing replay features or the learning objective. The results reveal a protocol-dependent separation between predictive support, retention, and new-task acquisition. Over a complete six-task run, additive reliability correction raises final Macro-F1 while reducing old-task retention. From a shared task-three checkpoint, the same correction instead improves retention. A separate, privileged diagnostic that filters MIR candidates using paired degradation scores improves new-task acquisition by 6.91 points with a 0.21-point retention decrease. These findings establish neither uniform forgetting reduction nor a universal plasticity benefit. They show that predictive reliability is insufficient as a proxy for replay utility, and motivate evaluating stability and plasticity separately under explicitly matched replay protocols.
- Primary area:
- transfer learning, meta learning, and lifelong learning