Indirect Attention: Turning Context Misalignment into a Feature
This addresses a specific bottleneck in attention-based models for scenarios with misaligned context, but it is incremental as it modifies an existing mechanism.
The paper tackled the problem of attention mechanisms failing when keys and values come from different sequences or modalities, by introducing Indirect Attention, which improved performance in synthetic and real-world tasks.
The attention mechanism has become a cornerstone of modern deep learning architectures, where keys and values are typically derived from the same underlying sequence or representation. This work explores a less conventional scenario, when keys and values originate from different sequences or modalities. Specifically, we first analyze the attention mechanism's behavior under noisy value features, establishing a critical noise threshold beyond which signal degradation becomes significant. Furthermore, we model context (key, value) misalignment as an effective form of structured noise within the value features, demonstrating that the noise induced by such misalignment can substantially exceed this critical threshold, thereby compromising standard attention's efficacy. Motivated by this, we introduce Indirect Attention, a modified attention mechanism that infers relevance indirectly in scenarios with misaligned context. We evaluate the performance of Indirect Attention across a range of synthetic tasks and real world applications, showcasing its superior ability to handle misalignment.