MP-IB
Cross-source consensus on MP-IB from 1 sources and 5 claims.
1 sources · 5 claims
How it works
Benefits
Comparisons
Highlighted claims
- MP-IB uses numerical precision as an information bottleneck to separate stable speaker traits from volatile agitation-related state information. — Mixed-Precision Information Bottlenecks for On-Device Trait-State Disentanglement in Bipolar Agitation Detection
- MP-IB's novelty is heterogeneous arithmetic precision across semantic heads rather than ordinary layer compression for one task. — Mixed-Precision Information Bottlenecks for On-Device Trait-State Disentanglement in Bipolar Agitation Detection
- MP-IB uses an FP16 64-dimensional trait head and an INT4 32-dimensional state head with an 8x capacity asymmetry. — Mixed-Precision Information Bottlenecks for On-Device Trait-State Disentanglement in Bipolar Agitation Detection
- MP-IB achieved the best Bridge2AI agitation prediction among the evaluated methods. — Mixed-Precision Information Bottlenecks for On-Device Trait-State Disentanglement in Bipolar Agitation Detection
- MP-IB's advantage is interpreted as coming from T-MAE pretraining, a low-capacity INT4 state bottleneck, and asymmetric trait-state precision. — Mixed-Precision Information Bottlenecks for On-Device Trait-State Disentanglement in Bipolar Agitation Detection