"SALSA: Soup-based Alignment Learning for Stronger Adaptation in RLHF"

Teaching neural nets to tell time without looking at the clock.

ISF predicts survival times without assumptions by learning implicit time representations.

This Time-aware neural networks is really innovative for survival prediction accuracy.

https://arxiv.org/abs/2411.01798

Original Problem 🎯:

Traditional survival analysis methods rely on strong assumptions about hazard rates or use discrete time spaces, limiting their ability to model complex real-world survival distributions accurately. Current deep learning approaches either assume time-invariant hazards or are restricted to preset time points.

-----

Solution in this Paper 🔧:

→ ISF (Implicit Survival Function) directly models conditional hazard rates without assumptions using neural networks

→ Uses positional encoding to capture time patterns through sinusoidal functions

→ Employs numerical integration to approximate cumulative distribution functions

→ Introduces a unified loss function handling both censored and uncensored samples

→ Implements parallel computation for faster processing across time points

-----

Key Insights 💡:

→ Weight space averaging enables better exploration during training

→ Sinusoidal positional encoding allows learning high-frequency functions

→ Model architecture is independent of time space discretization

→ Unified loss function simplifies handling different censoring types

-----

Results 📊:

→ Outperforms state-of-the-art methods on three public datasets (CLINIC, MUSIC, METABRIC)

→ Achieves highest Concordance Index (CI) scores: CLINIC (0.612), MUSIC (0.701), METABRIC (0.704)

→ Shows robustness across different censoring rates (13.2% to 55.2%)