Combining LoRAs generally fails to compose knowledge across disjoint tasks; reliable gains appear only when the target reasoning pattern or familiar entities are present in fine-tuning data.
Randomly Initialized One-Layer Neural Networks Make Data Linearly Separable
1 Pith paper cite this work. Polarity classification is still indexing.
abstract
Recently, neural networks have demonstrated remarkable capabilities in mapping two arbitrary sets to two linearly separable sets. The prospect of achieving this with randomly initialized neural networks is particularly appealing due to the computational efficiency compared to fully trained networks. This paper contributes by establishing that, given sufficient width, a randomly initialized one-layer neural network can, with high probability, transform two sets into two linearly separable sets without any training. Moreover, we furnish precise bounds on the necessary width of the neural network for this phenomenon to occur. Our initial bound exhibits exponential dependence on the input dimension while maintaining polynomial dependence on all other parameters. In contrast, our second bound is independent of input dimension, effectively surmounting the curse of dimensionality. The main tools used in our proof heavily relies on a fusion of geometric principles and concentration of random matrices.
citation-role summary
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
support 1representative citing papers
citing papers explorer
-
Position: Pause Recycling LoRAs and Prioritize Mechanisms to Uncover Limits and Effectiveness
Combining LoRAs generally fails to compose knowledge across disjoint tasks; reliable gains appear only when the target reasoning pattern or familiar entities are present in fine-tuning data.