Parameter-Efficient Transformer Embeddings
This addresses the memory efficiency problem for large-scale language model developers, though it is a proof-of-concept study requiring further experimentation.
The paper tackles the problem of embedding layers consuming excessive parameters in transformer NLP models without proportional performance gains by proposing a deterministic Fourier expansion method followed by a lightweight MLP. The result shows competitive performance on natural language inference tasks (SNLI, MNLI) and zero-shot STS-B evaluation with significantly fewer parameters, faster training, and no dropout needed.
Embedding layers in transformer-based NLP models typically account for the largest share of model parameters, scaling with vocabulary size but not yielding performance gains proportional to scale. We propose an alternative approach in which token embedding vectors are first generated deterministically, directly from the token IDs using a Fourier expansion of their normalized values, followed by a lightweight multilayer perceptron (MLP) that captures higher-order interactions. We train standard transformers and our architecture on natural language inference tasks (SNLI and MNLI), and evaluate zero-shot performance on sentence textual similarity (STS-B). Our results demonstrate that the proposed method achieves competitive performance using significantly fewer parameters, trains faster, and operates effectively without the need for dropout. This proof-of-concept study highlights the potential for scalable, memory-efficient language models and motivates further large-scale experimentation based on our findings.