Get started
Work

Language Embedding-Based Hardware-Aware Neural Architecture Representation Learning

Year: 2026
Type: article
Abstract: Neural architecture representation learning maps network architectures into feature vectors for predicting properties, such as latency and accuracy, and is, thus, central to practical deployment and design of models. Recently, transformer-based models integrated with graph neural networks have achieved significant progress in representation learning. However, current methods still exhibit two key limitations. First, they generally ignore hardware attributes, which is increasingly unrealistic given the growing diversity of deployment platforms and restricts their real-world applicability. Second, their encoders typically rely on static adjacency matrices to describe architectural topology, making them insensitive to structural differences among computational nodes and ultimately weakening representation quality. In this article, we present Le-Former, a framework that alleviates these limitations by coupling language-based semantic embeddings with dynamic graph representation learning. We first design a language embedding module that converts neural architectures and hardware platform specifications into compact, consistently ordered textual schemas and embeds them with a replaceable pretrained text encoder. Then, we introduce a dynamic graph-based transformer that models neural architectures with adaptive relational structure, yielding stronger architecture encodings than static-graph counterparts. Experiments on NNLQP show consistent latency-prediction gains and provide an initial demonstration of held-out hardware-precision transfer for latency prediction between NVIDIA P4 and T4 GPUs under the evaluated precision settings. Besides, our method achieves competitive performance on the cell-based neural architecture search (NAS)-Bench-101 and NAS-Bench-201 benchmarks. The source code is publicly available athttps://github.com/respACz/Le-Former.
DOI: 10.1109/tii.2026.3722260
Language: English
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7207584863
FWCI:
Cites:
Cited by:
Related to: 10