Yuqi Feng

Peijie Jiang, Yuqi Feng, Cunyin Peng et al.

In contemporary large language models (LLMs), the swish-gated linear unit (SwiGLU) activation function is widely adopted to regulate the information flow and introduce non-linearity. For large positive inputs, SwiGLU approximates the quadratic function $x^2$, providing strong nonlinearity and expressive capacity. However, this property also causes numerical instability as the input or model scale increases, particularly in low-precision LLM training. The main reason is its approximate quadratic amplification, which enlarges the output range and exacerbates outliers. To address this issue, we propose a stable activation function, Power Linear Unit (PowLU), for large-scale LLM pre-training. Specifically, PowLU employs a rational power function to achieve adaptive nonlinearity, thereby improving representation ability and enabling stable training in spike regions. Moreover, we provide theoretical justification for several key properties of PowLU. Scaling law experiments confirm that the performance is consistent across model sizes, and further experimental results with the Ling architecture (7.9B and 124B total parameters) demonstrate that PowLU achieves competitive results against SwiGLU and SwiGLU-Clip in large-scale training of LLMs. In addition, the experimental results also show that PowLU effectively improves the scalability of the large-scale training of LLMs.

Han Ji, Yuqi Feng, Jiahao Fan et al.

Performance predictors have emerged as a promising method to accelerate the evaluation stage of neural architecture search (NAS). These predictors estimate the performance of unseen architectures by learning from the correlation between a small set of trained architectures and their performance. However, most existing predictors ignore the inherent distribution shift between limited training samples and diverse test samples. Hence, they tend to learn spurious correlations as shortcuts to predictions, leading to poor generalization. To address this, we propose a Causality-guided Architecture Representation Learning (CARL) method aiming to separate critical (causal) and redundant (non-causal) features of architectures for generalizable architecture performance prediction. Specifically, we employ a substructure extractor to split the input architecture into critical and redundant substructures in the latent space. Then, we generate multiple interventional samples by pairing critical representations with diverse redundant representations to prioritize critical features. Extensive experiments on five NAS search spaces demonstrate the state-of-the-art accuracy and superior interpretability of CARL. For instance, CARL achieves 97.67% top-1 accuracy on CIFAR-10 using DARTS.

Han Ji, Yuqi Feng, Yanan Sun

Neural predictors are effective in boosting the time-consuming performance evaluation stage in neural architecture search (NAS), owing to their direct estimation of unseen architectures. Despite the effectiveness, training a powerful neural predictor with fewer annotated architectures remains a huge challenge. In this paper, we propose a context-aware neural predictor (CAP) which only needs a few annotated architectures for training based on the contextual information from the architectures. Specifically, the input architectures are encoded into graphs and the predictor infers the contextual structure around the nodes inside each graph. Then, enhanced by the proposed context-aware self-supervised task, the pre-trained predictor can obtain expressive and generalizable representations of architectures. Therefore, only a few annotated architectures are sufficient for training. Experimental results in different search spaces demonstrate the superior performance of CAP compared with state-of-the-art neural predictors. In particular, CAP can rank architectures precisely at the budget of only 172 annotated architectures in NAS-Bench-101. Moreover, CAP can help find promising architectures in both NAS-Bench-101 and DARTS search spaces on the CIFAR-10 dataset, serving as a useful navigator for NAS to explore the search space efficiently.

Yuwei Ou, Yuqi Feng, Yanan Sun

To defend deep neural networks from adversarial attacks, adversarial training has been drawing increasing attention for its effectiveness. However, the accuracy and robustness resulting from the adversarial training are limited by the architecture, because adversarial training improves accuracy and robustness by adjusting the weight connection affiliated to the architecture. In this work, we propose ARNAS to search for accurate and robust architectures for adversarial training. First we design an accurate and robust search space, in which the placement of the cells and the proportional relationship of the filter numbers are carefully determined. With the design, the architectures can obtain both accuracy and robustness by deploying accurate and robust structures to their sensitive positions, respectively. Then we propose a differentiable multi-objective search strategy, performing gradient descent towards directions that are beneficial for both natural loss and adversarial loss, thus the accuracy and robustness can be guaranteed at the same time. We conduct comprehensive experiments in terms of white-box attacks, black-box attacks, and transferability. Experimental results show that the searched architecture has the strongest robustness with the competitive accuracy, and breaks the traditional idea that NAS-based architectures cannot transfer well to complex tasks in robustness scenarios. By analyzing outstanding architectures searched, we also conclude that accurate and robust neural architectures tend to deploy different structures near the input and output, which has great practical significance on both hand-crafting and automatically designing of accurate and robust architectures.