As more and more pre-trained language models adopt on-cloud deployment, the privacy issues grow quickly, mainly for the exposure of plain-text user data (e.g., search history, medical record, bank account). Privacy-preserving inference of transformer models is on the demand of cloud service users. To protect privacy, it is an attractive choice to compute only with ciphertext in homomorphic encryption (HE). However, enabling pre-trained models inference on ciphertext data is difficult due to the complex computations in transformer blocks, which are not supported by current HE tools yet. In this work, we introduce $\textit{THE-X}$, an approximation approach for transformers, which enables privacy-preserving inference of pre-trained models developed by popular frameworks. $\textit{THE-X}$ proposes a workflow to deal with complex computation in transformer networks, including all the non-polynomial functions like GELU, softmax, and LayerNorm. Experiments reveal our proposed $\textit{THE-X}$ can enable transformer inference on encrypted data for different downstream tasks, all with negligible performance drop but enjoying the theory-guaranteed privacy-preserving advantage.
The sparse Mixture-of-Experts (MoE) model is powerful for large-scale pre-training and has achieved promising results due to its model capacity. However, with trillions of parameters, MoE is hard to be deployed on cloud or mobile environment. The inference of MoE requires expert parallelism, which is not hardware-friendly and communication expensive. Especially for resource-limited downstream tasks, such sparse structure has to sacrifice a lot of computing efficiency for limited performance gains. In this work, we observe most experts contribute scarcely little to the MoE fine-tuning and inference. We further propose a general method to progressively drop the non-professional experts for the target downstream task, which preserves the benefits of MoE while reducing the MoE model into one single-expert dense model. Our experiments reveal that the fine-tuned single-expert model could preserve 99.3% benefits from MoE across six different types of tasks while enjoying 2x inference speed with free communication cost.
Sparse mixture of experts provides larger model capacity while requiring a constant computational overhead. It employs the routing mechanism to distribute input tokens to the best-matched experts according to their hidden representations. However, learning such a routing mechanism encourages token clustering around expert centroids, implying a trend toward representation collapse. In this work, we propose to estimate the routing scores between tokens and experts on a low-dimensional hypersphere. We conduct extensive experiments on cross-lingual language model pre-training and fine-tuning on downstream tasks. Experimental results across seven multilingual benchmarks show that our method achieves consistent gains. We also present a comprehensive analysis on the representation and routing behaviors of our models. Our method alleviates the representation collapse issue and achieves more consistent routing than the baseline mixture-of-experts methods.
In this paper, we propose a simple yet effective method to stabilize extremely deep Transformers. Specifically, we introduce a new normalization function (DeepNorm) to modify the residual connection in Transformer, accompanying with theoretically derived initialization. In-depth theoretical analysis shows that model updates can be bounded in a stable way. The proposed method combines the best of two worlds, i.e., good performance of Post-LN and stable training of Pre-LN, making DeepNorm a preferred alternative. We successfully scale Transformers up to 1,000 layers (i.e., 2,500 attention and feed-forward network sublayers) without difficulty, which is one order of magnitude deeper than previous deep Transformers. Remarkably, on a multilingual benchmark with 7,482 translation directions, our 200-layer model with 3.2B parameters significantly outperforms the 48-layer state-of-the-art model with 12B parameters by 5 BLEU points, which indicates a promising scaling direction.
Knowledge-Enhanced Model have developed a diverse set of techniques for knowledge integration on different knowledge sources. However, most previous work neglect the language model's own ability and simply concatenate external knowledge at the input. Recent work proposed that Feed Forward Network (FFN) in pre-trained language model can be seen as an memory that stored factual knowledge. In this work, we explore the FFN in Transformer and propose a novel knowledge fusion model, namely Kformer, which incorporates external knowledge through the feed-forward layer in Transformer. We empirically find that simply injecting knowledge into FFN can enhance the pre-trained language model's ability and facilitate current knowledge fusion methods. Our results on two benchmarks in the commonsense reasoning (i.e., SocialIQA) and medical question answering (i.e., MedQA-USMLE) domains demonstrate that Kformer can utilize external knowledge deeply and achieves absolute improvements in these tasks.
The poor performance of the original BERT for sentence semantic similarity has been widely discussed in previous works. We find that unsatisfactory performance is mainly due to the static token embeddings biases and the ineffective BERT layers, rather than the high cosine similarity of the sentence embeddings. To this end, we propose a prompt based sentence embeddings method which can reduce token embeddings biases and make the original BERT layers more effective. By reformulating the sentence embeddings task as the fillin-the-blanks problem, our method significantly improves the performance of original BERT. We discuss two prompt representing methods and three prompt searching methods for prompt based sentence embeddings. Moreover, we propose a novel unsupervised training objective by the technology of template denoising, which substantially shortens the performance gap between the supervised and unsupervised setting. For experiments, we evaluate our method on both non fine-tuned and fine-tuned settings. Even a non fine-tuned method can outperform the fine-tuned methods like unsupervised ConSERT on STS tasks. Our fine-tuned method outperforms the state-of-the-art method SimCSE in both unsupervised and supervised settings. Compared to SimCSE, we achieve 2.29 and 2.58 points improvements on BERT and RoBERTa respectively under the unsupervised setting.
This report describes Microsoft's machine translation systems for the WMT21 shared task on large-scale multilingual machine translation. We participated in all three evaluation tracks including Large Track and two Small Tracks where the former one is unconstrained and the latter two are fully constrained. Our model submissions to the shared task were initialized with DeltaLM\footnote{\url{https://aka.ms/deltalm}}, a generic pre-trained multilingual encoder-decoder model, and fine-tuned correspondingly with the vast collected parallel data and allowed data sources according to track settings, together with applying progressive learning and iterative back-translation approaches to further improve the performance. Our final submissions ranked first on three tracks in terms of the automatic evaluation metric.
While pre-trained language models have achieved great success on various natural language understanding tasks, how to effectively leverage them into non-autoregressive generation tasks remains a challenge. To solve this problem, we present a non-autoregressive generation model based on pre-trained transformer models. To bridge the gap between autoregressive and non-autoregressive models, we propose a simple and effective iterative training method called MIx Source and pseudo Target (MIST). Unlike other iterative decoding methods, which sacrifice the inference speed to achieve better performance based on multiple decoding iterations, MIST works in the training stage and has no effect on inference time. Our experiments on three generation benchmarks including question generation, summarization and paraphrase generation, show that the proposed framework achieves the new state-of-the-art results for fully non-autoregressive models. We also demonstrate that our method can be used to a variety of pre-trained models. For instance, MIST based on the small pre-trained model also obtains comparable performance with seq2seq models.
Compared to monolingual models, cross-lingual models usually require a more expressive vocabulary to represent all languages adequately. We find that many languages are under-represented in recent cross-lingual language models due to the limited vocabulary capacity. To this end, we propose an algorithm VoCap to determine the desired vocabulary capacity of each language. However, increasing the vocabulary size significantly slows down the pre-training speed. In order to address the issues, we propose k-NN-based target sampling to accelerate the expensive softmax. Our experiments show that the multilingual vocabulary learned with VoCap benefits cross-lingual language model pre-training. Moreover, k-NN-based target sampling mitigates the side-effects of increasing the vocabulary size while achieving comparable performance and faster pre-training speed. The code and the pretrained multilingual vocabularies are available at https://github.com/bozheng-hit/VoCapXLM.