There is no theory yet as to whydeep neural networks (DNNs from here on) are as effectiveand efficient in practice as they evidently are.

State-of-the-art word embeddings presume a linear vector space, but this approach does not easily incorporate the nonlinearity that is necessary to represent polysemy. We thus propose a novel semantic FIeld REepresentation, called FIRE, which is a $D$-dimensional field in which every word is represented as a set of its locations and a nonlinear function covering the field. The strength of a word's relation to another word at a certain location is measured as the function value at that location. With FIRE, compositionality is represented via functional additivity, whereas polysemy is represented via the set of points and the function's multimodality. By implementing FIRE for English and comparing it with previous representation methods via word and sentence similarity tasks, we show that FIRE produces comparable or even better results. In an evaluation of polysemy to predict the number of word senses, FIRE greatly outperformed BERT and Word2vec, providing evidence of how FIRE represents polysemy. The code is available at https://github.com/kduxin/firelang.

We initiate the study of private inference on Transformer-based models in the client-server setting, where clients have private inputs and servers hold proprietary models. Our main contribution is to provide several new secure protocols for matrix multiplication and complex non-linear functions like Softmax, GELU activations, and LayerNorm, which are critical components of Transformers. Specifically, we first propose a customized homomorphic encryption-based protocol for matrix multiplication that crucially relies on a novel compact packing technique. This design achieves $\sqrt{m} \times$ less communication ($m$ is the number of rows of the output matrix) over the most efficient work. Second, we design efficient protocols for three non-linear functions via integrating advanced underlying protocols and specialized optimizations. Compared to the state-of-the-art protocols, our recipes reduce about half of the communication and computation overhead. Furthermore, all protocols are numerically precise, which preserve the model accuracy of plaintext. These techniques together allow us to implement \Name, an efficient Transformer-based private inference framework. Experiments conducted on several real-world datasets and models demonstrate that \Name achieves $3 \sim 14\times$ less communication and $3 \sim 11\times$ less runtime compared to the prior art.

Transformer models have gained significant attention due to their power in machine learning tasks.

State-of-the-art word embeddings presume a linear vector space, but this approach does not easily incorporate the nonlinearity that is necessary to represent polysemy. We thus propose a novel semantic FIeld REepresentation, called FIRE, which is a D -dimensional field in which every word is represented as a set of its locations and a nonlinear function covering the field. The strength of a word's relation to another word at a certain location is measured as the function value at that location. With FIRE, compositionality is represented via functional additivity, whereas polysemy is represented via the set of points and the function's multimodality. By implementing FIRE for English and comparing it with previous representation methods via word and sentence similarity tasks, we show that FIRE produces comparable or even better results.

We initiate the study of private inference on Transformer-based models in the client-server setting, where clients have private inputs and servers hold proprietary models. Our main contribution is to provide several new secure protocols for matrix multiplication and complex non-linear functions like Softmax, GELU activations, and LayerNorm, which are critical components of Transformers. Specifically, we first propose a customized homomorphic encryption-based protocol for matrix multiplication that crucially relies on a novel compact packing technique. This design achieves \sqrt{m} \times less communication ( m is the number of rows of the output matrix) over the most efficient work. Second, we design efficient protocols for three non-linear functions via integrating advanced underlying protocols and specialized optimizations.

Vision Transformer (ViT) acceleration with field programmable gate array (FPGA) is promising but challenging. Existing FPGA-based ViT accelerators mainly rely on temporal architectures, which process different operators by reusing the same hardware blocks and suffer from extensive memory access overhead. Pipelined architectures, either coarse-grained or fine-grained, unroll the ViT computation spatially for memory access efficiency. However, they usually suffer from significant hardware resource constraints and pipeline bubbles induced by the global computation dependency of ViT. In this paper, we introduce HG-PIPE, a pipelined FPGA accelerator for high-throughput and low-latency ViT processing. HG-PIPE features a hybrid-grained pipeline architecture to reduce on-chip buffer cost and couples the computation dataflow and parallelism design to eliminate the pipeline bubbles. HG-PIPE further introduces careful approximations to implement both linear and non-linear operators with abundant Lookup Tables (LUTs), thus alleviating resource constraints. On a ZCU102 FPGA, HG-PIPE achieves 2.78 times better throughput and 2.52 times better resource efficiency than the prior-art accelerators, e.g., AutoViTAcc. With a VCK190 FPGA, HG-PIPE realizes end-to-end ViT acceleration on a single device and achieves 7118 images/s, which is 2.81 times faster than a V100 GPU.