Homomorphic permutation is fundamental to privacy-preserving computations based on batch-encoding homomorphic encryption. It underpins nearly all homomorphic matrix operations and predominantly influences their complexity. Permutation decomposition as a potential approach to optimize this critical component remains underexplored. In this paper, we propose novel decomposition techniques to optimize homomorphic permutations, advancing homomorphic encryption-based privacy-preserving computations. We start by defining an ideal decomposition form for permutations and propose an algorithm searching for depth-1 ideal decompositions. Based on this, we prove the full-depth ideal decomposability of permutations used in specific homomorphic matrix transposition (HMT) and multiplication (HMM) algorithms, allowing them to achieve asymptotic improvement in speed and rotation key reduction. As a demonstration of applicability, substituting the HMM components in the best-known inference framework of encrypted neural networks with our enhanced version shows up to a 3.9× reduction in latency. We further devise a new method for computing arbitrary homomorphic permutations, specifically those with weak structures that cannot be ideally decomposed. We design a network structure that deviates from the conventional scope of decomposition and outperforms the state-of-the-art technique under a limited rotation key budget, achieving a speed-up of up to 1.69 ×.

Prosodic structure contributes to speech production and comprehension. One of the crucial problems in achieving natural-sounding synthesized speech is the prediction of appropriate phrase boundaries. Unfortunately, obtaining human annotations of prosodic phrases to train a supervised system can be laborious and costly. Active learning has been proven effective in reducing labeling efforts for supervised learning. This study explores active learning techniques with the objective to reduce the amount of human-annotated data needed to attain a given level of performance. It presents an approach based on active learning to predict the Chinese prosodic phrase boundaries in unrestricted Chinese text. Experiments show that for most of the cases considered, the active selection strategies for labeling the prosodic phrase boundaries are as good as or exceed the performance of random data selection.

Facial expressions can be mainly conveyed by only a few discriminative facial regions of interest. In this paper, we study the discriminative regions for facial expression recognition from video sequences. The goal of our method is to explore and make use of the discriminative regions for different facial expressions. For this purpose, we propose a Hidden Markov Model (HMM) Decision Forest (HMMDF). In this framework, each tree node is a discriminative classifier, which is constructed by combining weighted HMMs. Motivated by a psychological theory of "elimination by aspects", several HMMs on each node are modeled respectively for facial regions, which have discriminative capabilities for classification. The weights for these HMMs can be further adjusted according to the contributions of facial regions. Extensive experiments validate the effectiveness of discriminative regions on facial expression, and the experimental results show that the proposed HMMDF framework yields dramatic improvements in facial expression recognition compared to existing methods.

With the increasing demand for spoken language interfaces in human-computer interactions, automatic recognition of emotional states from human speeches has become of increasing importance. Unfortunately, obtaining human annotations of emotion corpus to train a supervised system can become a laborious and costly effort. To address this, we explore active learning techniques with the objective of reducing the amount of human-annotated data needed to attain a given level of performance. In this paper we proposed an approach for speech emotion recognition based on Active Conditional Random Fields. Experiments show that for most of the cases considered, active selection strategies when recognizing speech emotion are as good as or exceed the performance of random data selection.