Large-scale pre-trained language models such as BERT have contributed significantly to the development of NLP. However, those models require large computational resources, making it difficult to be applied to mobile devices where computing power is limited. In this paper we aim to address the weakness of existing input-adaptive inference methods which fail to take full advantage of the structure of BERT. We propose Dynamic Planning in BERT, a novel fine-tuning strategy that can accelerate the inference process of BERT through selecting a subsequence of transformer layers list of backbone as a computational path for an input sample. To do this, our approach adds a planning module to the original BERT model to determine whether a layer is included or bypassed during inference. Experimental results on the GLUE benchmark exhibit that our method reduces latency to 75% while maintaining 98% accuracy, yielding a better accuracy-speed trade-off compared to state-of-the-art input-adaptive methods.

This paper proposes a general framework called Markov stationary features (MSF) to extend histogram based features. The MSF characterizes the spatial co-occurrence of histogram patterns by Markov chain models, and finally yields a compact feature representation through Markov stationary analysis. Therefore, the MSF goes one step beyond histograms since it now involves spatial structure information of both within histogram bins and between histogram bins. Moreover, it still keeps simplicity, compactness, efficiency, and robustness. We demonstrate how the MSF is used to extend histogram based features like color histogram, edge histogram, local binary pattern histogram and histogram of oriented gradients. We evaluate the MSF extended histogram features on the task of TRECVID video concept detection. Results show that the proposed MSF extensions can achieve significant performance improvement over corresponding histogram features.

We introduce a new semiformal design validation framework to justify hard-to-reach corner-case states. We propose a cultural learning technique to identify the swarming of domain knowledge during the search. In addition, our guidance strategy abstracts sets of partitioned state variables, from which pre-images are computed to capture the expanded portions of the state spaces related to a target state. Experimental results show that our approach is very effective to reach hard- to-reach states than existing methods.

Local binary pattern histogram (LBPH) is one of the popular and excellent image texture descriptor. However, conventional LBPH lacks of the description of spatial structure information. This paper proposes an extension of LBPH called Markov chain local binary patterns (MCLBP) to alleviate this limitation. We apply MCLBP to the task of TRECVID video concept detection. Experimental results demonstrate that MCLBP achieves significant performance improvement over conventional LBPH features and other widely used visual features such as Gabor texture.

We present a new approach to intelligently mine three types of invariants from a sequential circuit to significantly improve SAT-based state justification. We adaptively generate mining databases targeting on the hard-to-reach corner-case states, from which global invariants, target state related invariants, and observability-don't-care extended invariants are mined. Each mined invariant involves two or more signals that span across multiple time-frames, which capture the knowledge of the state spaces related to a target state. These invariants are then checked for their validity, and they can significantly increase the deductive power of the instance by pruning a larger portion of the search space. Experimental results show that more than an order of magnitude performance improvement can be obtained when justifying hard-to-justify states.

We present a novel technique on mining relationships in a sequential circuit to discover global constraints. We utilize domain knowledge to prune the search space for our mining process, via which the mined relations may often span several time frames, thus improving the deductibility of the problem instance. Experimental results demonstrate that the application of these global constraints to satisfiability (SAT)-based bounded sequential equivalence checking can achieve a one to two orders-of-magnitude speedup.

Using DFT methods such as scan can improve testability and increase fault coverage. However, scan tests may scan in illegal or unreachable states during test application, which may result in incidental detection of functional untestable delay faults during the scan test. This paper presents novel mining techniques for fast top-down functional constraint extraction. The extracted functional constraints capture illegal states through internal signal relations. Imposing these relations as functional constraints to a commercial ATPG tool allows for the generation of effective pseudo-functional tests. We analyze its impact on minimizing the over-testing problem of the scan- based circuits. The experimental results on transition faults and path delay faults reveal that the proposed method produces a small fraction, yet extremely powerful functional constraints effective for constraining the state space.

This work introduces a new guidance strategy to justify hard-to-reach target states in sequential circuits, which also applies to reaching corner-case states in design validation. We propose data-mining methods to extract several partition sets of state variables directly from the gate-level netlist of the design, such that they enable the search to reach the target state. The partition sets are used to compute Partitioned Navigation Tracks (PNTs). PNTs capture the behavior of expanded portions of the state space as they relate to a target state of interest, thus providing a more accurate distance metric than traditional abstract guideposts. Moreover, the computation and storage costs of the PNTs are small, making our approach scalable to large circuits. With the proposed PNTs, we also need not refine the abstract models. Experiments showed that we are able to reach many more hard-to-reach states compared to state-of-the-art methods.

In this paper, we propose a novel technique on mining relationships in a sequential circuit to discover global constraints. In contrast to the traditional learning methods, our mining algorithm can find important relationships among several nodes efficiently. The nodes involved may often span several time-frames, thus improving the deductibility of the problem instance. Experimental results demonstrate that the application of these global constraints to SAT-based bounded sequential equivalence checking can achieve one to two orders of magnitude speedup. In addition, because it is orthogonal to the underlying SAT solver, it can help to enhance the efficacy of typical SAT based verification flows.

In this letter, we have analyzed and compared two different multiuser detection receiver structures that are appropriate for code-division multiple-access systems with multiple antennas. We compare the distributed decorrelator (where signal decorrelation is performed at each receiving antenna) and the centralized decorrelator (where decorrelation is performed once collectively). We show that they provide the same performance in additive Gaussian noise channels. The distributed decorrelator provides better performance in flat fading channels.