Searcharxiv⌕ Search

arXiv subjects

Hongjie Liu

Publications and source records attributed to Hongjie Liu.

8 recordsLinked to original sources
arXiv2026

Transfer Learning of CATE with Kernel Ridge Regression

The proliferation of data has sparked significant interest in leveraging findings from one study to estimate treatment effects in a different target population without direct outcome observations. However, the transfer learning process is frequently hindered by substantial covariate shift and limited overlap between (i) the source and target populations, as well as (ii) the treatment and control groups within the source. We propose a novel method for overlap-adaptive transfer learning of conditional average treatment effect (CATE) using kernel ridge regression (KRR). Our approach involves partitioning the labeled source data into two subsets. The first one is used to train candidate CATE models based on regression adjustment and pseudo-outcomes. An optimal model is then selected using the second subset and unlabeled target data, employing another pseudo-outcome-based strategy. We provide a theoretical justification for our method through sharp non-asymptotic MSE bounds, highlighting its adaptivity to both weak overlaps and the complexity of the CATE function. Extensive numerical studies confirm that our method achieves superior finite-sample efficiency and adaptability. We conclude by demonstrating the effectiveness and superiority of our approach on two real-world datasets.

stat.ME↗
arXiv2025

A 33.6-136.2 TOPS/W Nonlinear Analog Computing-In-Memory Macro for Multi-bit LSTM Accelerator in 65 nm CMOS

The energy efficiency of analog computing-in-memory (ACIM) accelerator for recurrent neural networks, particularly long short-term memory (LSTM) network, is limited by the high proportion of nonlinear (NL) operations typically executed digitally. To address this, we propose an LSTM accelerator incorporating an ACIM macro with reconfigurable (1-5 bit) nonlinear in-memory (NLIM) analog-to-digital converter (ADC) to compute NL activations directly in the analog domain using: 1) a dual 9T bitcell with decoupled read/write paths for signed inputs and ternary weight operations; 2) a read-word-line underdrive Cascode (RUDC) technique achieving 2.8X higher read-bitline dynamic range than single-transistor designs (1.4X better over conventional Cascode structure with 7X lower current variation); 3) a dual-supply 6T-SRAM array for efficient multi-bit weight operations and reducing both bitcell count (7.8X) and latency (4X) for 5-bit weight operations. We experimentally demonstrate 5-bit NLIM ADC for approximating NL activations in LSTM cells, achieving average error <1 LSB. Simulation confirms the robustness of NLIM ADC against temperature variations thanks to the replica bias strategy. Our design achieves 92.0% on-chip inference accuracy for a 12-class keyword-spotting task while demonstrating 2.2X higher system-level normalized energy efficiency and 1.6X better normalized area efficiency than state-of-the-art works. The results combine physical measurements of a macro unit-accounting for the majority of LSTM operations (99% linear and 80% nonlinear operations)-with simulations of the remaining components, including additional LSTM and fully connected layers.

cs.AR↗
arXiv2025

P/D-Device: Disaggregated Large Language Model between Cloud and Devices

Serving disaggregated large language models has been widely adopted in industrial practice for enhanced performance. However, too many tokens generated in decoding phase, i.e., occupying the resources for a long time, essentially hamper the cloud from achieving a higher throughput. Meanwhile, due to limited on-device resources, the time to first token (TTFT), i.e., the latency of prefill phase, increases dramatically with the growth on prompt length. In order to concur with such a bottleneck on resources, i.e., long occupation in cloud and limited on-device computing capacity, we propose to separate large language model between cloud and devices. That is, the cloud helps a portion of the content for each device, only in its prefill phase. Specifically, after receiving the first token from the cloud, decoupling with its own prefill, the device responds to the user immediately for a lower TTFT. Then, the following tokens from cloud are presented via a speed controller for smoothed TPOT (the time per output token), until the device catches up with the progress. On-device prefill is then amortized using received tokens while the resource usage in cloud is controlled. Moreover, during cloud prefill, the prompt can be refined, using those intermediate data already generated, to further speed up on-device inference. We implement such a scheme P/D-Device, and confirm its superiority over other alternatives. We further propose an algorithm to decide the best settings. Real-trace experiments show that TTFT decreases at least 60%, maximum TPOT is about tens of milliseconds, and cloud throughput increases by up to 15x.

cs.DC↗
arXiv2024

Topkima-Former: Low-energy, Low-Latency Inference for Transformers using top-k In-memory ADC

Transformer model has gained prominence as a popular deep neural network architecture for neural language processing (NLP) and computer vision (CV) applications. However, the extensive use of nonlinear operations, like softmax, poses a performance bottleneck during transformer inference and comprises up to 40% of the total latency. Hence, we propose innovations at the circuit, architecture, and algorithm levels to accelerate the transformer. At the circuit level, we propose topkima-combining top-k activation selection with in-memory ADC (IMA) to implement a low-energy and low-latency softmax without any sorting latency. Only the k largest activations are sent to the softmax calculation block, reducing the huge computational cost of softmax. Using a modified training scheme with top-k only in the forward pass, experimental results demonstrate only a 0.4% to 1.2% reduction in accuracy across ViT, distilBERT, and BERT-base models when evaluated on CIFAR-10, CIFAR-100, and SQuAD datasets with k=5. At the architecture level, an improved scale-free technique is introduced to reduce the computational cost of attention. The combined system, dubbed Topkima-Former, enhances 1.8x-84x speedup and 1.3x-35x energy efficiency (EE) over prior In-memory computing (IMC) accelerators. Compared to a conventional softmax macro and a digital top-k (Dtopk) softmax macro, our proposed tokima softmax macro achieves about 15x and 8x faster speed respectively.

cs.AR↗
arXiv2024

First Proof of Principle Experiment for Muon Production with Ultrashort High Intensity Laser

Muons, which play a crucial role in both fundamental and applied physics, have traditionally been generated through proton accelerators or from cosmic rays. With the advent of ultra-short high-intensity lasers capable of accelerating electrons to GeV levels, it has become possible to generate muons in laser laboratories. In this work, we show the first proof of principle experiment for novel muon production with an ultra-short, high-intensity laser device through GeV electron beam bombardment on a lead converter target. The muon physical signal is confirmed by measuring its lifetime which is the first clear demonstration of laser-produced muons. Geant4 simulations were employed to investigate the photo-production, electro-production, and Bethe-Heitler processes response for muon generation and their subsequent detection. The results show that the dominant contributions of muons are attributed to the photo-production/electro-production and a significant yield of muons up to 0.01 $μ$/$e^-$ out of the converter target could be achieved. This laser muon source features compact, ultra-short pulse and high flux. Moreover, its implementation in a small laser laboratory is relatively straightforward, significantly reducing the barriers to entry for research in areas such as muonic X-ray elemental analysis, muon spin spectroscopy and so on.

physics.acc-ph↗
arXiv2024

Modeling the Spread of COVID-19 in University Communities

Mathematical and simulation models are often used to predict the spread of a disease and estimate the impact of public health interventions, and many such models have been developed and used during the COVID-19 pandemic. This paper describes a study that systematically compared models for a university community, which has a much smaller but more connected population than a state or nation. We developed a stochastic agent-based model, a deterministic compartment model, and a model based on ordinary differential equations. All three models represented the disease progression with the same susceptible-exposed-infectious-recovered (SEIR) model. We created a baseline scenario for a population of 14,000 students and faculty and eleven other scenarios for combinations of interventions such as regular testing, contact tracing, quarantine, isolation, moving courses online, mask wearing, improving ventilation, and vaccination. We used parameter values from other epidemiological studies and incorporated data about COVID-19 testing in College Park, Maryland, but the study was designed to compare modeling approaches to each other using a synthetic population. For each scenario we used the models to estimate the number of persons who become infected over a semester of 119 days. We evaluated the models by comparing their predictions and evaluating their parsimony and computational effort. The agent-based model (ABM) and the deterministic compartment model (DCM) had similar results with cyclic flow of persons to and from quarantine, but the model based on ordinary differential equations failed to capture these dynamics. The ABM's computation time was much greater than the other two models' computation time. The DCM captured some of the dynamics that were present in the ABM's predictions and, like those from the ABM, clearly showed the importance of testing and moving classes on-line.

q-bio.PE↗
arXiv2023

Optimizing Bus Travel: A Novel Approach to Feature Mining with P-KMEANS and P-LDA Algorithms

Customizing services for bus travel can bolster its attractiveness, optimize usage, alleviate traffic congestion, and diminish carbon emissions. This potential is realized by harnessing recent advancements in positioning communication facilities, the Internet of Things, and artificial intelligence for feature mining in public transportation. However, the inherent complexities of disorganized and unstructured public transportation data introduce substantial challenges to travel feature extraction. This study presents a bus travel feature extraction method rooted in Point of Interest (POI) data, employing enhanced P-KMENAS and P-LDA algorithms to overcome these limitations. While the KMEANS algorithm adeptly segments passenger travel paths into distinct clusters, its outcomes can be influenced by the initial K value. On the other hand, Latent Dirichlet Allocation (LDA) excels at feature identification and probabilistic interpretations yet encounters difficulties with feature intermingling and nuanced sub-feature interactions. Incorporating the POI dimension enhances our understanding of travel behavior, aligning it more closely with passenger attributes and facilitating easier data analysis. By incorporating POI data, our refined P-KMENAS and P-LDA algorithms grant a holistic insight into travel behaviors and attributes, effectively mitigating the limitations above. Consequently, this POI-centric algorithm effectively amalgamates diverse POI attributes, delineates varied travel contexts, and imparts probabilistic metrics to feature properties. Our method successfully mines the diverse aspects of bus travel, such as age, occupation, gender, sports, cost, safety, and personality traits. It effectively calculates relationships between individual travel behaviors and assigns explanatory and evaluative probabilities to POI labels, thereby enhancing bus travel optimization.

cs.LG↗