Continual Learning with Pre-trained Models holds great promise for efficient adaptation across sequential tasks. However, most existing approaches freeze PTMs and rely on auxiliary modules like prompts or adapters, limiting model plasticity and leading to suboptimal generalization when facing significant distribution shifts. While full fine-tuning can improve adaptability, it risks disrupting crucial pre-trained knowledge. In this paper, we propose Mutual Information-guided Sparse Tuning (MIST), a plug-and-play method that selectively updates a small subset of PTM parameters, less than 5%, based on sensitivity to mutual information objectives. MIST enables effective task-specific adaptation while preserving generalization. To further reduce interference, we introduce strong sparsity regularization by randomly dropping gradients during tuning, resulting in fewer than 0.5% of parameters being updated per step. Applied before standard freeze-based methods, MIST consistently boosts performance across diverse continual learning benchmarks. Experiments show that integrating our method into multiple baselines yields significant performance gains.

Implicit sentiment analysis (ISA) is particularly sensitive to user characteristics due to the absence of explicit sentiment cues. While existing approaches leverage explicit user attributes and social relationships, they neglect the implicit interest preferences embedded in user content and multimodal commonsense knowledge. This article introduces a novel personalized ISA framework that systematically integrates heterogeneous user knowledge with multimodal commonsense to address this limitation. Our core innovation lies in a multi-stage knowledge integration pipeline that first captures rich semantic representations through a large language model, then constructs a comprehensive user profile by fusing multiple views of implicit interests derived from user-multimodal commonsense-content interactions. Specifically, we employ graph neural networks to distill structured knowledge from automatically constructed multimodal commonsense graphs, which enhances semantic understanding. The different perspectives of user interests are then systematically fused to capture implicit preference characteristics. Finally, we introduce an adaptive gated fusion mechanism that dynamically incorporates heterogeneous user knowledge and multimodal commonsense into implicit sentiment semantics, enabling personalized analysis capabilities. Extensive experiments on two public personalized ISA Chinese datasets demonstrate that our method outperforms baselines by at least 2.86% and 3.03%, respectively, validating its effectiveness in comprehensive and personalized modeling of implicit sentiment.

We present an all-digital programmable machine learning accelerator chip for image classification, underpinning on the Tsetlin machine (TM) principles. The TM is an emerging machine learning algorithm founded on propositional logic, utilizing sub-pattern recognition expressions called clauses. The accelerator implements the coalesced TM version with convolution, and classifies booleanized images of$28\times 28$pixels with 10 categories. A configuration with 128 clauses is used in a highly parallel architecture. Fast clause evaluation is achieved by keeping all clause weights and Tsetlin automata (TA) action signals in registers. The chip is implemented in a 65 nm low-leakage CMOS technology, and occupies an active area of 2.7 mm2. At a clock frequency of 27.8 MHz, the accelerator achieves 60.3 k classifications per second, and consumes 8.6 nJ per classification. This demonstrates the energy-efficiency of the TM, which was the main motivation for developing this chip. The latency for classifying a single image is$25.4~\mu $s which includes system timing overhead. The accelerator achieves 97.42%, 84.54% and 82.55% test accuracies for the datasets MNIST, Fashion-MNIST and Kuzushiji-MNIST, respectively, matching the TM software models.

The subtlety of emotional expressions makes implicit emotion analysis (IEA) particularly sensitive to user-specific characteristics. Current studies personalize emotion analysis by focusing on the author but neglect the impact of the intended reader on implicit emotional feedback. In this paper, we introduce Personalized IEA (PIEA) and present the RAPPIE model, which addresses subjective variability by incorporating reader feedback. In particular, (1) we create reader agents based on large language models to simulate reader feedback, overcoming the issue of “spiral of silence effect” and data incompleteness of real reader reaction. (2) We develop a role-aware multi-view graph learning to model the emotion interactive propagation process in scenarios with sparse reader information. (3) We construct two new PIEA datasets covering English and Chinese social media with detailed user metadata, addressing the text-centric limitation of existing datasets. Extensive experiments show that RAPPIE significantly outperforms state-of-the-art baselines, demonstrating the value of incorporating reader feedback in PIEA.

Tracking multiple objects in videos captured by unmanned aerial vehicles is challenging due to sudden viewpoint changes, non-linear motion trajectories, and rapid variations in target size and appearance. Existing methods often struggle to handle these complexities, as they rely heavily on handcrafted geometric constraints and fail to adapt to significant field-of-view changes. To address these issues, this paper presents the Dynamic Field-Aware Multi-Object Tracker (DFA-MOT), a joint detection and tracking framework that integrates detection and motion prediction into a unified model, enhancing tracking performance in dynamic UAV environments. The proposed Dynamic Field-of-View Consistency Learning (DFCL) module mitigates geometric distortions caused by UAV movement by leveraging optical flow and learnable deformation operations to achieve progressive spatial alignment. A Scale-Aware Tracking (SAT) mechanism is explored, which enables to accurately predict of both position and scale variations, enhancing the model’s adaptability to variations in target size. By combining detection with predictive motion modeling, DFA-MOT effectively overcomes the limitations of traditional manual constraints. Extensive experiments on the VisDrone2019 and UAVDT datasets demonstrate that DFA-MOT significantly outperforms state-of-the-art tracking methods in UAV scenarios.

Physically Unclonable Functions (PUFs) need error correction whilst regenerating Secret Keys in cryptography. The proposed 8-Transistor (8T) PUF, which coordinates with the power gating technique, can significantly accelerate a single evaluation cycle 1000 times faster than$\mathbf{6}\mathbf{T}$-SRAM PUF does with a 12.8% area increase. This design enables multiple evaluations even in the key regeneration phase in field, hence greatly reducing the number of errors and the hardware penalty for error correction. The$\mathbf{8T}$PUF derives from the$\mathbf{6}\mathbf{T}$SRAM. It is built to eliminate data retention swiftly and maximise physical mismatches. And a two-phase power gating module is designed to provide controllable power-on/off cycles rapidly for the chosen PUF clusters in order to facilitate statistical measurements and curb the in-rush current, thereby enhancing PUF entropy and security. An architecture of the power-gated PUF is developed to accommodate fast multiple evaluations. Post-layout Monte Carlo simulations were performed with Cadence, and the extracted PUF Responses were processed with Matlab to evaluate the 8T PUF performance and statistical metrics for subsequent inclusion into PUF Responses.

In-memory computing for Machine Learning (ML) applications remedies the von Neumann bottlenecks by organizing computation to exploit parallelism and locality. Non-volatile memory devices such as Resistive RAM (ReRAM) offer integrated switching and storage capabilities showing promising performance for ML applications. However, ReRAM devices have design challenges, such as nonlinear digital-analog conversion and circuit overheads. This paper proposes an In-Memory Boolean-to-Current Inference Architecture (IMBUE) that uses ReRAM-transistor cells to eliminate the need for such conversions. IMBUE processes Boolean feature inputs expressed as digital voltages and generates parallel current paths based on resistive memory states. The proportional column current is then translated back to the Boolean domain for further digital processing. The IMBUE architecture is inspired by the Tsetlin Machine (TM), an emerging ML algorithm based on intrinsically Boolean logic. The IMBUE architecture demonstrates significant performance improvements over binarized convolutional neural networks and digital TM in-memory implementations, achieving up to a 12.99x and 5.28x increase, respectively.

RGBT tracking can be accomplished by constructing Bayesian estimators that incorporate fusion prior distributions for the visible (RGB) and thermal (T) modalities. Such estimators enable the computation of a posterior distribution for the variables of interest to locate the target. Incorporating rich prior information can improve the performance of predictors. However, current RGBT trackers face limited fusion prior data. To mitigate this issue, we propose a novel tracker, BD$^{2}$Track, which employs a diffusion model. Firstly, this paper introduces a dumbbell diffusion model, and employ convolution networks and the dumbbell model to derive the fusion feature prior information from various index frames in the same tracking video sequence. Secondly, we propose a plug-and-play channel augmented joint learning strategy to derive the images prior distribution. This strategy not only homogeneously generates modality-relevant prior information but also increases the distance between positive and negative samples within the modality, while reducing the distance between modalities during fusion. Results demonstrate promising performance in the GTOT, RGBT234, LasHeR, and VTUAV-ST datasets, surpassing other state-of-the-art trackers.