Vectorizing Classical Tamil: Representation Learning for Verse-Commentary Pairs
研究通过构建古典泰米尔语诗-注释对语料库,使用多种深度学习模型探索信息表示学习,以解决文本内容和结构的自动理解与生成问题。
研究通过构建古典泰米尔语诗-注释对语料库,使用多种深度学习模型探索信息表示学习,以解决文本内容和结构的自动理解与生成问题。
为了解决多智能体系统团队变更导致的评估难题,提出了一种名为COVER的方法,通过固定公共信息边界等手段准确衡量路由效果。
This study investigates whether bilingual Mixture-of-Experts (MoE) routing exhibits linguistic structure by adopting a declarative-procedural framework. Utilizing probing techniques and mutual information to quantify expert specialization, the research integrates cognitive theory into interpretability analysis to reveal emergent language organization patterns under non-curriculum training and the suppressive effects of curriculum learning on monolingual dominance. Results demonstrate that mixed training yields stronger aggregated specialization compared to sequential approaches, with the non-curriculum model achieving a mutual information score of 0.2599. Conversely, curriculum training facilitates stable language-balanced routing. These findings offer novel insights into the internal mechanisms of multilingual models, highlighting the interplay between training strategies and linguistic representation in MoE architectures.
Next-generation tactical networks face a fundamental trilemma under global passive adversaries: strong anonymity, strict isochronicity, and low bandwidth overhead are mutually incompatible. To resolve this, we propose a cross-layer anonymous communication framework that jointly integrates Lyapunov drift-penalty control, robust discrete-time control barrier functions (RaCBFs), and convex sidelobe time-varying modulation (SLTM). Our approach leverages rapid entropy injection via physical-layer antenna sidelobes to achieve near-isochronous, low-overhead anonymous transmission. Crucially, we introduce physical-layer equivalence into anonymity theory—formally proving that entropy growth drives both delay and dummy-packet overhead asymptotically to zero. FPGA-based prototyping demonstrates a 40% increase in anonymity set size, deterministic jitter <30 ms (100% compliance), only a 5% throughput reduction, and significantly lower interception probability compared to state-of-the-art LPI/LPD schemes.
In signed multiplication for edge detection and similar applications, frequent occurrences of constant-1 terms and negative partial products severely degrade energy efficiency. To address this, we propose a sign-aware approximate signed multiplier architecture. Our method introduces two novel compression units—A+B+C+1 and A+B+C+D+1—that explicitly model sign logic to simplify negative-term handling; it further combines truncation of the least-significant (N−1) columns of the partial-product matrix with an error-compensation mechanism, achieving substantial computational savings under controllable accuracy loss. Integrated into a custom convolutional layer, the proposed multiplier, implemented in 8-bit precision, reduces power-delay product by 29.21% and power consumption by 14.39% over the state-of-the-art. Experimental validation on real-world edge detection tasks confirms its superior trade-off between accuracy and energy efficiency.
研究通过构建古典泰米尔语诗-注释对语料库,使用多种深度学习模型探索信息表示学习,以解决文本内容和结构的自动理解与生成问题。
为了解决多智能体系统团队变更导致的评估难题,提出了一种名为COVER的方法,通过固定公共信息边界等手段准确衡量路由效果。
This study investigates whether bilingual Mixture-of-Experts (MoE) routing exhibits linguistic structure by adopting a declarative-procedural framework. Utilizing probing techniques and mutual information to quantify expert specialization, the research integrates cognitive theory into interpretability analysis to reveal emergent language organization patterns under non-curriculum training and the suppressive effects of curriculum learning on monolingual dominance. Results demonstrate that mixed training yields stronger aggregated specialization compared to sequential approaches, with the non-curriculum model achieving a mutual information score of 0.2599. Conversely, curriculum training facilitates stable language-balanced routing. These findings offer novel insights into the internal mechanisms of multilingual models, highlighting the interplay between training strategies and linguistic representation in MoE architectures.
Next-generation tactical networks face a fundamental trilemma under global passive adversaries: strong anonymity, strict isochronicity, and low bandwidth overhead are mutually incompatible. To resolve this, we propose a cross-layer anonymous communication framework that jointly integrates Lyapunov drift-penalty control, robust discrete-time control barrier functions (RaCBFs), and convex sidelobe time-varying modulation (SLTM). Our approach leverages rapid entropy injection via physical-layer antenna sidelobes to achieve near-isochronous, low-overhead anonymous transmission. Crucially, we introduce physical-layer equivalence into anonymity theory—formally proving that entropy growth drives both delay and dummy-packet overhead asymptotically to zero. FPGA-based prototyping demonstrates a 40% increase in anonymity set size, deterministic jitter <30 ms (100% compliance), only a 5% throughput reduction, and significantly lower interception probability compared to state-of-the-art LPI/LPD schemes.
In signed multiplication for edge detection and similar applications, frequent occurrences of constant-1 terms and negative partial products severely degrade energy efficiency. To address this, we propose a sign-aware approximate signed multiplier architecture. Our method introduces two novel compression units—A+B+C+1 and A+B+C+D+1—that explicitly model sign logic to simplify negative-term handling; it further combines truncation of the least-significant (N−1) columns of the partial-product matrix with an error-compensation mechanism, achieving substantial computational savings under controllable accuracy loss. Integrated into a custom convolutional layer, the proposed multiplier, implemented in 8-bit precision, reduces power-delay product by 29.21% and power consumption by 14.39% over the state-of-the-art. Experimental validation on real-world edge detection tasks confirms its superior trade-off between accuracy and energy efficiency.