A statistical approach to bias in zero-shot learning: the lens of handwriting recognition
本文针对零样本学习中的分类偏差问题，提出了一种统计方法，通过两阶段架构结合经典零样本学习模型与轻量级蒙特卡洛偏差校正器来解决大规模词汇手写识别中的偏差问题。
0 citationsRead paper
Selected work
Representative Papers
Emergent Models: Intelligence from Tiny Substrates
This study addresses the limitations of differentiable feedforward mappings by exploring a novel paradigm for intelligence emergence based on simple computational substrates. Through evolutionary search, we train local recurrent computations in minimal systems such as cellular automata and propose the theory of "Latent Universality," demonstrating that arbitrary partially computable functions can be realized under fixed rules solely by varying initial conditions. Experiments show that models with merely dozens of parameters achieve exact arithmetic extrapolation, behavioral control, and online adaptation. This work validates the feasibility of emergent intelligence within minimal substrates, effectively expanding both the design space and theoretical boundaries of machine learning beyond conventional differentiable architectures.
0 citationsRead paper
Mass Conservation as an Inductive Bias for Self-Organized Criticality in NCA Reservoirs
This work investigates how to effectively induce self-organized criticality (SOC) in neural cellular automata (NCA) reservoirs to enhance information processing capabilities without compromising downstream task performance. To this end, mass conservation is introduced for the first time as an inductive bias into the NCA evolution dynamics, guiding the reservoir to spontaneously approach a critical state. Criticality is quantified via power-law fitting, and performance is evaluated on three benchmark tasks: 5-bit sequence memory, MNIST classification, and CartPole-v1 control. Results demonstrate that mass-conserving NCA evolves 1.27× faster, achieves desirable power-law distributions more consistently, and matches standard NCA performance across all tasks, with the optimally critical reservoir attaining the highest score in the temporal control task.
0 citationsRead paper
Pair2Score: Pairwise-to-Absolute Transfer for LLM-Based Essay Scoring
This work addresses the challenge of automatic essay scoring, which requires predicting absolute scores—a task difficult to model directly. To overcome this, the authors propose Pair2Score, a two-stage framework that first trains a directional siamese ranker to perform pairwise comparisons and then efficiently transfers the learned ranking knowledge to an absolute scoring model via configurable transfer strategies, such as warm-start initialization and embedding fusion. This approach innovatively integrates pairwise comparison with absolute scoring and demonstrates that the choice of transfer configuration critically influences performance. Experimental results across grammatical, lexical, and syntactic tasks show that the best-performing transfer variants consistently and significantly outperform pure absolute-scoring baselines, achieving notable gains in Quadratic Weighted Kappa (QWK) while enabling parameter-efficient fine-tuning based on LLaMA.
0 citationsRead paper
Benchmarking the State of Networks with a Low-Cost Method Based on Reservoir Computing
This study addresses the need for low-cost, non-intrusive monitoring of communication and mobile network states. We propose a novel reservoir computing–based approach that models weighted networks—constructed from readily available aggregated mobile traffic data in Norway—as untrained echo state networks (ESNs), leveraging their intrinsic dynamics as a fixed reservoir. A single-layer linear readout performs a neuroscience-inspired proxy task to assess network dynamical states with minimal energy consumption and without full-network retraining. Experiments demonstrate that model performance degrades significantly under network perturbations, enabling sensitive identification of structural vulnerabilities. Crucially, the method requires neither raw signaling data nor hardware instrumentation, supports near-real-time monitoring, and generalizes to other complex networked systems (e.g., transportation). It thus establishes a scalable, lightweight computational paradigm for resilience assessment of large-scale critical infrastructure.
0 citationsRead paper
Recent publications
Latest Papers
A statistical approach to bias in zero-shot learning: the lens of handwriting recognition
本文针对零样本学习中的分类偏差问题，提出了一种统计方法，通过两阶段架构结合经典零样本学习模型与轻量级蒙特卡洛偏差校正器来解决大规模词汇手写识别中的偏差问题。
0 citationsRead paper
Emergent Models: Intelligence from Tiny Substrates
This study addresses the limitations of differentiable feedforward mappings by exploring a novel paradigm for intelligence emergence based on simple computational substrates. Through evolutionary search, we train local recurrent computations in minimal systems such as cellular automata and propose the theory of "Latent Universality," demonstrating that arbitrary partially computable functions can be realized under fixed rules solely by varying initial conditions. Experiments show that models with merely dozens of parameters achieve exact arithmetic extrapolation, behavioral control, and online adaptation. This work validates the feasibility of emergent intelligence within minimal substrates, effectively expanding both the design space and theoretical boundaries of machine learning beyond conventional differentiable architectures.
0 citationsRead paper
Mass Conservation as an Inductive Bias for Self-Organized Criticality in NCA Reservoirs
This work investigates how to effectively induce self-organized criticality (SOC) in neural cellular automata (NCA) reservoirs to enhance information processing capabilities without compromising downstream task performance. To this end, mass conservation is introduced for the first time as an inductive bias into the NCA evolution dynamics, guiding the reservoir to spontaneously approach a critical state. Criticality is quantified via power-law fitting, and performance is evaluated on three benchmark tasks: 5-bit sequence memory, MNIST classification, and CartPole-v1 control. Results demonstrate that mass-conserving NCA evolves 1.27× faster, achieves desirable power-law distributions more consistently, and matches standard NCA performance across all tasks, with the optimally critical reservoir attaining the highest score in the temporal control task.
0 citationsRead paper
Pair2Score: Pairwise-to-Absolute Transfer for LLM-Based Essay Scoring
This work addresses the challenge of automatic essay scoring, which requires predicting absolute scores—a task difficult to model directly. To overcome this, the authors propose Pair2Score, a two-stage framework that first trains a directional siamese ranker to perform pairwise comparisons and then efficiently transfers the learned ranking knowledge to an absolute scoring model via configurable transfer strategies, such as warm-start initialization and embedding fusion. This approach innovatively integrates pairwise comparison with absolute scoring and demonstrates that the choice of transfer configuration critically influences performance. Experimental results across grammatical, lexical, and syntactic tasks show that the best-performing transfer variants consistently and significantly outperform pure absolute-scoring baselines, achieving notable gains in Quadratic Weighted Kappa (QWK) while enabling parameter-efficient fine-tuning based on LLaMA.
0 citationsRead paper
Benchmarking the State of Networks with a Low-Cost Method Based on Reservoir Computing
This study addresses the need for low-cost, non-intrusive monitoring of communication and mobile network states. We propose a novel reservoir computing–based approach that models weighted networks—constructed from readily available aggregated mobile traffic data in Norway—as untrained echo state networks (ESNs), leveraging their intrinsic dynamics as a fixed reservoir. A single-layer linear readout performs a neuroscience-inspired proxy task to assess network dynamical states with minimal energy consumption and without full-network retraining. Experiments demonstrate that model performance degrades significantly under network perturbations, enabling sensitive identification of structural vulnerabilities. Crucially, the method requires neither raw signaling data nor hardware instrumentation, supports near-real-time monitoring, and generalizes to other complex networked systems (e.g., transportation). It thus establishes a scalable, lightweight computational paradigm for resilience assessment of large-scale critical infrastructure.
0 citationsRead paper