In an era overwhelmed by vast amounts of data, the effective curation of web-crawl datasets is essential for optimizing model performance. This paper tackles the challenges associated with the unstructured and heterogeneous nature of such datasets. Traditional heuristic curation methods often inadequately capture complex features, resulting in biases and the exclusion of relevant data. We introduce an advanced, learning-driven approach, Ensemble Curation Of DAta ThroUgh Multimodal Operators, called EcoDatum, which employs a novel quality-guided deduplication method to balance feature distribution. EcoDatum strategically integrates various unimodal and multimodal data curation operators within a weak supervision ensemble framework, utilizing automated optimization to effectively score each data point. EcoDatum, which significantly improves the data curation quality and efficiency, outperforms existing state-of-the-art (SOTA) techniques, ranking 1st on the DataComp leaderboard with an average performance score of 0.182 across 38 diverse evaluation datasets. This represents a 28% improvement over the DataComp baseline method, demonstrating its effectiveness in improving dataset curation and model training efficiency.

Distributed edge computing platforms are of great significance for the implementation of brain-like computing research. Due to the limited power consumption and real-time requirements, hardware acceleration of computing units is a challenging task. Taking advantage of both scalable hardware framework and lower cost, this article designs a multicore distributed computing platform for spiking neural networks. Particularly, a shared memory partition structure is utilized to participate in hardware acceleration. Through the spike-queue-based synaptic mapping mechanism, each parallel computing unit deals with efficient point-to-point connections. In addition, this platform provides a basic community unit (BCU) that encapsulates a standard neuron model library and rich peripheral interfaces. With the support of GUI, users can quickly build large-scale systems. The experimental results show that a single BCU can accommodate more than 10 000 neurons updated in real-time at a power consumption of 273.6 mW. The extended BCU group is able to perform network dynamic simulations in the basal ganglia-thalamus plausible biological network composed of Hodgkin–Huxley neurons as well as MNIST dataset classification in the leaky-integrate-fire network. The outstanding flexibility and real-time performance of the proposed hardware architecture provide great potential for embedded applications of neural computation.