Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
FLAME: Fast Long-context Adaptive Memory for Event-based Vision
Authors: Biswadeep Chakraborty, Saibal Mukhopadhyay
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct empirical evaluation of FLAME to demonstrate its effectiveness and efficiency in processing asynchronous, event-driven data from event cameras across various challenging benchmarks. Our experiments are designed to assess: (1) performance on demanding event-based vision datasets, including high-resolution and complex activity scenarios; (2) computational efficiency in terms of FLOPs, parameters, and inference latency, including profiling on diverse hardware backends; and (3) the specific contributions of key architectural components through targeted ablation studies. |
| Researcher Affiliation | Collaboration | Biswadeep Chakraborty Georgia Institute of Technology, Mercuria Energy Trading EMAIL Saibal Mukhopadhyay School of Electrical and Computer Engineering Georgia Institute of Technology |
| Pseudocode | Yes | Algorithm 1 FLAME Model Training |
| Open Source Code | Yes | The paper confirms that the full codebase used for all experiments including the complete Event-Aware Hi PPO module with NPLR and FFT support, the full EAL implementation, spatial pooling routines, and all necessary training/evaluation scripts is available in the author s github. |
| Open Datasets | Yes | We evaluate FLAME on a diverse set of public benchmarks to assess its capabilities in event-based vision. For event-based vision, these include DVS Gesture [7] (gesture recognition), HARDVS [61] (human activity), Celex-HAR [2] (high-resolution human activity), CIFAR10-DVS [62] (neuromorphic image classification), and N-Caltech101 [63] (neuromorphic object recognition). Additional evaluations cover event-based speech datasets, namely Spiking Heidelberg Digits (SHD) and Spiking Speech Commands (SSC) [64], and Sequential CIFAR-10/100 [65] for sequential image classification, with further details provided in Appendix B. |
| Dataset Splits | No | Comprehensive details on all dataset characteristics and specific task setups are available in Appendix B.1. |
| Hardware Specification | Yes | Main training and inference for event datasets are conducted on NVIDIA A100 GPUs. We measured inference latency on an NVIDIA A100 GPU (40GB VRAM). The performance of FLAME-Tiny, in terms of inference accuracy and latency, was measured on several common hardware backends, including a standard CPU and various NVIDIA GPUs. The results... are summarized in Table 10. Table 10 lists: Intel Xeon CPU (2 v CPUs @2.2GHz), NVIDIA T4 GPU, NVIDIA V100 GPU (32GB), NVIDIA A100 GPU (40GB). |
| Software Dependencies | No | FLAME models are implemented in Py Torch. |
| Experiment Setup | No | Further details on hyperparameters and training procedures are in Appendix B. |