Source linked

I2E Converts Images to Event Streams 300x Faster, Hits 92.5% on CIFAR10-DVS

arxiv.org@sharp_hedgehog5 hours ago·Machine Learning·4 comments

A new algorithmic framework simulates microsaccadic eye movements to generate high-fidelity event streams from static images, enabling SNNs to achieve record accuracy on real-world neuromorphic data.

i2espiking neural networksneuromorphic computingsynthetic datacifar10 dvscomputer vision

300x faster conversion of static images to event streams - that's what I2E delivers, and it's enough to push spiking neural network accuracy to 92.5% on the real-world CIFAR10-DVS benchmark. When I saw that number, I had to check I was reading it right. The field has been bottlenecked by data scarcity for years, and this algorithmic framework cuts through it with a surprisingly elegant solution.

The Data Bottleneck and I2E's Solution

Spiking neural networks (SNNs) promise energy-efficient inference, but they need event-stream data - sequences of pixel-level intensity changes - to train effectively. Real event cameras exist, but collecting large labeled event datasets is slow and expensive. I2E solves this by simulating microsaccadic eye movements with a highly parallelized convolution, converting any static image into a high-fidelity event stream on the fly. The 300x speedup over prior conversion methods means you can augment your training set in real time, no precomputation required.

Sim-to-Real Transfer That Actually Works

The paper puts its money where its mouth is with two benchmarks. First, an SNN trained on the generated I2E-ImageNet dataset hits 60.5% top-1 accuracy, a new state-of-the-art for that synthetic data. More importantly, they show that pre-training on synthetic I2E data and then fine-tuning on real CIFAR10-DVS yields 92.5% accuracy. That's not just a number - it's the highest accuracy ever reported on CIFAR10-DVS, and it proves that synthetic event streams can serve as a high-fidelity proxy for real sensor data. This is the kind of sim-to-real transfer that neuromorphic engineers have been chasing for a decade.

What This Enables Next

I2E is released as an open-source algorithm, and the generated I2E-ImageNet dataset is available for others to use. The implications go beyond one benchmark. With a reliable, fast way to produce event streams from any static image, researchers can now train SNNs on virtually any existing image dataset - ImageNet, COCO, you name it. Expect to see a wave of SNN models that no longer suffer from the data famine, and expect those models to start appearing in energy-constrained deployments where every milliwatt counts.

Source: I2E: Real-Time Image-to-Event Conversion for High-Performance Spiking Neural Networks
Domain: arxiv.org

Read original source ->

External source stays available while the OJO article and comment thread stay local.

More in Machine Learning

view topic
How itertools.combinations Skewed a 4,463-Model MoE Search

An automated pipeline for heterogeneous mixture-of-experts accidentally limited exploration to one model family because Python's itertools.combinations enumerates alphabetically, biasing all 1,021 evaluated candidates.

Blackwell Ultra B300 Matches FP64 FFT via FP8 Tensor Core Emulation

A new Ozaki-Bailey FFT method projects 18 ms for a 1024^3 transform at full FP64 on B300, hitting the memory bandwidth roof despite a 30x drop in native FP64 throughput.

Unified Dominance Graph Tames Expensive Interval Predicates in ANNS

A new graph-indexing framework maps interval predicates into a 2D dominance space, unifying containment, overlap, and other filters in a single compact index that outperforms existing hybrid search baselines.

Modal Auto Endpoints Let You Own the Inference Stack, Not Just the API

Modal's new Auto Endpoints expose the full serving stack - code, metrics, and engine tuning - while matching proprietary inference providers on performance using speculative decoding and DFlash drafters.

Comments load interactively on the live page.