Spiking neural networks offer dramatic energy efficiency compared to conventional deep learning, but training them has been constrained by a fundamental problem. SNNs evolved to process event streams, sparse temporal data from neuromorphic sensors like event cameras. Static image datasets like ImageNet cannot be directly used. Event based datasets exist but remain small and expensive to collect. This data scarcity has limited SNN development.

A November 2025 paper introduces I2E, an algorithmic framework that converts static images into high fidelity event streams. The method is inspired by microsaccadic eye movements, the tiny involuntary motions that human eyes make constantly. By simulating these movements at high speed, I2E generates synthetic event data 300x faster than previous conversion methods. More importantly, networks trained on this synthetic data transfer successfully to real world neuromorphic sensors, achieving state of the art accuracy.

Event Cameras and the Data Problem

Event cameras, also called Dynamic Vision Sensors, operate differently from traditional cameras. Instead of capturing frames at fixed intervals, they record changes in brightness asynchronously. Each pixel operates independently, triggering an event when brightness crosses a threshold. The output is not a sequence of images but a stream of events, each tagged with pixel location, polarity (brightness increase or decrease), and timestamp.

This design mirrors biological retinas more closely than frame based cameras. It provides high temporal resolution (microsecond scale), low latency, high dynamic range, and extreme efficiency. The camera only transmits data when something changes in the scene.

For spiking neural networks, event streams are a natural input format. SNNs process information through spike timing, making them well matched to the temporal dynamics of event data. Training SNNs on event streams can leverage their biological inspiration more directly than training on static frames.

However, large scale event datasets are scarce. The CIFAR10-DVS dataset, a neuromorphic version of CIFAR-10, contains only 10,000 recordings. N-Caltech101 has 8,709 samples. These are orders of magnitude smaller than the millions of labeled images available for standard computer vision tasks.

Creating event datasets requires either specialized hardware to record event camera data or algorithmic conversion from static images. Recording real event data is time consuming and expensive. Algorithmic conversion methods existed but were slow and often produced low fidelity results.

Technology Readiness Level: TRL 4 (validated in laboratory). The framework has been tested on multiple benchmark datasets with reproducible results, but deployment in production neuromorphic systems requires further engineering.

Microsaccadic Eye Movement Simulation

The I2E framework addresses the conversion speed and fidelity problems by modeling how biological vision systems process static scenes. When humans fixate on a stationary image, our eyes are not actually still. They make continuous small movements called microsaccades, typically 0.1 to 2 degrees of visual angle.

These microsaccades serve multiple functions. They prevent neural adaptation (neurons stop firing when stimulated constantly). They enhance edge detection and fine detail perception. They refresh the retinal image continuously, generating temporal variation from a static scene.

I2E simulates this process algorithmically. It takes a static RGB image, converts it to intensity values, then generates shifted versions by simulating eye movements. Computing the difference between successive shifted intensity maps produces temporal changes analogous to brightness changes that would trigger events in a Dynamic Vision Sensor.

The key innovation is parallelization through convolution operations. Previous methods simulated pixel by pixel comparisons sequentially, making them computationally expensive. I2E reformulates the process as a series of convolutions, which can be executed efficiently on GPUs. This achieves 300x speedup compared to prior approaches and up to 30,000x faster than physically recording event data with actual cameras.

The output is an eight timestep event stream from a single static image. Each event includes spatial coordinates and polarity. The generated streams preserve edge information, motion cues, and temporal dynamics that SNNs can exploit.

Benchmark Performance

To validate the approach, the researchers created I2E-ImageNet by converting the entire ImageNet-1K dataset to event streams. An SNN trained on this synthetic event data achieved 60.50% top-1 accuracy on ImageNet classification. This sets a new state of the art for SNNs on large scale image recognition, demonstrating that synthetic event data can support training of competitive models.

More striking is the sim-to-real transfer result. The team pre-trained an SNN on synthetic I2E-CIFAR10 data, then fine-tuned it on the real world CIFAR10-DVS dataset recorded from an actual event camera. The final model achieved 92.5% accuracy, surpassing the previous best result by 7.7 percentage points.

This validates a critical hypothesis. Synthetic event streams generated by I2E serve as a high fidelity proxy for real sensor data. An SNN can learn meaningful representations from simulated events and transfer that knowledge to real neuromorphic hardware. This opens the possibility of leveraging large static image datasets to overcome the data bottleneck in neuromorphic computing.

The framework also enables on the fly data augmentation during SNN training. Because conversion is fast, different event stream versions of the same image can be generated in real time with varied parameters. This provides regularization and improves generalization without requiring additional labeled data.

Sim-to-Real Transfer in Neuromorphic Vision

Sim-to-real transfer, training on synthetic data and deploying on real systems, is a longstanding challenge in robotics and computer vision. The domain gap between simulation and reality often causes performance degradation.

For neuromorphic vision, this gap has been particularly problematic. Event cameras have unique noise characteristics, temporal jitter, and pixel response variations that are difficult to model perfectly. Previous attempts at simulation produced event streams that looked visually similar but caused trained networks to fail on real hardware.

The I2E results suggest that the microsaccadic movement model captures essential properties of event generation with sufficient fidelity for neural networks to learn robust representations. The fact that an SNN pre-trained on I2E data and fine-tuned on a small real dataset outperforms models trained only on real data indicates the synthetic events provide useful inductive biases.

Recent work on sim-to-real transfer in other domains highlights similar patterns. A June 2025 study on event camera simulation using CARLA found that models trained solely on synthetic data performed well on synthetic test sets but degraded significantly on real data. However, when combined with domain adaptation techniques or fine tuning on real examples, sim-to-real gaps narrowed substantially.

The I2E framework demonstrates that with appropriate biological inspiration (microsaccades) and careful algorithmic design, synthetic neuromorphic data can bridge the simulation-reality divide effectively.

Implications for Neuromorphic Engineering

The data bottleneck has been one of the primary obstacles to widespread adoption of spiking neural networks and neuromorphic computing. Hardware platforms exist. Intel’s Loihi 2, IBM’s TrueNorth, and specialized chips from companies like SynSense provide neuromorphic processors capable of running SNNs efficiently. But without large scale datasets to train on, developing high performance SNN models remained difficult.

I2E provides a scalable solution. Researchers can now leverage existing computer vision datasets, converting them to event streams as needed. This accelerates SNN research by making training data abundant rather than scarce.

The framework also enables new applications. Real time robotics, autonomous vehicles, and drone vision systems benefit from the low latency and efficiency of event cameras. But deploying these systems required custom data collection pipelines. I2E allows developers to prototype using standard datasets, then transfer to real event hardware.

The 300x speedup makes the conversion process practical for large scale use. Generating event streams from ImageNet-1K, a dataset with 1.28 million images, becomes feasible on a single GPU cluster rather than requiring specialized infrastructure.

Combined with recent advances in spiking neural network architectures, the I2E framework positions neuromorphic computing for practical deployment. SNNs can now be trained at scale, achieving competitive accuracy while maintaining the energy efficiency advantages that make them attractive for edge computing and mobile applications.

Connection to Brain Emulation

Event cameras and spiking neural networks are both inspired by biological vision and neural processing. The retina encodes visual information through changes in activity rather than absolute values. Retinal ganglion cells fire when brightness changes in their receptive field, similar to event camera pixels.

Microsaccades play a functional role in biological vision that I2E exploits algorithmically. The constant eye movements ensure that retinal neurons receive temporally varying input, preventing adaptation and maintaining sensitivity. This biological mechanism solves the same problem I2E addresses: generating temporal dynamics from static scenes.

For whole brain emulation, the question is what level of biological detail is necessary. If cognitive processes depend on spike timing with millisecond precision, then event based processing and spiking networks capture essential computational properties. The success of SNNs on vision tasks supports this level of abstraction.

However, if consciousness requires quantum processes in microtubules or other subcellular mechanisms, then even biologically inspired SNNs may miss critical dynamics. The I2E framework demonstrates that certain aspects of biological vision, specifically microsaccadic movements and event based encoding, can be abstracted into algorithms that preserve functional performance.

This suggests a pragmatic path for brain emulation. Start with higher level functional abstractions like spike timing and event driven processing. Test whether these capture behaviorally relevant computation. If they do, detailed subcellular simulation may not be necessary for many cognitive functions.

Path Forward

The I2E framework is open source, with code and datasets released to the research community. This enables reproducibility and accelerates follow up work. Immediate next steps include scaling to larger models, testing on additional neuromorphic benchmarks, and integrating with hardware deployment pipelines.

A key open question is whether I2E generated event streams fully capture the richness of real event camera data. The 92.5% CIFAR10-DVS accuracy suggests high fidelity, but edge cases may exist where synthetic data differs from real sensors. Further analysis of the domain gap and development of techniques to minimize it will improve transferability.

Another direction is extending I2E beyond static images. Video provides temporal information that could be leveraged to generate even more realistic event streams. Combining frame based video with microsaccadic simulation might produce event data that better matches real world camera motion and scene dynamics.

Hardware co-design represents another opportunity. As neuromorphic processors become more capable, optimizing the I2E conversion process for on-chip execution could enable real time synthetic data generation during training or inference. This would allow systems to dynamically adapt to new visual environments without requiring extensive pre-recorded datasets.

The convergence of efficient data generation (I2E), scalable architectures (BrainTransformers), and specialized hardware (Loihi 2, neuromorphic ASICs) positions neuromorphic computing for practical impact. Energy efficient vision systems that operate at millisecond latency with minimal power consumption become feasible for deployment in drones, robots, wearables, and embedded devices.

Whether these systems approach biological neural processing or simply borrow useful principles remains an open question. But the ability to train brain inspired models at scale brings us closer to answering it through experimentation rather than speculation.

Official Sources

Primary Research Paper:

• Ma, R., et al. (2025). “I2E: Real-Time Image-to-Event Conversion for High-Performance Spiking Neural Networks.” arXiv preprint arXiv:2511.08065. arXiv | HTML Version | Hugging Face

Event Camera and DVS Research:

• Li, H., et al. (2017). “CIFAR10-DVS: An Event-Stream Dataset for Object Classification.” Frontiers in Neuroscience, 11, 309. Frontiers Article | Dataset

• Gallego, G., et al. (2020). “Event-based Vision: A Survey.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1), 154-180. RPG ETH Zurich

Neuromorphic Hardware and Applications:

• Chen, Y., et al. (2024). “Hardware, Algorithms, and Applications of the Neuromorphic Vision Sensor: A Review.” Sensors, 25(19), 6208. MDPI Article | arXiv Version

• Zhang, X., et al. (2025). “Dynamic Vision Sensor-Driven Spiking Neural Networks for Low-Power Event-Based Tracking and Recognition.” Sensors, 25(19), 6048. MDPI Article

Sim-to-Real Transfer:

• Wei, Z., et al. (2025). “How Real is CARLA’s Dynamic Vision Sensor? A Study on the Sim-to-Real Gap in Traffic Object Detection.” arXiv preprint arXiv:2506.13722. arXiv

• Martinez-Cantin, R., et al. (2025). “Physical AI: Bridging the Sim-to-Real Divide Toward Embodied, Ethical, and Autonomous Intelligence.” Machine Learning for Computational Science and Engineering. Springer

Related SNN Research:

• Lin, Y., et al. (2022). “Rethinking Pretraining as a Bridge from ANNs to SNNs.” arXiv preprint arXiv:2203.01158. arXiv | Hugging Face

News Coverage:

• Quantum Zeitgeist. (2025). “I2E Enables 300x Faster Image-to-Event Conversion, Achieving 60.50% Accuracy For High-Performance Spiking Neural Networks.” Article

Microsaccade Research:

• Delbruck, T., & Lichtsteiner, P. (2023). “Microsaccade-inspired event camera for robotics.” Science Robotics, 8(82). Science Robotics