Context:
Recently, scientists from Tampere University in Finland and Université Marie et Louis Pasteur in France showed that optical fibres—thin glass wires that carry light—can be used in special ways to perform AI tasks like image recognition.
About Light-Based Computing:
Light-based computing is a new way of doing calculations using light particles (photons) instead of electricity (electrons). This method can process data faster, more efficiently, and handle many tasks at once, which is difficult for current electronic computers.
Advantage of light based computing:
-
- Faster signal propagation: Photons travel at the speed of light, and optical signals don’t suffer resistance (ohmic losses) like electrons do. This can reduce latency significantly.
- Higher bandwidth / parallelism: Optical systems can use multiple wavelengths, polarizations, etc., to send many data streams in parallel through the same physical medium. This means more data throughput for e.g. batch processing, communication between AI components.
- Lower energy losses & less heat: Since photons don’t generate heat via resistance, optical data transmission, interconnects, and possibly computation can reduce energy overhead and cooling requirements.
- Overcoming interconnect bottlenecks: Optical interconnects (e.g., silicon photonics) can allow faster, more efficient communication between chips and modules. For AI clusters, where moving data is often the bottleneck, this is a big plus.
- Faster signal propagation: Photons travel at the speed of light, and optical signals don’t suffer resistance (ohmic losses) like electrons do. This can reduce latency significantly.
Key Challenges and Limitations:
-
- Many AI models require nonlinear activations and logistic functions. Optical systems are good at linear operations (e.g. matrix multiplications, Fourier transforms), but nonlinearities are harder to implement purely optically; often one must convert to electrical domain or use special materials.
- Optical memory is underdeveloped compared to electronic RAM or flash. Storing data optically, retaining it reliably, and reading it without converting to electronics is difficult.
- Light traveling through optical media suffers from losses, dispersion (spreading of pulses), and noise. Keeping signal quality high, especially in optical interconnects and computing units, can be challenging.
- Many AI models require nonlinear activations and logistic functions. Optical systems are good at linear operations (e.g. matrix multiplications, Fourier transforms), but nonlinearities are harder to implement purely optically; often one must convert to electrical domain or use special materials.
Applications and Use Cases:
-
- Image Recognition
Achieved over 91% and 93% accuracy using an Extreme Learning Machine (ELM) model with optical fibres. - Neural Network Computation
Light-based systems can perform matrix multiplications, data classification, and signal transformation without electronics. - High-Speed AI Inference
Optical neural networks (ONNs) could one day replace or augment GPUs in AI systems for faster real-time analysis.
- Image Recognition
Conclusion:
Light-based computing could greatly boost AI by overcoming limits like energy use, heat, and speed in electronic systems. However, challenges in design, materials, and cost remain. In the near future, hybrid systems—combining optical and electronic parts—are likely good. As photonic tech improves, more of AI may shift to light-based processing.
