A team of researchers from the George Washington University, Queens University, University of British Columbia and Princeton University have discovered that photonic chips can speed up the training in machine learning.
Training a machine learning model involves feeding it a large amount of data and adjusting the model’s parameters based on the performance of the model itself.
The training process can be time-consuming, especially for large and complex datasets, like in the case of Tesla’s autopilot. This process requires technology akin to supercomputers and it can cost up to several millions of dollars in electricity to run.
To fill the gap between computer hardware and recent demand for AI, the multi-institution research team has developed an optical chip that can train a machine learning system significantly faster.
What is a photonic chip?
A photonic chip, also known as an optical chip, is a type of integrated circuit that uses light signals instead of electrical signals to transmit and process data.
Optic chips are made from various materials such as silicon, indium phosphide, and silicon-on-insulator. The photonic chips are used in a variety of applications, including high-speed communications, military systems, and medical imaging.
Why use optical chips?
There are several advantages of photonic chips over traditional electronic chips. They provide:
- Faster data transmission speeds: photonic chips can transmit data using light signals, which can travel much faster than electrical signals. This enables photonic chips to transmit data at much higher speeds than traditional electronic chips.
- Lower power consumption: photonic chips consume less power than traditional electronic chips because they do not require electrical current to transmit data. This makes them more energy-efficient and suitable for use in portable or low-power devices.
- Longer transmission distances: photonic chips can transmit data over longer distances without signal loss, which makes them useful for applications that require long-distance communication, such as satellite communications or intercontinental data transmission.
- Resistance to electromagnetic interference: photonic chips are resistant to electromagnetic interference, making them suitable for use in environments where electronic chips may be disrupted by electromagnetic fields.
- Ability to perform multiple calculations simultaneously: photonic chips can perform multiple calculations simultaneously, improving the efficiency of certain types of computations.
Conclusions, future research
Photonic chips may offer faster data transmission speeds compared to the traditional electronic ones. They have the potential to improve the performance of machine learning systems in many ways:
- By implementing neural networks on the chip itself, which are traditionally implemented using transistors, the speed of the operations can be much faster and energy-efficient.
- The development of machine learning algorithms that are specifically designed to run on photonic hardware.
- In the case of autonomous vehicles or surveillance systems which require the ability to analyze and respond to real time data, photonic chips can outperform traditional electronic communication methods.
- In addition to these applications, photonic chips may also be used to improve the efficiency of machine learning systems by reducing the amount of power required for computations and by enabling the transmission of data over long distances without loss of quality.
Although ML models are currently trained using conventional digital circuits, there has been significant interest in utilizing photonic processors. The latest developments in the use of photonic chips suggest that using a silicon photonic architecture offer highly parallel, effective, and quick data processes.
Learn more
- Teaching photonic chips to ’learn’ (on Science Daily)
- Silicon Photonics Platform Provides On-Chip Neural Network Training (on Photonics Media)
- Research paper: “Silicon photonic architecture for training deep neural networks with direct feedback alignment” (on Optica)
- Related studies: “A photonic chip-based machine learning approach for the prediction of molecular properties” (on ArXiv)
