Artificial intelligence is faster in the home through light

Pregúntele mediate voz a un dispositivo inteligente de su hogar what time is it, y el aparato tardará varios segundos en responder. One of the reasons why this latency is produced is that the devices connected do not have sufficient memory and power to store and execute the huge models of automatic learning (a modality of artificial intelligence) necessary for the device to understand what the user pide The model is stored in a data center that can be hundreds of kilometers away, where the answer is calculated and sent to the device.

A team that includes Alexander Sludds and Dirk Englund, both from the Instituto Tecnológico de Massachusetts (MIT) in the United States, has created a new method to calculate directly in these devices and drastically reduce this latency. Su técnica desplaza los pasos que más memoria require a un servidor central donde los componentes del modelo se codifican en ondas de luz.

Las ondas se transmiten a un dispositivo conectado mediate fibra óptica, lo que permite enviar a large amount of data at a huge speed through a network. El receptor emplea entonces un simple dispositivo optico que realiza rapide los calculus utilizando las parts de un modelado por esas ondas de luz.

This technique allows the energy efficiency to be multiplied by more than one hundred in comparison with other methods. También podría mejorar la seguridad, ya que los datos de un usuario no necesitan ser transferidos a un centro de datos para su calculus.

This method could allow a self-driving car to make decisions in real time using only a small percentage of the energy that high-consumption computers currently need. It could also allow a user to maintain a conversation without latency with his smart home device, be used for live video processing over cellular networks or even allow the classification of images at high speed on board a spaceship at millions of kilometers de la Tierra.

“Every time you want to run a neural network, you have to execute the program, and the speed with which you can run the program depends on the speed with which you can channel the program from memory. Our pipeline is enormous: it’s equivalent to sending a full movie to the internet every millisecond, more or less. Asi de rápido llegan los datos a nuestro sistema. Y puede calcular tan rápido como eso,” explains Englund.

[Img #67231]

The new intelligent transceiver uses silicon photonic technology to accelerate spectacularly one of the phases that require more memory from the process of executing an automatic learning model. (Illustration: Alex Sludds, with modifications from MIT News. CC BY-NC-ND 3.0)

Neuronal networks are models of automatic learning that use layers of connected nodes and artificial neurons to recognize patterns in sets of data and perform tasks such as classifying images and recognizing speech. Pero estos models pueden contener miles de milliones de valores numéricos que transforman los datos de entrada a medida que se procesan. Estos valores deben amasserense en la memoria. At the same time, the data transformation process involves thousands of algebraic calculations, whose realization requires a large amount of energy.

El proceso de sacar los datos de la memoria y trasladarlos a las partes de un ordenador que realizan el calculo es uno de los majores factors que limitan la velocidad y la eficiencia energetica.

With Netcast, the neural network architecture developed by Sludds and his colleagues, the storage of the aforementioned numerical values ​​is realized in a central server that is connected to a new hardware component called an intelligent transceiver. This smart transceiver, a chip the size of a thumb that can receive and transmit data, uses a technology known as silicon photonics to obtain billions of numerical values ​​from memory every second.

Recibe los valores como señales eléctricas y los plasma en ondas de luz. Como los datos están codificados en forma de bits (ceros y unos), el transceptor los convertivo mediate la conmutación laser; un láser se enciende para un 1 y se apaga para un 0. Combina estas ondas de luz y luego las transferiere periodicamente través de una red de fibra óptica al dispositivo cliente.

Once the light waves reach the client device, a simple optical component known as a wide band Mach-Zehnder modulator is used to realize an ultra-fast analog calculation. This includes codifying device input data, such as sensor information, in numerical values. Then it sends each individual wavelength to a receptor that detects light and measures the result of the calculation.

Sludds and his colleagues exponen los detalles de su nuevo sistema en la revista académica Science, bajo el título “Delocalized Photonic Deep Learning on the Internet’s Edge”. (Fuente: NCYT de Amazings)

.

Leave a Comment

Your email address will not be published. Required fields are marked *