An EU project is trying to mimic the processing of biological nervous systems

The results will be used to build neuromorphic computing systems that can efficiently process real-world sensory signals and natural time series data in real time. Target applications include multidimensional distributed environmental monitoring, implantable chips for medical diagnosis, wearable electronics and human-computer interfaces

Artificial intelligence is considered the enabling computing technology of the technological innovations in the coming years. The Internet of Things already makes extensive use of deep learning computing paradigms to enable Internet information search services or to identify audio-visual information, while the emerging Internet of Things will manage and provide services that process data from billions of networked sensors.

CEA-Leti has announced its participation in the new EU MeM-Scales project, which aims to develop a class of algorithms, devices and circuits that mimic the processing at multiple time scales of biological neural systems.

The results will be used to build neuromorphic computing systems that can efficiently process real-world sensory signals and natural time series data in real time and to demonstrate the ideas through a practical prototype in the laboratory. Target applications include multidimensional distributed environmental monitoring, implantable chips for medical diagnosis, wearable electronics and human-computer interfaces. 

In order to interact with the real world, the brain processes and perceives the sensory signals on multiple time scales, noted Elisa Vianello, director of the artificial intelligence program at CEA-Leti, in an interview with EE Times Europe.

"Memory of this interaction is formed on time scales ranging from hundredths of a second (short-term memory) to months and years (long-term structural changes)," Vianello said. "To design systems that interact with the real world, neuromorphic circuits need to mimic the processing at multiple time scales of the brain. Therefore, these circuits are the critical components in the processing pipeline."

In a standard model of a neural network, input data is first sent to the input neurons and then transmitted through hidden layers of other neurons through connections called synapses. The data is modified at each step, and the output from one layer is used as the input of the next layer.

The data eventually reaches the final output layer, which provides the prediction - for example, a classification into a category or a numerical value in regression. There is no element of real time here - the input data is all transmitted at the same time, passes through each of the hidden layers in order, and is emitted all at once.

But what if the input data doesn't all arrive at the same time in a clean way - what if it's a time series or data that depends on time in some other way, such as real-time input from sensors in a self-driving car? And what if this is also the case with the results - and what if the results are also based on time, like instructions given to a self-driving car when to turn and when to increase or decrease speed?

Neural networks with sharp increases (SNNs) are a solution to this problem. They can receive time-based inputs and produce time-based outputs. Instead of neat layers, the content contains more complex structures for transferring data between nerve cells, such as loops or multi-directional connections. Because they are more complex, they require different types of training and learning algorithms, such as making modifications to backpropagation-like approaches to adapt to spike behavior.

More of the topic in Hayadan:

• What is the relationship between bacteria and neural networks?
• We are developing neural networks that will allow us to process a huge amount of texts and summarize their meaning
• Neurons search for meaning