The Institute for Fundamentals of Information Processing at TU Graz and Intel Labs have experimentally demonstrated for the first time that a large neural network can process sequences (such as sentences) four to sixteen times more efficiently on neuromorphic hardware than on conventional hardware. The new research results are based on the neuromorphic research chip Loihi (1st generation) from Intel Labs. Loihi uses findings from neuroscience to create chips modeled on the biological brain. The research results have now been published in Nature Machine Intelligence.

The human brain as a model

Smart machines and intelligent computers that can independently recognize and deduce objects and relationships between different objects are the subject of global artificial intelligence (AI) research. Energy consumption is a major obstacle on the way to a broader application of such AI methods. It is hoped that neuromorphic technology will provide a boost in the right direction. It is modeled on the human brain, which is the world champion when it comes to energy efficiency: its hundred billion neurons only consume around 20 watts to process information, which is not much more energy than an average energy-saving light bulb.

In its work, the team from TU Graz and Intel Labs focused on algorithms that work with temporal processes. For example, the system had to answer questions about a previously told story and identify the relationships between objects or people from the context. The hardware tested consisted of 32 Loihi chips.

Loihi chip: up to sixteen times more energy efficient

"Our system is four to sixteen times more energy efficient than other AI models on conventional hardware," says Philipp Plank, PhD student at the TU Graz Institute for Information Processing Fundamentals. Plank expects further efficiency gains when these models are migrated to the next generation of Loihi hardware, which significantly improves the performance of chip-to-chip communication.

"Intel's Loihi research chips promise advances in AI, particularly by reducing high energy costs," said Mike Davies, Director of the Intel Neuromorphic Computing Lab. "Our work with TU Graz provides further evidence that neuromorphic technology can improve the energy efficiency of today's deep learning workloads by rethinking their implementation from a biology perspective."

Imitation of human short-term memory

In their concept, the group reproduced a presumed method of the human brain, as Wolfgang Maass, doctoral supervisor of Philipp Plank and professor emeritus at the Institute for the Foundations of Information Processing, explains: "Simulations suggest that a fatigue mechanism of a subset of neurons is essential for short-term memory."
The network only needs to test which neurons are currently fatigued in order to reconstruct what information it has previously processed. In other words, previous information is stored in non-activity of neurons, and non-activity consumes the least energy.