Loihi – How Intel Is Replicating the Human Brain in Silicon
The applications for AI reach into almost every aspect of everyday life. Some seem a little trivial and frivolous, while others have the power to make enormous contributions to society. Some of the aspects holding AI projects back are limitations in terms of size, power consumption, price or the lack of specific conditions to make AI implementations even possible.
This could all change with the recent launch of Loihi (spoken low-ee-hee), a neuromorphic research chip created by Intel. The device is designed to operate in the same manner as the brain and features around 130,000 neurons as well as 130 million synapses which however still puts it below the complexity of a common house fly’s brain.
So, how will this device make a difference? Mostly, it is the processing approach it uses to support AI algorithms. Today’s optimal AI processing platforms make use of digital graphics processing units (GPUs). Due to their highly parallelized architectures, they are excellent at executing the core math algorithms that underpin AI. They use SIMD processing (single instruction, multiple data) which enables them to execute a single instruction that utilizes multiple input data sources in a single clock cycle. Despite the cleverness of this technology, the brain’s neurons and synapses are still essentially being simulated in AI software.
Loihi changes the approach completely by offering an electronic chip that replicates the structure of the brain with neurons linked to one another by synapses. This is achieved by combining both digital and analog electronic technologies. Just like the human brain, there is no need for a central clock to synchronize instruction execution. Instead, a neuromorphic processor architecture is used, a machine that is comprised of many simple memory structures that communicate using simple messages. The simple messages function in a similar way to the brain’s synapses, replicating the way ions in the brain flow between cells for communication.
Another improvement is its huge reduction in power consumption compared to existing alternative solutions – it is namely a factor of up to 1,000 times lower. This is likely to benefit applications found in the area of autonomous vehicles for instance. The fully autonomous vehicle prototypes that exist nowadays can require as much as 2,500 watts of energy to function. This results in either higher fuel consumption for a traditional vehicle, or a reduced range for an electric vehicle. Loihi could therefore be critical in making autonomous vehicles a reality.
A Loihi-based AI could also learn and compensate for changes while in use, unlike most of today’s AI implementation. Today’s AI algorithms, once trained, can be confused by slight changes to the input data. For example, if trained to recognize faces, the addition of glasses or someone with a beard may cause the AI to fail to successfully recognize a face. To resolve this issue the AI would need to be taken off-line and retrained with appropriate images. The Loihi AI, just like the human brain, would be able to undertake this learning on-the-fly. Just like us, once it has been told what is “normal”, it could learn and accommodate for small anomalies on its own.
So, is Loihi the technology that will transform AI and bring it to the masses? Well, it will take some time to determine how best to leverage the capability of these new neuromorphic devices, but knowing the pace of research in the field of AI it won’t take them long to discover where their true strengths lie.
 As a reference, the human brain has around 100 billion neurons.