Interesting excerpt from an AI TIME JOURNAL interview with Peter Van Der Made dated 26.12.2019. It provides a good insight into what's about to be delivered and what sets Akida apart from the rest. Well worth reading.


“AI to create an AI” how do you emphasize this statement from the perspective of human intelligence associated with it?

Current AI is not intelligent. Deep Learning systems don’t learn, they are trained using a sequential optimizing routine that feeds back errors and corrects weights, not unlike successive approximation routines; it makes a guess, checks the error, and correct the guess by half of what the error was until the output value and the guessed value converge. Convolutional Neural Networks are computational constructs that have very little in common with the way the brain works.

In the brain, time is of the essence. Information is encoded in the timing of ‘spikes’ – short bursts of electrical energy that are sent between neural cells. The interval between spikes, the intensity of spikes and location where spikes occur all contain information. Synapses store information that is released by incoming spikes. The information stored in synapses is constantly updated. Learning is a function of the timing of spikes. The brain is a very dynamic system that is changing all the time. Intelligence is shaped by its environment through constant learning.

As I stated earlier, the brain has a very defined structure, which varies in different brain regions. The brain is not one homogeneous mass of neural cells. We have a right and left hemisphere, which look much the same. But we also have a cerebellum and a hippocampus, the limbic system and other brain regions each with their own structure specific to their function. Even insect brains are far more structured and intricate than our current neural networks. Brains predict the next action before sensory stimuli arrive. None of that is present in today’s neural networks. To say that today’s neural networks exhibit the intelligence of a honeybee is a blatant exaggeration.

The BrainChip Akida technology is using a brain-inspired spiking neural network to perform inference. It can do all the things that today’s convolutional neural networks can do, but it can also run completely native spiking neural networks that resemble the learning method and processing method of the brain. For instance, to perform cyber-security threat recognition and incremental learning. Future versions of the Akida technology will incorporate more of the structure of the brain, with the aim to make AI more intelligent. This is no threat to human intelligence. With our 86 billion neural cells, 100 trillion synapses and 300 billion glial cells we are way ahead of any intelligent AI for some foreseeable future. By comparison, the largest AI networks today are up to a few million neuron equivalents and miss all of the structure that makes the brain intelligent.

To sustain the market’s hunger, it is very essential for any technology organization to excel in innovation & initiatives. How BrainChip is bracing up for this?

BrainChip’s aim, from inception, has always been to create better AI. We don’t follow the general trend in the market with their massive parallel multipliers, and up to 200 layers that are promoted as AI chips. We defined our event-based Spiking Neural Network technology before 2008, when we filed our first patent. We have accomplished everything that Deep Learning has to offer, using a very different philosophy, a philosophy that has a clear path to the future. Where standard CNN technology runs into a brick wall when they try to go beyond image classification, our event-based Spiking NN excels.

Our philosophy is based on copying the function and structure of the brain, and to apply that technology to solve today’s problems. That is why Akida can be used to process, after a simple conversion process, today’s Deep Learning based CNNs, as well as going forward from that point to do incremental learning and on-chip training. This has many advantages, beyond incremental learning. Akida can learn from the environment in which it exists, an ability that we will expand in future generations of Akida with episodic memory, that is to remember sequences of events. An example of sequence memory is when you retrace your steps to find your car in a busy parking lot. Sequence memory has real-world applications in text and speech interpretation and robotics.

It can also process large amounts of video or data at an extremely low power consumption, which is good for the environment. It was stated that training one Deep Learning CNN uses enough power to run five electric cars for a lifetime. With Akida, that power requirement is reduced to the power needed to run a flashlight. Event-based processing is a ‘green’ technology.