Artificial intelligence has traditionally been derived from the dynamics of the human brain. However, compared to deep learning, brain learning has several important limitations (DL). First, effective DL wire topologies (architectures) have several tens of feedforward (consecutive) layers, while brain dynamics have only a few feedforward layers.
Second, DL architectures typically consist of many successive layers of filters that are necessary to identify one of the input classes. For example, if the input is a car, the first filter identifies the wheels, the second the door, the third lights up, and after many more filters it is clear that the input object is indeed a car. Conversely, brain dynamics contains only a single filter located close to the retina. The last necessary part is a mathematically complex DL training procedure, which is obviously far beyond biological realization.
Can the brain, with its limited implementation of precise mathematical operations, compete with advanced artificial intelligence systems implemented on fast and parallel computers? We know from our daily experience that for many tasks the answer is yes! Why is this, and given this positive answer, can a new type of effective brain-inspired artificial intelligence be created? In a paper published today in Scientific Reports, researchers from Bar-Ilan University in Israel tackle this puzzle.
Prof. Ido Kanter of Bar-Ilan’s Department of Physics and Gonda (Goldschmied) Multidisciplinary Brain Research Center, who led the research said “We have shown that efficient learning on an artificial tree architecture, where each weight has a single path to the output unit, can achieve better classification success than previously achieved by DL architectures consisting of multiple layers and filters. This finding opens the way for efficient, biologically inspired new hardware and artificial intelligence algorithms”.
Highly pruned tree architectures represent a step towards a plausible biological realization of efficient learning of dendritic trees by one or a few neurons with reduced complexity and energy consumption, and a biological realization of the backpropagation mechanism that is currently a central technique in AI.
The efficient learning of dendritic trees is based on previous research by Kanter and his experimental research team—and led by Drs. Roni Vardi – suggesting evidence for sub-dendritic adaptations using neuronal cultures, along with other anisotropic properties of neurons such as different spike curves. , refractory periods and maximum transmission rates.
Efficient implementation of highly pruned tree training requires a new type of hardware that differs from emerging GPUs that are better suited to the current DL strategy. In order to effectively mimic the dynamics of the brain, the emergence of new hardware is necessary.
written by: Vaishali Verma
Read Now :What is visual positioning system (VPS) and how it can replace GPS in future
