The Brain as a Template for AI

The human brain solves many tasks that Artificial Intelligence (AI) aspires to, and its design therefore inspires many aspects of Machine Learning, including artificial neurons, convolutional networks, and other network architectures. Another example of the influences of neurobiology on AI is the layer-wise training of deep learning models, reflecting the progressive maturation and learning in the human brain. While many insights from the exploration of the human brain have already flowed into AI and Machine Learning, there are still unexplored areas such as the role of randomness in neural networks, which could influence future AI developments. Noise, originally seen as a disturbance signal, plays a key role in resonance phenomena and can be useful for neural information processing, especially through the phenomenon of stochastic resonance. Current findings suggest that Artificial Intelligence (AI) and deep neural networks can improve their performance by adding noise as additional input, especially in tasks such as speech recognition. Random, untrained connections in neural networks, as discovered in the olfactory system of the fruit fly, can contribute surprisingly effectively to information processing. Artificial neural networks with fixed binary random connections have shown improved accuracy in classifying noisy input data, and the lottery ticket hypothesis suggests that more efficient networks can be achieved by identifying and training favorably initialized subnetworks. In addition, reservoir computingReservoir Computing, which is based on randomly generated, highly recurrent neural networks, provides further insights into the usefulness of random network architectures.