Deep learning has shown considerable promise in recent years, producing tremendous results and significantly improving the accuracy of a variety of challenging problems when compared to other machine learning methods. However, they require high performance computing systems (such as supercomputer clusters and GPU arrays) due to their highly complex and large computational architectures. Additionally, deep neural networks require machine learning experts to delicately design and fine-tune the large, complex architectures. This issue of complexity has increased greatly over time, driven by the demand for increasingly deeper and larger networks to boost cognitive accuracy. As such, it has become near impossible to take advantage of such powerful yet complex deep neural networks in scenarios where computational and energy resources are scarce, such as in embedded systems, as well as increasingly more difficult to hand-craft their architectures. Inspired by nature, the team at VIP lab have developed several pioneering strategies for enabling powerful yet operational deep intelligence by considering a radically different idea: Can deep neural networks evolve naturally over generations to become not only highly efficient but also powerful?
We have introduced the concept of evolutionary deep intelligence, where we evolve deep neural networks over multiple generations to become more efficient yet smart. The 'DNA' of each generation of deep neural networks is encoded computationally and used, along with simulated environmental factors such as those encouraging computational and energy efficiency through natural selection, to 'give birth' to its offspring deep neural networks, with the process repeating generation after generation. These 'evolved' offspring deep neural networks will naturally have more efficient, more varied architectures than their ancestor deep neural networks (due to natural selection and random mutations) while achieving powerful cognitive capabilities. Experimental results from a study using the MSRA-B and HKU-IS datasets demonstrated that the synthesized offspring deep neural networks can achieve state-of-the-art F-beta scores while having network architectures that are significantly more efficient, with a staggering ~48X fewer synapses by the fourth generation compared to the original, first-generation ancestor network. This level of performance was further reinforced by experimental results from a study using the MNIST dataset, which demonstrated synthesized offspring deep neural networks can achieve state-of-the-art accuracy (>99%) while having network architectures that are significantly more efficient, with a staggering ~40X fewer synapses by the seventh generation compared to the original, first-generation ancestor network. More remarkably, an accuracy of ~98% was still achieved by thirteen-generation offspring deep neural networks with an incredible ~125X fewer synapses compared to the original, first-generation ancestor network.