Machine Intelligence
AI research in the better part of last few decades was focused on doing exceptionally well in a single task better than any human i.e shallow AI. However these networks were incapable of performing tasks that are different from the one they were trained originally. This is because the algorithms are not able to generalize the learning and apply the knowledge to a slightly different problem. Humans, on the other hand can generalize knowledge really well and can transfer learning across multiple problems. Hence let’s try to understand the key aspects human brain that could help us build a deep AI network along with the problems and challenges involved.
The three most important parts that makes a human brain an universal learning machine are:
Old brain
Old brain in humans takes care of basic survival functions such as breathing, moving, resting, and feeding and creates our experiences of emotion. Similarly, in a deep AI network, an old brain equivalent can be used for high priority functions such as recharge, basic movements and safeguarding personal and human safety. The old brain equivalent should have higher response times and should be programmed with high priority rules that could possible override other rules in the event of a rule clash. The purpose, aspirations and motivations of the deep AI should also be ingrained in the old brain equivalent.
Neocortex
Neocortex in humans forms the majority of human brain makes predictions using models continually about the world around us. It consists of around 150,000 cortical columns that makes predictions using a common algorithm. Different areas of neocortex that are similar in size and shape makes predictions using different sensory data(audio, video and sensory input). This is good news for AI researchers as making huge amounts of predictions using a common algorithm is commonplace in AI and machine learning. Similarly a neocortex equivalent in a deep AI system should be able to process information and solve problems using a common algorithm.
Embodiment
One of the key ways by which humans learn is through movement. We have developed motor capabilities that allows us to move our sensors on will. This makes it easy to learn quickly and effectively about the world around us. Hence any deep AI system requires an embodiment with varied sensory array that can be moved as and when required to enable continual learning about the environment around it. Sensors that can emulate vision, touch and hearing can be good choices for the deep AI system.
Miscellaneous parts
In addition to the above important parts, a deep AI machine requires grid and place cells similar to one’s found in humans for cognitive mapping and spatial memory. Sophisticated deep AI networks also need a consensus mechanism to decide the best prediction among a variety of predictions made by the network
Representation problem
One of the biggest challenges for the creating of a deep AI system is the representation problem which states how the sensory data that is sent to brain is structured and processed. Research is still going to determine the representation of sensory and other data within the brain. For example one such research proposes a solution where the brain creates a low-dimensional representational space of the sensory data that reflects the statistical regularities of object co-occurrence. A similar sophisticated solution is required in the deep AI network for processing the data coming from various sensory mechanisms.
Complete representation
So we have discussed how we can use our existing knowledge about the human brain to build a deep AI system. Based on the discussion, the system could have most or all of the following mechanisms built into it.
I hope you liked this blog where we discussed about one possible implementation of a deep AI system. If you are interested to know more, check out the links in references section. Thank you for reading and I’ll catch you guys in the next one.