The human brain provides a working example of human-level intelligence. Therefore attempts to reverse engineer the detailed workings of the human brain carry some risk of enabling the development of human-level or smarter-than-human AI. Many brain-related research areas could have some undesirable consequences, but the risks vary widely.
In Europe the Human Brain Project is a large scale project focused on better understanding the human brain, with the explicit aims of better treating diseases and developing new computing technologies.
The founding Blue Brain sub-project aims to simulate the human brain at an unprecedented level of detail, and to date has used the multi-compartment electrochemical simulation NEURON software. Multi-compartment electrochemical simulation is slow/expensive and therefore the short term risks of this work seem low. (Amazon EC2 2017 on demand price: $1.00 per simulated hour per neuron).
In the United States the BRAIN Initiative is also a large scale project focused on better understanding the human brain, with the primary goal of better treating diseases. Most of the sub-projects are focused on developing and using tools to better understand the operation of the brain. An exception is:
The $100m IARPA MICrONS project which "seeks to revolutionize machine learning by reverse-engineering the algorithms of the brain".
Another IARPA project is Knowledge Representation in Neural Systems (KRNS) which seeks to understand how the human brain encodes knowledge.
Separate from the BRAIN Initiative is the $76m (through 2011) DARPA SyNAPSE project which is providing money to IBM's TrueNorth project and HRL's Brain-Machine Intelligence lab to develop spiking neuron inspired hardware.
The above research could lay the groundwork for brain inspired AI. There are a number of steps required before human-level AI might be achieved:
- Develop a hardware platform (SyNAPSE)
Understand the wiring pattern of a single particular cortical column (MICrONS)
- Mapping the wiring pattern of a particular cortical column to a set of general wiring principles
- Developing wiring principles for other brain regions
- Possibly understand how the brain encodes information (KRNS)
- Understanding how synapses are dynamic connections that vary over time
- Gain a better understanding of higher level phenomena like motivation, learning, and short-term memory
It seems unlikely human-level AI will be achieved by scanning a complete brain. The cost of that appears prohibitive (see cost estimate in Uploading). Instead it might be achieved by scanning small sections of brain tissue to deduce general wiring principles, and building a computer model based on that.
It may or may not be necessary to understand how the brain encodes information in order to create brain inspired AI. However projects like KRNS that attempt to figure this out are extremely valuable in that understanding how information is encoded by the brain may make it possible to "probe" or "audit" brain inspired AI systems for safety. For biological reasons it is virtually impossible to tell what humans know or are thinking, but with a neuromorphic system in which every synaptic weight is known, and various scenarios can be replayed over and over, it might be much easier.
If programs like MICrONS and TrueNorth might be successful in advancing AI towards human brain like performance it is imperative that they be matched dollar for dollar in funding for AI safety. AI safety is not something that can be added later on.
It might be argued that until we are certain we can develop advanced AI safely that we focus on areas of brain research that have less risk of undesirable consequences than some of the above projects. Some possible examples:
- Glial cells are as numerous as neurons in the brain, indicating that they play an important role in brain functioning, yet they receive little attention.
- Drug companies are loath to do head-to-head trials against other drugs, preferring to benchmark against a placebo. Either a mandate requiring this, or publicly funded clinical trials of psychiatric drugs would allow the better treatment of psychiatric disorders without requiring any additional fundamental research.
- Thinking freely and broadly, sleep is a brain phenomena that exerts a burden far greater than mental illness, yet little brain research funding goes into studying sleep, discovering if it is necessary, or ways to reduce the need for sleep.
- Generally a focus on the cellular and molecular levels and less on the systems of neurons level might prove beneficial.