Brain Versus Machine

Abstract

Many biologists, especially those who study the biochemistry or cell biology of neural tissue are sceptical about claims to build a human brain on a computer. They know from first hand how complicated living tissue is and how much there is that we still do not know. Most importantly a biologist recognizes that a real brain acquires its functions and capabilities through a long period of development. During this time molecules, connections, and large scale features of anatomy are modified and refined according to the person’s environment. No present-day simulation approaches anything like the complexity of a real brain, or provides the opportunity for this to be reshaped over a long period of development. This is not to deny that machines can achieve wonders: they can perform almost any physical or mental task that we set them—faster and with greater accuracy than we can ourselves. However, in practice present day intelligent machines still fall behind biological brains in a variety of tasks, such as those requiring flexible interactions with the surrounding world and the performance of multiple tasks concurrently. No one yet has any idea how to introduce sentience or self-awareness into a machine. Overcoming these deficits may require novel forms of hardware that mimic more closely the cellular machinery found in the brain as well as developmental procedures that resemble the process of natural selection.

It is not impossible to build a human brain and we can do it in 10 years.

Henry Markram. TED Conference Oxford, 2009.

Randal Koene on Bray’s “Brain Versus Machine”

Emulation Versus Understanding

In his article, “Brain versus Machine”, Dennis Bray discusses and critiques claims that “reverse engineering of the brain is within reach”, and he compares this with developments in artificial intelligence. It is clear from his exposition that Bray does not agree with the optimistic outlook espoused by Terry Sejnowski at the 2010 Singularity Summit. As in all disagreements though, it is important to be careful about the subject matter that is the apparent cause of disagreement, the terminology used and how that terminology is understood. We will see that a more precise use of terms constrains the sweeping conclusions drawn by Bray.

What is meant by “reverse engineering” the brain? And does the validity of the reverse engineering claim affect the topics subsequently addressed in Bray’s article? The Merriam-Webster dictionary says: “Reverse engineer: to disassemble and examine or analyze in detail (as a product or device) to discover the concepts involved in manufacture usually in order to produce something similar.” According to Bray, we should be able to diagnose and cure ALS, schizophrenia, Huntington’s disease and Alzheimer’s if we are indeed shortly able to reverse engineer the brain. But the definition makes no claims about repairing existing systems. Even in the case of simple consumer electronics, repairing is much harder than replacing (which is what reverse engineering to “produce something similar” enables). For example, we do not tend to fix malfunctioning integrated circuits, even though we can design them. A fix would require additional sophisticated tools.

There is data acquisition about a system; there is the replication of a functioning system based on such data; and there is intervention in an existing system, with the aim to modify or correct its operation. As written, these three different accomplishments are likely listed in order of increasing difficulty. It is for this reason that I have in prior writings and presentations made an effort to compare side by side the probable degrees of difficulty in efforts to repair biology and thereby greatly extend life or to acquire relevant brain data in order to replicate function in a whole brain emulation. Given the number of possible points of break-down and the tendency of interventions to lead to unexpected downstream side-effects in vivo, the former may be much more difficult than the latter.

It is true that neuroscience has not given us a strong understanding of the various strategies at different levels that together make up the mind. Sejnowski’s optimism comes from the bottom up. He, like many others in the rapidly expanding fields of computational neuroscience and neuroinformatics, deals predominantly with functions carried out at the neuronal mechanistic level. That is where neuroscience has spent most of the last 100 years learning to identify elements and measure compounds and signals.

Bray’s assertion that experts in neurophysiology would view attempts to build brains on a computer as irrelevant is a bit of a strawman argument. Computational modeling is used widely in neurophysiology (see for example the Computational Neurophysiology lab at Boston University, at which I used to work). The scale of this work in neuroinformatics is increasing rapidly. Any representation, any model is an effort at system identification where you pick your level of detail to match the input and output you are interested in. For many representations of brain mechanisms we are interested in the timing of spikes of activity at individual neurons. Sensory input arrives as trains of spikes, spikes drive muscles (e.g., for speech), and inter-spike timing is crucial for memory formation at synapses.

Obviously, no one insists on representing a synapse with a line of code, as Bray posits. A much more likely approach is to represent each neuron by what is known as a compartmental model, where each compartment can be thought of as an electric circuit and implements the Hodgkin-Huxley equations for membrane channel dynamics. Through system identification, you translate underlying physics into functions. For example, the modulating effects of microRNA that Bray points to implement “mode switches” at many neurons in a diffuse manner. You can identify these modes by observing behaving neurons and a functional version of such a switch could be a 1-bit flag.

There are really two different questions:

• 1. What does it take to acquire data from a brain and replicate its unique function?

• 2. What does it take to understand the system strategy from the top down to the cells so that you could build similar systems based on that strategy?

Bray’s arguments focus on question 2. But we have already achieved excellent results aimed at question 1, such as those published by Briggman et al. and by Bock et al. (both in Nature, 2011). Each team demonstrated proofs of principle for system identification by reconstructing detailed individual neural circuitry, one in retina and one in visual cortex. Kozloski and Wagner (2011) showed how to take this to large-scale neural tissue simulation. Some researchers are indeed attempting to learn about emergent properties of neural networks by using large numbers of generic cell simulations. That is an endeavor separate from the building of tools with which to acquire sufficient structural and functional data from a specific piece of neural tissue to solve the system identification task within that constrained context. Examples of such tools are coming out of the labs of Winfried Denk (Max Planck Institute) and Jeff Lichtman (Harvard University). Although Bray refers to work by Izhikevich and Edelman in 2008, he omits references to some of Eugene Izhikevich’s more famous work, developing neuron representations now commonly known as Izhikevich Neurons. Those are a good example of system identification. They can produce the output of a wide range of different types of neurons without having to model any of the deeper neurophysiology.

Just as modern astronomy became possible by developing telescopes, what our goal requires is also the development of the right measurement instruments. If you capture the input–output perspective of each neuron with correct system identification then you capture everything that the system of neurons is responsible for, including a mind’s sense of dignity, sense of self-worth, respect, humor, and so forth.