The Neuroscience of Intelligence
Page 25
6.4 Bridging Human and Machine Intelligence Circuit by Circuit
The goal of artificial intelligence (AI) research is to create computer software and hardware that mimics human intelligence. There are many wildly successful applications of “smart” technology that continue to change everyday life throughout the world. There are computer programs that beat chess grandmasters, Jeopardy champions, and poker players. Engineers have developed most advances in AI with limited input from neuroscientists, mostly related to methods based on computational models of neural networks. However, an even more ambitious goal is to create intelligent machines with algorithms based on how neurons communicate in actual brain circuits explicated by basic neuroscience researchers. This is “real” intelligence.
A popular book by computer engineer and entrepreneur Jeff Hawkins makes a compelling case for building an intelligent machine using this neuroscience-based approach (Hawkins & Blakeslee, 2004). A key point is that computers and brains work on entirely different principles. For example, computers must be programmed and brains are self-learning. His core idea is that the cerebral cortex works fundamentally as a hierarchical system for storing and applying memory, especially memory of sequences, to make predictions about the world, and that this system is the essence of intelligence. One key insight is that the elements of this system are integrated by a single, all-purpose cortical learning algorithm (CLA). Therefore, Hawkins believes that the AI approach of designing separate elements of the system for machines is inherently limited. He believes that it is possible to design machines based on an all-purpose CLA, and that such machines might exceed human mental abilities. Here is how he states the challenge: “For half a century we’ve been bringing the full force of our species’ considerable cleverness to trying to program intelligence into computers. In the process we’ve come up with word processors, databases, video games, the Internet, mobile phones, and convincing computer-animated dinosaurs. But intelligent machines still aren’t anywhere in the picture. To succeed, we will need to crib heavily from nature’s engine of intelligence, the neocortex. We have to extract intelligence from within the brain. No other road will get us there” (p. 39). Hawkins has created the Redwood Neurosciences Institute and a company called Numenta to make brain-informed intelligent machines a reality. Numenta markets software based on algorithms that identify patterns, trends, and anomalies in large data sets. The approach is controversial, especially because it challenges the AI computational approaches widely used by companies like Facebook, Microsoft, and Google. It is too early to evaluate how the hierarchical CLA concept might relate to the hierarchical g-factor, but it clearly fits the theme of this chapter on neuroscience approaches to intelligence and where they might lead (see an informative online interview with Hawkins from March 20, 2014 with Jack Clark of The Register, www.theregister.co.uk/2014/03/29/hawkins_ai_feature).
The concept of building machines based on the way the brain works also is informing the design of microchips. A number of research groups are working on building microchips to perform brain functions, especially related to perception, based on actual neural circuitry data. The general effort is known as neuromorphic chip technology. Some of these chips are designed to have direct interface with the brain. Already chips are available to enhance hearing and vision. These efforts may one day expand to cognitive processes, but so far I am unaware of any neuromorphic successes related to specific mental abilities let alone general intelligence. This too is an area ripe for fertile imagination.
In Chapters 2 and 4 we introduced some multicenter consortia that are pooling genetic data to create very large samples for statistical analyses that maximize the discovery of small effects related to intelligence that are hard to detect in smaller samples typical of individual studies. There are also other large collaborative research programs that share data from many sources with the goal of mapping the structure and function of the human brain and how it develops. Current technology can produce maps at the neural circuit level. These maps can inform studies of aging, brain disorders, and brain diseases. They may also inform questions about learning, memory, and other cognitive processes. Such studies would be a prelude to addressing how individual differences in mental abilities, including intelligence factors, arise from differences among brains.
A bold goal was announced in 2005 by a group of scientists working in Switzerland. They undertook to create an artificial brain by building biologically realistic models of neurons and networks. Working with an IBM Big Blue supercomputer, they simulated brain activity starting with about 10,000 virtual neurons. This ambitious “Blue Brain” project expanded dramatically in 2009 when the European Union provided additional funding of $1.3 billion and many additional collaborators joined the endeavor, renamed the Human Brain Project. The stated goal is to simulate a human brain, all 80–100 billion neurons with 100 trillion connections. No neuroscience project has ever received this level of support. There is no shortage of controversy about every aspect of this project, but the most important issue for us is that cognitive neuroscience was excluded from the project. This likely will be reversed given the outcry from the cognitive neuroscience community (Fregnac & Laurent, 2014). Even with their return, however, intelligence research is not on the agenda. Hopefully, at some point someone with access to a simulated brain will wonder about just how smart the virtual brain may be.
In the USA, there are more modest initiatives with similar goals of building simulated brains. DARPA (The Defense Advanced Research Projects Agency) funded the SyNAPSE Program (Systems of Neuromorphic Adaptive Plastic Scalable Electronics) in 2008 through 2016 (sounds like someone badly wanted to call this SYNAPSE and worked backwards with a committee). The ultimate aim is to build a microprocessor system that emulates a mammalian brain. In 2013, the White House announced the BRAIN Initiative (Brain Research through Advancing Innovative Neurotechnologies), which provided funding for projects that will lead to detailed functional and structural maps of the brain. This initiative builds on other collaborations already funded like the Human Connectome Project, which is one of the few to include cognitive tests from which a g-factor can be derived. An intriguing first report from the Human Connectome Project related to intelligence is based on resting state fMRIs from 461 participants. Functional connectivity computations among 200 brain areas incorporated 158 demographic and psychometric variables in a single analysis. A g-factor was not derived, but the main result showed that intelligence variables were among the strongest related to overall connectivity among brain areas, such that greater connectivity was associated with higher test scores (Smith et al., 2015). This is a short report, but their sophisticated analyses suggest intelligence is related to resting-state connectivity in the default network and in PFIT areas.
Just as I was finishing the final draft of this book, another study was published from the Human Connectome Project that expanded the fMRI study just described (Finn et al., 2015). This is a breathtaking study not only for the findings, but also because it fulfills a dream I have had for 40 years about using brain profiles to describe individuals and their mental abilities. Here’s what they reported based on analyses of connectivity patterns among brain areas (like those shown in Figure 4.1). They started with fMRI data from 126 people collected during six sessions, including four task and two resting conditions. The typical analysis would have compared the average connectivity for the entire group among the task and rest conditions. These researchers, however, focused on individual differences. The simple question was whether connectivity patterns were stable within a person. To address this question, functional connectivity patterns among 268 brain nodes (making up 10 networks) were calculated for each person separately for each session. Not only was the connectivity pattern stable within a person when the two resting conditions were compared, it was also stable across the four different tasks. In addition, each person’s pattern was unique enough that it could be used to identify the person. Because these remarkable results combined stabilit
y and uniqueness, the connectivity pattern was characterized as a brain fingerprint. Of particular interest to us, individual brain fingerprints predicted individual differences in fluid intelligence. It gets even better. The strongest correlations with fluid intelligence were in frontoparietal networks. And, best of all, cross-validation was included in the report. The authors note, “These results underscore the potential to discover fMRI-based connectivity ‘neuromarkers’ of present or future behavior that may eventually be used to personalize educational and clinical practices and improve outcomes.” They conclude, “Together, these findings suggest that analysis of individual fMRI data is possible and indeed desirable. Given this foundation, human neuroimaging studies have an opportunity to move beyond population-level inferences, in which general networks are derived from the whole sample, to inferences about single subjects, examining how individuals’ networks are functionally organized in unique ways and relating this functional organization to behavioral phenotypes in both health and disease.” I believe this is a landmark study. I wish I had done it. It is the perfect end to this section and the perfect beginning for a new phase of neuroimaging research on intelligence, if, of course, there is independent replication. Whew.
All these major funding initiatives speak well for the future of neuroscience research. They have accelerated the enthusiasm generated more than 25 years ago when President George H. Bush declared the 1990s would be the Decade of the Brain. Unfortunately, at that time, intelligence was not mentioned among the targets for research (Haier, 1990). There is an understandable general justification for basic research that may have practical implications for understanding brain diseases and disorders. At some point, these newest, twenty-first-century efforts to map and simulate brains may also recognize that attention to intelligence is equally worthy. We now apparently have brain fingerprints that predict intelligence. Once a realistic virtual human brain exists, can creating real intelligence be far behind?
6.5 Consciousness and Creativity
This is a good place to comment briefly on consciousness and creativity. Like intelligence, both are among the highest-order functions of the human brain. If intelligence can be simulated, why not simulate creativity or consciousness? The idea that consciousness has a neuroscientific basis has become mainstream, in large part based on the popularity of Francis Crick’s book, The Astonishing Hypothesis (Crick, 1994); Crick shared the Nobel Prize for discovering the molecular structure of DNA. Some of the research efforts to understand the neural basis of consciousness include neuroimaging studies of humans in varying degrees of consciousness induced by different anesthetic drugs. My friend and colleague Michael Alkire, an anesthesiologist, and I published some of the earliest PET imaging studies that investigated this (Alkire & Haier, 2001; Alkire et al., 1995, 1999, 2000). We were trying to establish which brain circuits were the last to deactivate as the participant became unconscious. From these studies, we hoped to infer the mechanisms of action for different anesthetic drugs and pinpoint the brain mechanisms responsible for consciousness. No luck so far, but this ambitious goal remains the greatest unsolved mystery of neuroscience.
I raise the topic here briefly because during our early PET experiments, we wondered if there might be a link between consciousness and intelligence. We tend to regard everyone who is awake as conscious, but are there degrees of “awakeness”? Are some people more conscious (aware) than others and could such differences be related to intelligence? We have no clear way to assess individual differences in consciousness in individuals who are awake. One hypothesis that could be tested is whether high-IQ people need more (or less) anesthetic drug to render them unconscious for surgery, assuming there is a valid measure of depth of anesthesia. We have not pursued this question, but it seems reasonable to suspect that the two highest-order activities of the human brain may have circuits in common. The mechanisms of action of anesthetic drugs remain unclear, but if there are common circuits between consciousness and intelligence, we might speculate that new drugs that work in opposite ways than anesthetic drugs may produce hyper-consciousness or hyper-awareness, possible aspects of higher intelligence.
Similarly, I want to discuss briefly neuroscience studies of creativity as they relate to intelligence. My friend and colleague Rex Jung is a neuropsychologist who specializes in neuroimaging studies of creativity. We have pursued whether intelligence and creativity may have common neurocircuits. There is some overlap between creativity and intelligence (Haier & Jung, 2008; Jung, 2014). Creativity and the creative process are even more difficult to define and assess for empirical research than intelligence and reasoning. The same general approach, however, is applicable. A battery of tests that assesses different aspects of creativity is given and a creativity index is derived either by summing the scores of individual tests (like IQ scores) or by extracting a latent creativity variable like the g-factor. Aspects of creativity include, for example, measures of originality, fluency of ideas, and divergent thinking. However, whether there is a g-like factor of general creativity that transcends different specialty fields is still an open question. Creative artistic ability in dance, or painting, or music might have quite different neural elements and they might not overlap at all with neural aspects of creativity in fields of science or literature or architecture. There is also the question of genius, a concept equally challenging to define for research. Can a creative genius have lower than average IQ, as it seems in some cases of savants? Can an intellectual genius have no creativity? Does the rare “true” genius require both high intelligence and high creativity? There are not yet clear empirical answers, but neuroscience approaches may help resolve these basic issues. Creativity research is a large field, so we will limit our focus here to illustrative neuroimaging studies.
I am not aware of any verified cases where brain damage or illness resulted in increased intellectual ability. However, there are apparently rare cases where people have demonstrated a dramatic new creative ability, often artistic, after they develop frontotemporal dementia (FTD), a degenerating illness similar to Alzheimer’s disease. This observation is not typical for FTD patients (Miller et al., 1998, 2000; Rankin et al., 2007). This is intriguing because it raises the possibility that creativity might be unleashed in more people if only certain brain conditions changed, although dementia is hardly a positive change. The general idea, however, is that dis-inhibition (i.e., deactivation) of neural circuits and networks caused by the disease process is a key factor, because dis-inhibition allows more associations among brain areas that do not routinely communicate. There are many ways to dis-inhibit the brain in general, like drinking alcohol or developing FTD, but dis-inhibition targeted to particular neural networks related to creativity may be possible without affecting other networks necessary for balance, coordination, memory, and judgment. Are there creativity networks in the brain?
Functional neuroimaging studies have tried to capture brain activity during the creative process. There are now many studies, for example, that have imaged people with fMRI while they performed musical improvisation as an expression of creativity (Bengtsson et al., 2007; Berkowitz & Ansari, 2010; Donnay et al., 2014; Limb & Braun, 2008; Liu et al., 2012; Pinho et al., 2014; Villarreal et al., 2013). Music improvisation is a manageable paradigm in experimental studies, whereas imaging studies of the creative process in dance, architecture, and other domains is not as practical. One early study, for example, scanned six male professional jazz pianists with fMRI while they performed two tasks that required either improvisation or over-learned musical sequences (Limb & Braun, 2008). The results of this small study suggested that compared to the over-learned sequence, improvisation was associated with a combination of bilateral deactivation in some areas, especially parts of the prefrontal cortex (including BAs 8, 9, and 46), along with bilateral activation in other areas distributed across the brain including in the frontal lobe (BA 10). These findings are shown in Figure 6.2. A similar fMRI study of 12 male freestyle rap musicians compared spont
aneous creation of rap lyrics to previously memorized sequences, both conditions using the same musical background (Liu et al., 2012). The results also suggested a pattern of deactivations and activations, as shown in Figure 6.3, which is mostly consistent with the results of the Limb and Braun study of jazz pianists. Another study of 39 pianists with varying degrees of improvisation experience reported an association between length of experience and connectivity among brain areas that suggested more efficient information flow during creative expression for the more experienced musicians (Pinho et al., 2014). Less activity in frontal and parietal areas was associated with more improvisation experience as shown in Figure 6.4. A meta-analysis of musical improvisation studies like these tried to integrate findings and explain inconsistencies among studies (Beaty, 2015), but this research is still at an early stage and little consistency is apparent. Whether it will provide a model for creativity in general remains to be seen.
Figure 6.2 Brain activations (red/yellow) and deactivations (blue/green) in jazz pianists during improvisation.
Adapted from Limb & Braun (2008) (Open Access).
Figure 6.3 fMRI comparison of rappers during improvised and conventional conditions. Yellow represents significant increases in fMRI blood flow during improvisation; blue represents significant decreases. Top row shows cortical surface; bottom row shows medial (inside) surface.