A statistical mechanics of neuronal interactions (SMNI) is explored as providing some substance to a physiological basis for the g factor. Some specific elements of SMNI, previously used to develop a theory of short-term memory (STM) and a model of electroencephalography (EEG), are key to providing this basis. Specifically, Hick's Law, an observed linear relationship between reaction time (RT) and the information storage of STM, in turn correlated to a RT-g relationship, is derived.
1. My specific interest in reviewing "The g Factor" by Arthur Jensen (1998, 1999) is to see whether some anatomical and/or physiological processes at the columnar level of neuronal interactions can account for the g factor.
2. From circa 1978 to the present, a series of papers on the statistical mechanics of neocortical interactions (SMNI) has been developed to model columns and regions of neocortex, spanning mm to cm of tissue. Most of these papers have dealt explicitly with calculating properties of short-term memory (STM) and scalp EEG in order to test the basic formulation of this approach. SMNI derives aggregate behaviour of experimentally observed columns of neurons from statistical electrical-chemical properties of synaptic interactions. While not useful to yield insights at the single neuron level, SMNI has demonstrated its pability to describe large-scale properties of short-term memory and electroencephalographic (EEG) systematics [FOOTNOTE 1].
3. One must assume that Jensen faced very difficult problems in choosing just how much technical detail to give in his broad text, e.g., discussing the extent of the expert statistical analyses that have been brought to bear upon the g factor. However, I do see reason to criticise some general features of the simple statistical algorithms presented, especially those that overlap with my own mathematical and physics expertise.
The simple approach to factor analysis initiated on page 23,
X = t + e {1}
ftp://www.cogsci.soton.ac.uk/pub/psycoloquy/1999.volume.10/Pictures/ing1.html
where e is the residual "noise" of fitting the variable X to the independent variable t, has some serious flaws not addressed by additional material presented thereafter. For example, in this context, I find the arguments in the long footnote 16 on pages 101-103 unconvincing, but I agree with its conclusion:
"But the question is mainly of scientific interest, and a really satisfactory answer ... will become possible only as part and parcel of a comprehensive theory of the nature of g. ... The distribution of obtained measurements should conform to the characteristics of the distribution dictated by theoretical considerations."
I think it clear that any such "theoretical considerations" must themselves be well tested against experimental evidence at each spatial-temporal scale purported to be modelled.
4. It must be understood that a quite explicit model of the real world is being assumed here -- that of a simple normal Gaussian process. The real issue in many physical/biological systems is that most often the real multivariable world is much more aptly described by something like
X = tX(X,Y) + sX(X,Y) eX {2}
ftp://www.cogsci.soton.ac.uk/pub/psycoloquy/1999.volume.10/Pictures/ing2.html
Y = tY(X,Y) + sY(X,Y) eY {3}
ftp://www.cogsci.soton.ac.uk/pub/psycoloquy/1999.volume.10/Pictures/ing3.html
When the t's and s's are constants, then simple statistics can determine their values and cross-correlations between the s's.
5. Simple statistical methods can be useful even if the t's are relatively simple quasi-linear parametrised functions. Such simple methods fail quite miserably if the ts are highly nonlinear functions, especially if care is not taken to use sophisticated optimisation algorithms. The most terrible flaws often occur because, for the sake of making life easier for the analyst, any model faithful to the real system is butchered and sacrificed, and the s's are taken to be constants. There can be a lot of "signal" in the (generally nonlinear) functionality of the "noise" terms, and no amount of fancy quasi-linear statistical analysis can substitute for a proper theory/model of the real system.
6. In general, the proper treatment of the problem is quite difficult, which is of course no excuse for poor treatment. The solution in many disciplines is to go a level or two deeper in some reductionist sense, to develop plausible models at the top scale being analysed. Indeed, this was the call I saw and responded to in the advertisement for reviewers of Jensen's (1999) work:
"Commentary is invited from psychometricians, statisticians, geneticists, neuropsychologists, psychophysiologists, cognitive modellers, evolutionary psychologists and other specialties concerned with cognitive abilities, their measurement, and their cognitive and neurobiological basis."
In this context, the successes of SMNI and its agreement with general STM observations are due to processing stochastic nonlinearities of the forms described above. Attempts to avoid dealing with these nonlinearities, derived from lower-level synaptic and neuronal activity, have not been as successful as SMNI in detailing STM (Ingber, 1995b).
6. Since the early 1980s, a series of papers on the statistical mechanics of neocortical interactions (SMNI) has been developed to model columns and regions of neocortex, spanning mm to cm of tissue. Most of these papers have dealt explicitly with calculating properties of short-term memory (STM) and scalp EEG in order to test the basic formulation of this approach [FOOTNOTE 2]. This model was the first physical application of a nonlinear multivariate calculus developed by other mathematical physicists in the late 1970s (Graham, 1977; Langouche et al, 1982).
7. Studies of SMNI have detailed a physics of short-term memory and (short-fibre contribution to) EEG phenomena (Ingber, 1984), in terms of MG firings, where G represents E or I, ME represents contributions to columnar firing from excitatory neurons, and MI represents contributions to columnar firing from inhibitory neurons. About 100 neurons comprise a minicolumn (twice that number in the visual cortex); while a macrcolumn is comprised of around 1000 minicolumns. A mesocolumn is developed by SMNI to reflect the convergence of short-ranged (as well as long-ranged) interactions of macrocolumnar input on minicolumnar structures, in terms of synaptic interactions taking place among neurons (about 10,000 synapses per neuron). The SMNI papers give more details on this derivation.
8. In this SMNI development, a Lagrangian is explicitly defined from a derived probability distribution of mesocolumnar firings in terms of the MG and electric potential variables, PHIG. Several examples have been given to illustrate how the SMNI approach is complementary to other models. For example, a mechanical string model was first discussed as a simple analogue of neocortical dynamics to illustrate the general ideas of top down and bottom up interactions (Nunez, 1989; Nunez & Srinivasan, 1993). SMNI was applied is to this simple mechanical system, illustrating how macroscopic variables are derived from small-scale variables (Ingber & Nunez, 1990). [See also Nunez 2000.]
9. To the extent that SMNI offers stochastic bounds for STM phenomena during focused selective attention, SMNI provides a model of STM. This 7+-2 "rule" is well verified by SMNI for acoustical STM (Ingber, 1984; Ingber, 1985b; Ingber, 1994), transpiring on the order of tenths of a second to seconds, limited to the retention of 7+-2 items (Miller, 1956). The 4+-2 "rule" is also well verified by SMNI for visual or semantic STM, which typically require longer times for rehearsal in a hypothesised articulatory loop of individual items, with a capacity that appears to be limited to 4+-2 (Zhang & Simon, 1985). SMNI has detailed these constraints in models of auditory and visual cortex (Ingber, 1984; Ingber, 1985b; Ingber, 1994; Ingber & Nunez, 1995).
10. Another interesting phenomenon of STM capacity explained by SMNI is the primacy versus recency effect in STM serial processing (Ingber, 1985b), wherein first-learned items are recalled most error-free, with last-learned items still more error-free than those in the middle (Murdock, 1983). The basic assumption is that a pattern of neuronal firing which persists for many tau cycles, tau on the order of 10 msec, is a candidate for the storage of the "memory" of activity that gave rise to this pattern. If several firing patterns can simultaneously exist, then there is the capability that several memories can be stored. The short-time probability distribution derived for the neocortex is the primary tool in the search for such firing patterns. The highest peaks of this probability distribution can be more readily accessed than the others, and they sustain their patterns against fluctuations more accurately. The more recent memories or newer patterns may be presumed to be those having synaptic parameters which have been more recently tuned and/or more actively rehearsed.
11. It has been noted that experimental data on velocities of propagation of long-ranged fibres (Nunez, 1981; Nunez, 1995) and derived velocities of propagation of information across local minicolumnar interactions (Ingber, 1982) yield comparable time scales of interactions across minicolumns of tenths of a second. Therefore, such phenomena as STM likely are inextricably dependent on interactions at local and global scales.
12. Previous SMNI studies have detailed that maximal numbers of attractors lie within the physical firing space of MG. This is consistent with experimentally observed capacities of auditory and visual short- term memory (STM), when a "centering" mechanism is enforced by shifting background noise in synaptic interactions, itself consistent with experimental observations under conditions of selective attention (Mountcastle et al, 1981; Ingber, 1984; Ingber, 1985b; Ingber, 1994; Ingber & Nunez, 1995). This leads to all attractors of the short-time distribution lying along a diagonal line in MG space, effectively defining a narrow parabolic trough containing these most likely firing states. This essentially collapses the 2 dimensional MG space down to a 1 dimensional space of most importance. Thus, the predominant physics of short-term memory and of (short-fiber contribution to) EEG phenomena takes place in a narrow "parabolic trough" in MG space, roughly along a diagonal line (Ingber, 1984).
13. Using the power of this formal structure, sets of EEG and evoked potential data, collected to investigate genetic predispositions to alcoholism, were fitted to an SMNI model to extract brain "signatures" of short-term memory (Ingber, 1997; Ingber, 1998). These results give quantitative support for an intuitive picture portraying neocortical interactions as having common algebraic or physics mechanisms across quite disparate spatial scales and functional or behavioural phenomena, i.e., describing interactions among neurons, columns of neurons, and regional masses of neurons.
14. For future work, I have described how bottom-up neocortical models can be developed into eigenfunction expansions of probability distributions appropriate to describe short-term memory in the context of scalp EEG (Ingber, 2000). The mathematics of eigenfunctions are similar to the top-down eigenfunctions developed by some EEG analysts, although they have different physical manifestations. The bottom-up eigenfunctions are at the local mesocolumnar scale, whereas the top- down eigenfunctions are at the global regional scale. However, these approaches have regions of substantial overlap (Ingber & Nunez, 1990; Ingber, 1995a), and future studies may expand top-down eigenfunctions into the bottom-up eigenfunctions, yielding a model of scalp EEG that is ultimately expressed in terms of columnar states of neocortical processing of attention and short-term memory.
15. An optimistic outcome of future work might be that these EEG eigenfunctions, baselined to specific STM processes of individuals, could be a more direct correlate to estimates of g factors.
16. In Chapter 6, "Biological Correlates of g," Jensen puts these issues in perspective: "First, psychometric tests were never intended or devised to measure anything other than behavioural variables... at this point most explanations are still conjectural." Most of the chapter, however, falls back on similar too-simple statistical models of correlations between measured variables and behavioural characteristics.
17. The sections on "Biological Correlates of g" dealing with EEG recordings are incomplete, and appear superficial in comparison with most of the other parts of the book. The introduction of "complexity" as a possible correlate to IQ is based on faddish studies that do not have any theoretical or experimental support (Nunez, 1995). If such support does emerge, it will most likely be developed on the basis of more complete stochastic models and much better EEG recordings.
18. The outline of test categories giving rise to high versus low g loadings on page 35 provides me with some immediate candidates for further investigations of a physiological basis for the g factor. I think a good working hypothesis is that these two categories are marked by having the high g loadings more correlated than the low g loadings to statistical interactions among the peaks of the probability distributions described above in the context of SMNI's theory of STM. The high g categories clearly require relatively more processing of several items of information than do the low g categories.
19. This seems to be a correlation similar to that drawn by Spearman (1904, 1927), as described by Jensen, in having "education of relations" and "education of correlates" more highly correlation with high g categories than the "apprehension of experience".
20. From the SMNI perspective, control of selective attention generally is highly correlated with high utilisation of STM, e.g., to tune the "centering mechanism". This seems similar to Spearman's correlation of "mental energy" with high g categories.
21. There are several mechanisms that could be used to distinguish how well individuals might perform on high g category tasks. The ability to control the "centering mechanism" is required to sustain a high degree of statistical processes of multiple most probable states of information. The particular balance of general chemical-electrical activity directly shapes the distribution of most probable states. For some tasks, processing across relatively more of these most probable states might be required; for other tasks processing among larger focused peaks of most probable states may be more important.
22. The SMNI approach to STM gives a reasonable foundation to discuss RT and items in STM storage. These previous calculations support the intuitive description of items in STM storage as peaks in the 10-millisecond short-time (Ingber, 1984; Ingber, 1985b) as well as the several-second long-time (Ingber, 1994; Ingber & Nunez, 1995) conditional probability distribution of correlated firings of columns of neurons. These columnar firing states of STM tasks were also correlated to EEG observations of evoked potential activities (Ingber, 1997; Ingber, 1998). This distribution is explicitly calculated by respecting the nonlinear synaptic interactions among all possible combinatoric aggregates of columnar firing states (Ingber, 1982; Ingber, 1983).
23. The RT necessary to "visit" the states under control during the span of STM can be calculated as the mean time of "first passage" between multiple states of this distribution, in terms of the probability P as an outer Integral dt (sum) over refraction times of synaptic interactions during STM time t, and an inner Integral Integral dM (sum) taken over the mesocolumnar firing states M (Risken, 1989) which has been explicitly calculated to be within observed STM time scales (Ingber, 1984),
RT = - Integral dt t Integral dM dP/dt {4}
ftp://www.cogsci.soton.ac.uk/pub/psycoloquy/1999.volume.10/Pictures/ing4.html
24. As demonstrated by previous SMNI STM calculations, within tenths of a second, the conditional probability of visiting one state from another P, can be well approximated by a short-time probability distribution expressed in terms of the previously mentioned Lagrangian L as
P = 1/sqrt(2 pi dt g) exp(-Ldt) {5}
ftp://www.cogsci.soton.ac.uk/pub/psycoloquy/1999.volume.10/Pictures/ing5.html
where g is the determinant of the covariance matrix of the distribution P in the space of columnar firings. This expression for RT can be approximately rewritten as
RT = K Integral dt Integral dM P lnP {6}
ftp://www.cogsci.soton.ac.uk/pub/psycoloquy/1999.volume.10/Pictures/ing6.html
where K is a constant when the Lagrangian is approximately constant over the time scales observed. Since the peaks of the most likely M states of P are to a very good approximation well-separated Gaussian peaks (Ingber, 1984), these states can be treated as independent entities under the Integral. This last expression is essentially the "information" content weighted by the time during which processing of information is observed.
25. The calculation of the heights of peaks corresponding to most likely states includes the combinatoric factors of their possible columnar manifestations, as well as the dynamics of synaptic and columnar interactions. In the approximation that we only consider the combinatorics of items of STM as contributing to most likely states measured by P, i.e., that P measures the frequency of occurrences of all possible combinations of these items, we obtain Hick's Law, the observed linear relationship of RT versus STM information storage, first discussed by Jensen in Chapter 8. For example, when the bits of information are measured by the probability P being the frequency of accessing a given number of items in STM, the bits of information in 2, 4 and 8 states are given as approximately multiples of ln2 of items, i.e., ln2, 2ln2 and 3ln2 respectively. (The limit of taking the logarithm of all combinations of independent items yields a constant times the sum over pi lnpi, where pi is the frequency of occurrence of item i.)
26. Jensen's book provides the motivation to explore a more fundamental basis of the g factor. I have examined this work in the narrow focus of some specific elements of SMNI, previously used to develop a theory of STM and a model of EEG. I have focused on how bottom-up SMNI models can be developed into eigenfunction expansions of probability distributions appropriate to describe STM. This permits RT to be calculated as an expectation value over the STM probability distribution of stored states, and in the good approximation of such states being represented by well separated Gaussian peaks, this yield the observed linear relationship of RT versus STM information storage. This SMNI STM approach also suggests several other studies that can be performed in the context of examining an underlying basis for the g factor.
[1] See Ingber, 1982; Ingber, 1983; Ingber, 1984; Ingber, 1991; Ingber, 1994; Ingber, 1995a; Ingber & Nunez, 1995; Ingber, 1996a; Ingber, 1997.
[2] See Ingber, 1981; Ingber, 1982; Ingber,1983; Ingber, 1984; Ingber, 1985a; Ingber, 1985b; Ingber, 1986; Ingber & Nunez, 1990; Ingber, 1991; Ingber, 1992; Ingber, 1994; Ingber & Nunez, 1995; Ingber, 1995a; Ingber, 1995b; Ingber, 1996b; Ingber, 1997; Ingber, 1998.
Graham, R. (1977) Covariant formulation of non-equilibrium statistical thermodynamics. Z. Physik B26:397-405.
Ingber, L. (1981) Towards a unified brain theory. J. Social Biol. Struct. 4:211-224.
Ingber, L. (1982) Statistical mechanics of neocortical interactions.
Ingber, L. (1983) Statistical mechanics of neocortical interactions. Dynamics of synaptic modification. Phys. Rev. A 28:395-416. http://www.ingber.com/smni83dynamics.ps.gz
Ingber, L. (1984) Statistical mechanics of neocortical interactions. Derivation of short-term-memory capacity. Phys. Rev. A 29:3346-3358. http://www.ingber.com/smni84stm.ps.gz
Ingber, L. (1985a) Statistical mechanics of neocortical interactions. EEG dispersion relations. IEEE Trans. Biomed. Eng. 32:91-94. http://www.ingber.com/smni85eeg.ps.gz
Ingber, L. (1985b) Statistical mechanics of neocortical interactions: Stability and duration of the 7+-2 rule of short-term-memory capacity. Phys. Rev. A 31:1183-1186. http://www.ingber.com/smni85stm.ps.gz
Ingber, L. (1986) Statistical mechanics of neocortical interactions. Bull. Am. Phys. Soc. 31:868.
Ingber, L. (1991) Statistical mechanics of neocortical interactions: A scaling paradigm applied to electroencephalography. Phys. Rev. A 44:4017-4060. http://www.ingber.com/smni91eeg.ps.gz
Ingber, L. (1992) Generic mesoscopic neural networks based on statistical mechanics of neocortical interactions. Phys. Rev. A 45:R2183-R2186. http://www.ingber.com/smni92mnn.ps.gz
Ingber, L. (1994) Statistical mechanics of neocortical interactions: Path-integral evolution of short-term memory. Phys. Rev. E 49:4652-4664. http://www.ingber.com/smni94stm.ps.gz
Ingber, L. (1995a) Statistical mechanics of multiple scales of neocortical interactions, In: Neocortical Dynamics and Human EEG Rhythms, ed. P.L. Nunez. Oxford University Press, 628-681. http://www.ingber.com/smni95scales.ps.gz
Ingber, L. (1995b) Statistical mechanics of neocortical interactions: Constraints on 40 Hz models of short-term memory. Phys. Rev. E 52:4561-4563. http://www.ingber.com/smni95stm40hz.ps.gz
Ingber, L. (1996a) Nonlinear nonequilibrium nonquantum nonchaotic statistical mechanics of neocortical interactions. Behavioral and Brain Sci. 19:300-301. [Invited commentary on Dynamics of the brain at global and microscopic scales: Neural networks and the EEG, by J.J. Wright and D.T.J. Liley.] http://www.ingber.com/smni96nonlinear.ps.gz
Ingber, L. (1996b) Statistical mechanics of neocortical interactions: Multiple scales of EEG, In: Frontier Science in EEG: Continuous Waveform Analysis (Electroencephal. clin. Neurophysiol. Suppl. 45), ed. R.M. Dasheiff & D.J. Vincent. Elsevier, 79-112. [Invited talk to Frontier Science in EEG Symposium, New Orleans, 9 Oct 1993.] http://www.ingber.com/smni96eeg.ps.gz
Ingber, L. (1997) Statistical mechanics of neocortical interactions: Applications of canonical momenta indicators to electroencephalography. Phys. Rev. E 55:4578-4593. http://www.ingber.com/smni97cmi.ps.gz
Ingber, L. (1998) Statistical mechanics of neocortical interactions: Training and testing canonical momenta indicators of EEG. Mathematical Computer Modelling 27:33-64. http://www.ingber.com/smni98cmitest.ps.gz
Ingber, L. (2000) Statistical mechanics of neocortical interactions: EEG eigenfunctions of short-term memory. Behavioural and Brain Sciences (to be published). [Invited commentary on Toward a Quantitative Description of Large-Scale Neo-Cortical Dynamic Function and EEG, by P.L. Nunez.] http://www.ingber.com/smni00eegstm.ps.gz
Ingber, L. & Nunez, P.L. (1990) Multiple scales of statistical physics of neocortex: Application to electroencephalography. Mathematica. Computer Modelling 13:83-95.
Ingber, L. & Nunez, P.L. (1995) Statistical mechanics of neocortical interactions: High resolution path-integral calculation of short- term memory. Phys. Rev. E 51:5074-5083. http://www.ingber.com/smni95stm.ps.gz
Jensen, A. (1998). The g Factor: The Science of Mental Ability. Praeger
Jensen, A. (1999). Precis of: "The g Factor: The Science of Mental Ability" PSYCOLOQUY 10 (023). ftp://ftp.princeton.edu/pub/harnad/Psycoloquy/1999.volume.10/ psyc.99.10.023.intelligence-g-factor.1.jensen http://www.cogsci.soton.ac.uk/cgi/psyc/newpsy?10.023
Langouche, F., Roekaerts, D. & Tirapegui, E. (1982) Functional Integration and Semiclassical Expansions. Reidel, Dordrecht, The Netherlands.
Miller, G.A. (1956) The magical number seven, plus or minus two. Psychol. Rev. 63:81-97.
Mountcastle, V.B., Andersen, R.A. & Motter, B.C. (1981) The influence of attentive fixation upon the excitability of the light-sensitive neurons of the posterior parietal cortex. J. Neurosci. 1:1218-1235.
Murdock, B.B., Jr. (1983) A distributed memory model for serial-order information. Psychol. Rev. 90:316-338.
Nunez, P.L. (1981) Electric Fields of the Brain: The Neurophysics of EEG. Oxford University Press, London.
Nunez, P.L. (1989) Towards a physics of neocortex, Vol. 2, In: Advanced Methods of Physiological System Modeling, ed. V.Z. Marmarelis. Plenum, 241-259.
Nunez, P.L. (1995) Neocortical Dynamics and Human EEG Rhythms. Oxford University Press, New York, NY.
Nunez, Paul L. (2000) Toward a Quantitative Description of Large Scale Neocortical Dynamic Function and EEG Behavioral and Brain Sciences 23(3) (in press). ftp://ftp.princeton.edu/pub/harnad/BBS/.WWW/bbs.nunez.html
Nunez, P.L. & Srinivasan, R. (1993) Implications of recording strategy for estimates of neocortical dynamics with electroencephalography. Chaos 3:257-266.
Risken, H. (1989) The Fokker-Planck Equation: Methods of Solution and Applications. Springer-Verlag, Berlin.
Spearman, C. (1904). 'General intelligence, objectively determined and measured.' American Journal of Psychology, 15, 201-293.
Spearman, C. (1927). The abilities of man: Their nature and measurement. New York: Macmillan.
Zhang, G. & Simon, H.A. (1985) STM capacity for Chinese words and idioms: Chunking and acoustical loop hypotheses. Memory & Cognition 13:193-201.