Directional tuning became firmly established during the 1980s. Two key issues were raised and began to be addressed during that period. The first had to do with the unique neural coding of movement direction: given that the directional tuning curve is broad and symmetric, it can provide unambiguous information only at its peak (i.e. at the cell’s preferred direction), but it is unreasonable to suppose that only that point in the curve is used and all else discarded. This problem was solved by the neuronal population vector, an ensemble coding scheme (Georgopoulos et al. 1983, 1986, 1988) which became a nodal factor for the resurgence of computational neuroscience. The neuronal population vector has proved an effective way for decoding directional tuned motor cortical recordings currently used in neuroprosthetics (Collinger et al. 2013; Courtine et al. 2013) and for monitoring dynamic, time-varying cognitive operations (Georgopoulos et al. 1993; Pellizzer et al. 1995).

The second problem had to do with the cortical representation of the preferred direction (PD). Cells had different PDs which covered the whole directional range (Figs. 8.3 and 8.5) (Schwartz et al. 1988). Naturally, we asked the question of whether the preferred direction would be a feature (or, rather, the feature) of motor cortical columnar organization. Our first approach was similar to that used by Powell and Mountcastle (1959), namely to note the location of cells with specific PD along histologically identified penetrations and then observe possible en block changes in PD in penetrations at an angle with anatomical cortical columns. I presented our first results (Georgopoulos et al. 1984) at a meeting of the Neuroscience Research Program (NRP) held at the Salk Institute in 1983, after the NRP moved from the Rockefeller University to the Neurosciences Institute in La Jolla, CA. The results provided strong evidence for a columnar organization of the PD: in penetrations at the exposed cortex, PDs stayed very similar (Fig. 8.6), whereas they changed en block in penetrations at an angle with the anatomical columns (Amirikian and Georgopoulos 2003, Fig. 1). We went a step further and analyzed this relation quantitatively. For that purpose, we correlated two measures: one was this angle φ between the penetration and the anatomical columns that it crossed (Fig. 8.7) and the other was the spread of PDs along that penetration, measured as the circular standard deviation s ₀ (Mardia 1972) of the PD distribution. If the columnar organization of PD holds, we argued that, at the one extreme, when φ ≈ 0 (i.e. for penetrations along a column), then s ₀ ≈ 0, whereas, at the other extreme, when φ ≈ 90° (i.e. for penetrations perpendicular to columns, as in a bank), then s ₀ ≈ 180°, with intermediate values in-between. Indeed, we found a statistically significant positive correlation between φ and s ₀ (r = 0.756, p < 0.01) (Georgopoulos et al. 1984). This finding provided strong support for the columnar organization of the PD. Interestingly, at the same meeting, Bruce Dow presented the results of a similar analysis done for orientation selectivity in the visual cortex; their correlation coefficient was r = 0.6 (Dow et al. 1984). Therefore, our quantitative evidence for a columnar organization of the PD in the motor cortex was as good as, or better than, that for orientation selectivity in the visual cortex!

The next major step was to extend the testing of the columnar hypothesis for PD to 3-D reaching movements. Our first attempt provided clear quantitative evidence in that direction (Amirikian and Georgopoulos 2003). However, it was clear that we needed an experimental arrangement designed specifically for this problem. Specifically, we needed to (a) have a 3-D reaching task, (b) insert microelectrodes in a regular grid on motor cortical surface, (c) identify (or approximate) the trajectory of microelectrode penetrations, and (d) record neural activity during 3-D reaching at regular spatial intervals (depths) along a penetration. We successfully implemented those objectives as follows (see Naselaris et al. 2005, 2006a for details). We (a) employed the original 3-D reaching task (Schwartz et al. 1988), (b) constructed precise location-aligned templates for inserting and advancing 16 microelectrodes simultaneously using the Eckhorn Multielectrode matrix, (c) used dyes to identify the edges of the penetration matrix, (d) recorded neural activity simultaneously from 16 electrodes every 150 μm during task performance, (e) approximated the location of recording sites along the penetrations, (f) flattened the cortex, and (g) projected the PD in the recording sites onto the flattened cortical surface (Fig. 8.8). Thus, a 2-D cortical map was constructed with the PD color coded, after they were binned to octants (Fig. 8.9). Figure 8.9 shows that PDs were repeatedly represented on the motor cortical surface such that, within a given locale, practically the full range of the PD continuum was represented. This meant that an accurate neuronal population estimate of the movement direction could be derived from within any one of these locales (Naselaris et al. 2006b).

The next challenge, of course, was to find out whether PDs are organized in a columnar fashion. For that purpose, we used the full precision of the PD determination (i.e. without binning) and carried out a spectral analysis of the distribution of the PDs on the cortical surface (Georgopoulos et al. 2007). We identified 3 major peaks in the periodogram, namely one at a period of ~240 μm, another at a period of ~90 μm, and a smaller peak at ~30 μm (see Figs. 8.4 and 8.5 in Georgopoulos et al. 2007). These findings suggested a columnar organization of the PD with an estimated unit-column width of ~ 30 μm and a repetition of the full PD range every ~240 μm (Fig. 8.10). In fact, a regression analysis revealed an orderly increasing angular difference of PDs away from a given locus, up to 120 μm, suggesting an orderly representation of a series of PDs on the cortical surface (Georgopoulos et al. 2007). Altogether, these findings suggested a lattice representation of PDs, as illustrated in Fig. 8.10.

There is no doubt that the column is the organizing unit of the cerebral cortex. Although the computer metaphor of the brain has been discredited in many attempts, prominently fought against relentlessly by the late giant of neuroscience and dear friend Gerry Edelman, there is something to be said for the operational usefulness of the analogy.¹ Assuming the metaphor for the sake of argument, one can consider the cortical column as a microprocessor, and the brain as a set of massively interacting such microprocessors, i.e. a high-performance computer cluster. Accordingly, intra-columnar processing (Anderson et al. 2010; Opris et al. 2011; Apicella et al. 2012; Chadderdon et al. 2014) would correspond to operations within the microprocessor, whereas inter-columnar operations would correspond to serial and parallel operations in the cluster. The most immediate case concerns local inter-columnar operations, i.e. within an area of a radius of ~500 μm around a column (Gatter and Powell 1978; Georgopoulos and Stefanis 2010). Such local interactions would serve, e.g., to shape the directional tuning curve (Lee et al. 2012; Mahan and Georgopoulos 2013; Georgopoulos 2014). Long-range interactions among columns (Caminiti et al. 1985, 1988), and with spinal systems (Georgopoulos 1996), would correspond to large-scale, parallel computer-cluster type of operations. Figure 8.11 gives an overall picture of columnar and intercolumnar/hemispheric organization, based on known anatomical and physiological facts, as follows. (a) Cortical layers (laminae) are color-coded and labeled with Roman numerals (layer IV is omitted for this agranular cortex). (b) Neurons in different layers of the same column interact and are also synchronized (Opris et al. 2011). (c) Neurons from different layers project predominantly to different targets: layer II → ipsilateral cortex; layer III → contralateral cortex; layer V → subcortical structures (basal ganglia, brainstem, spinal cord); layer VI → thalamus. The extensive cortical synchronization observed in studies using various technologies, including fMRI (Christova et al. 2011), magnetoencephalography (Leuthold et al. 2005; Langheim et al. 2006) and local field potentials (Merchant et al. 2014), is probably due to multiple factors, namely (i) local mechanisms (Stefanis and Jasper 1964a, b), (ii) specific and non-specific thalamic afferents (Jones 2001), and (iii) synchronization among cortical layers, carrying over to their projections (Opris et al. 2011).