A major goal of learning and consciousness is to automate behavior--i.e., to transition from ‘thinking slow’ to ‘thinking fast’ (Kahneman 2011)--so that when an organism is subjected to a specific context that an automatic response will be executed with minimal participation from volitional circuits (i.e., in the neocortex). When one needs to enter a secure area, it is common for one to be confronted with a keypad upon which one must punch out the code to gain entry. At the beginning of learning the code, one is given a number, e.g., ‘3897’, which must be put to declarative memory. After having entered the facility on numerous occasions, one no longer needs to remember the number, but just the spatial sequence of the finger presses. Thus, the code has been automated by the brain. In fact, often the number is no longer required, since the nervous system automatically punches out the number using implicit memory (something like never needing to recall the rules of grammar to write correct sentences).

So, how does the brain automate behavior? The first clue to this question comes from studies on express, saccadic eye movements (Schiller and Tehovnik 2015). Express saccades are eye movements generated briskly to single targets at latencies between 80 and 125 ms. In contrast, regular saccades are saccadic eye movements generated to a single or to multiple targets (as used in discrimination learning such as match-to-sample) whose latencies vary from 125 to 200 ms, or greater depending on task difficulty (see Figure 14). The behavioral context for the elicitation of express saccades is to have a gap between the termination of the fixation spot and the onset of a single punctate visual target (Fischer and Boch 1983). The distributions of express saccades and regular saccades are bimodal, suggesting that two very different neural processes are in play when these eye movements are being evoked. After carrying out lesions of different parts of the visual system (i.e., the lateral geniculate nucleus parvocellular, the lateral geniculated nucleus magnocellular, area V4, the middle temporal cortex, the frontal eye fields, the medial eye fields, or the superior colliculus) it was found that lesions of the superior colliculus abolished express saccades, and for all other lesion types the express saccades were spared. Thus, a posterior channel starting in V1 and passing through the superior colliculus mediates express saccades (Schiller and Tehovnik 2015). Furthermore, the minimal latency for express saccades (i.e., 80 ms) is accounted for by the summed, signal latency between the retina and area V1 (i.e., 30 ms), the signal latency between area V1 and the superior colliculus (i.e., 25 ms), and the signal latency between the superior colliculus, the saccade generator, and the ocular muscles (i.e., 25 ms, Tehovnik et al. 2003)[1]. What this indicates is that express saccade behavior bypasses the frontal cortex and the posterior association areas of the neocortex (i.e., V4 and the medial temporal cortex), and is transmitted directly from V1 to the brain stem[2].

For oculomotor control, parallel pathways occur between (1) the posterior and the anterior regions of the neocortex (i.e., including, respectively, V1 and the frontal eye fields[3]) and (2) the brain stem ocular generator, which mediates ocular responses in mammals (Figure 15, Tehovnik et al. 2021). The idea that parallel pathways between the neocortex and brain stem mediate specific responses, such as the V1-collicular pathway subserving ocular automaticity, is not new. Ojemann (1983, 1991) has proposed that a multitude of parallel pathways subserves language, since once a language is mastered, it becomes a highly automated act, and electrical perturbation of a focal neocortical site affects a specific component of a language, but not an entire language string, as long as the remaining parallel pathways are intact. Global aphasia occurs when all the parallel pathways of Wernicke’s and Broca’s areas are damaged (Kimura 1993; Ojemann 1991; Penfield and Roberts 1966).

Why is it that express saccades and regular saccades alternate across trials in a quasi-random order (Schiller and Tehovnik 2015)? Lisberger (1984) has studied latency oscillations across trials for the vestibuloocular reflex by measuring the onset of an eye movement after the beginning of a head displacement. He found latency values as low as 12 ms and as high as 20 ms (Lisberger 1984; Miles and Lisberger 1981). At a 12-ms latency, the signal would need to bypass the cerebellar cortex and be transmitted from the vestibular nerve through the vestibular nucleus (which is a cerebellar nucleus) to the abducens (oculomotor) nucleus to contract the eye muscles within 12 ms (Lisberger 1984). At a 20-ms latency, the signal would pass from the vestibular nerve to the cerebellar cortex by way of the granular-Purkinje synapses and then to the vestibular and abducens nuclei to arrive at the muscles within 20 ms. The difference between the fast and slow pathway is 8 ms, and it is the additional 8 ms through the cerebellar cortex that allows for any corrections to be made to the efference-copy code[4].

In the case of regular versus express saccades, the minimal latency difference is 45 ms (i.e., 125 ms – 80 ms = 45 ms, Schiller and Tehovnik 2015). So, what could explain this difference? Regular saccades utilize both posterior and anterior channels in the neocortex, for paired lesions of the superior colliculus and the frontal eye fields are required to abolish all visually guided saccades (Schiller et al. 1980). Perhaps, the longer latency of regular saccades as compared to express saccades is due to transmission by way of the frontal eye fields for regular saccades, as well as having the signal sent through the cerebellar cortex via the pontine nuclei and inferior olive to update any changes to the efference-copy code. Express saccades, on the other hand, utilize a direct pathway between V1 and the saccade generator, with access to the cerebellar nuclei (i.e., the fastigial nuclei[5], Noda et al. 1991; Ohtsuka and Noda 1991) for completion of a response at a latency approaching 80 ms—a latency that is too short for frontal lobe/temporal lobe participation and the conscious evaluation of the stimulus (at least 125 ms is required for a frontal/temporal lobe signal to arrive in V1, Ito, Maldonado et al. 2023)[6]. Utilizing the fast pathway would not permit any changes to the efference-copy code and furthermore there would be no time for the conscious evaluation of the stimulus conditions. This general scheme for slow versus fast ‘thinking’ (Kahneman 2011) can be applied to any behavior, as the behavior changes from a state of learning and consciousness to a state of automaticity and unconsciousness[7].

While thinking slow, the human cerebellum can update as many as 50,000 independent efference-copy representations (Heck and Sultan 2002; Sultan and Heck 2003). And we know that during task execution the entire cerebellar cortex is engaged including circuits not necessary for task execution (Hasanbegović 2024). This global reach assures that all aspects of a behavior are perfected through continuous sensory feedback; hence, evolution left nothing to chance.

The number of neurons dedicated to a behavioral response decreases as a function of automaticity. This translates into a reduction in energy expenditure per response for the neurons as well as for the muscles[8]. The first evidence for this idea came from the work of Chen and Wise (1995ab) on their studies of neurons in the medial and frontal eye fields of primates (see Figure 15, monkey). Monkeys were trained on a trial-and-error association task, whereby an animal fixated a central spot on a TV monitor, and arbitrarily associated a visual object with a specific saccade direction by evoking a saccadic eye movement to one of four potential targets (up, down, left, or right) to get a reward (see Figure 16, left-top panel, the inset). An association was learned to over 95% correctness within 20 trials; unit recordings were made of the neurons in the medial and frontal eye fields during this time. The performance of an animal improved on a novel object-saccade association, such that the neurons exhibited either an increase in unit spike rate over an increase in the proportion of correct trials (Figure 16, novel, top panel), or an increase followed by a decrease in unit spike rate as the proportion of correct trials increased (Figures 16, novel, bottom panel, and Figure 17, novel, top panel). When the neurons were subjected to a familiar association, the discharge often assumed the same level of firing achieved following the asymptotic performance on novel associations: namely, high discharge and modulated (Figure 16, familiar, top panel) or low discharge and unmodulated (Figure 16, familiar, bottom panel; Figure 17, familiar, top panel). Accordingly, many neurons studied exhibited a decline in activity when subjected to familiar objects[9]. Although 33% of the neurons (33 of 101 classified as having learning-related activity) exhibited a declined and a de-modulation in activity during the presentation of a familiar object (e.g., Figure 17, familiar, top), this proportion is likely an underestimation, since many such neurons may have been missed given that unit recording is biased in favor of identifying responsive neurons. For example, a neuron that exhibited a burst of activity on just one trial could have been missed due to data averaging of adjacent trials, using a 3-point averaging method (Chen and Wise 1995ab).

For cells that had the properties shown in figure 16 (novel, top panel) for novel objects—i.e., showing an increase in activity with an increase in task performance—there was no delay in trials between the change in neural firing and the change in performance, as indicated by the downward arrow in the figure representing ‘0’ trials between the curves; this suggests that these cells were tracking the performance. Also, there was a group of cells that exhibited an increase and a decrease in unit firing such that their response to novel and familiar objects declined with the number of trials as well (Figure 16, bottom panels, novel and familiar). This indicates that the decline in activity was being replayed when the object became familiar. Finally, for neurons that exhibited an increase and decrease in spike activity over trials, the declining portion of the neural response (at 50% decline) always followed the increase in task performance by more than half a dozen trials, as indicated by the gap between the downward arrows of figure 16 (novel, bottom) and figure 17 (novel, top), illustrating that these neurons anticipated peak performance. Some have suggested that the short-term modulation in the frontal lobes is channels to the caudate nucleus for long-term storage (Hikosaka et al. 2014; Kim and Hikosaka 2013). More will be said about this in the next chapter.

Imaging experiments (using fMRI) have shown that as one learns a new task, the number of neurons modulated by the task declines. Human subjects were required to perform a novel association task (associate novel visual images with a particular finger response) and to perform a familiar association task (associate familiar visual images with a particular finger response) (Toni et al. 2001). It was found that as compared to the novel association task, the familiar association task activated less tissue in the following regions: the medial frontal cortex and anterior cingulate, the prefrontal cortex, the orbital cortex, the temporal cortex and hippocampal formation, and the caudate nucleus. Furthermore, the over-learning of a finger sequencing task by human subjects from training day 1 to training day 28 was associated with a decline in fMRI activity in the following subcortical areas: the substantia nigra, the caudate nucleus, and the cerebellar cortex and dentate nucleus (Lehericy et al. 2005). Also, there was a decrease in activity in the prefrontal and premotor cortices, as well as in the anterior cingulate.

Finally, it is well-known that a primary language as compared to a secondary language is more resistant to the effects of brain damage of the neocortex and cerebellum, and a primary language, unlike a secondary language, is more difficult to interrupt by focal electrical stimulation of the neocortex (Mariën et al. 2017; Ojemann 1983, 1991; Penfield and Roberts 1966). Accordingly, the more consolidated a behavior, the fewer essential neurons dedicated to that behavior. Once a behavior is automated, there is no need to recall the details: e.g., punching out a code on a keypad no longer requires an explicit recollection of the numbers. This is why a good scientist is also a good record keeper, which further minimizes the amount of information stored in the brain (Clark 1998). By freeing up neural space, the brain is free to learn about and be conscious of new things (Hebb 1949, 1968).

Summary:

1. Automaticity is mediated by parallel channels originating from the neocortex and passing to the motor generators in the brain stem; behaviors triggered by this process are context dependent and established through learning and consciousness.

2. Express saccades are an example of an automated response that depends on a pathway passing through V1 and the superior colliculus to access the saccade generator in the brain stem. The context for triggering this behavior is a single visual target presented with a gap between the termination of the fixation spot and the presentation of the target.

3. The rhythmical activity between express behavior and non-express activity across trials is indicative of the express behavior bypassing the cerebellar cortex and non-express behavior utilizing the cerebellar cortex to adjust the efference-copy code.

4. Express saccades or express fixations are too short in duration (< 125 ms) for a target to be consciously identified. It takes at least 125 ms for a signal to be transmitted between the frontal/temporal lobes and area V1 to facilitate identification.

5. Automaticity reduces the number of neurons participating in the execution of a behavioral response; this frees up central nervous system neurons for new learning and consciousness.

Footnotes:

[1] The long delay of 25 ms between V1 and the superior colliculus is partly due to the tonic inhibition of the colliculus by the substantia nigra reticulata, which originates from the frontal cortex (Schiller and Tehovnik 2015).

[2] Cooling area V1 of monkeys disables the deepest layers of the superior colliculus, thereby making it impossible for signals to be transmitted between V1 and the saccade generator in the brain stem (see figure 15-11 of Schiller and Tehovnik 2015).

[3] In rodents, the frontal eye field homologue is the anteromedial cortex, and the neurons in this region elicit ocular responses using eye and head movements (Tehovnik et al. 2021). In primates, the frontal eye fields control eye movements independently of head movements hence the name ‘frontal eye field’ (Chen and Tehovnik 2007).

[4] These short latencies are for highly automated vestibular responses. Astronauts returning from space have severe vestibular (and other) problems, and it takes about a week for full adaptation to zero-G conditions (Carriot et al. 2021; Demontis et al. 2017; Lawson et al. 2016). It would be expected that the latencies would far surpass 20 ms, since now vestibular centers of the neocortex (to engage learning and consciousness) would be recruited in the adaptation process (Gogolla 2017; Guldin and Grüsser 1998; Kahane, Berthoz et al. 2003). Patients suffering from vestibular agnosia would be unaware of the adaptation process, as experienced by astronauts (Calzolari et al. 2020; Hadi et al. 2022).

[5] The discharge of monkey fastigial neurons begins to fire 7.7 ms before the execution of a saccadic eye movement (Fuchs and Straube 1993). This nucleus is two synapses away from the ocular muscles.

[6] Presenting an unfamiliar object during an express fixation of an object (i.e., a fixation of less than 125 ms; fixations between electrically-evoked staircase saccades evoked from the superior colliculus are about 90 ms, Schiller and Tehovnik 2015) should fail to be identified consciously by a primate; on the other hand, the identification of a familiar object will only occur using ‘subconscious’ pathways during an express fixation, which are pathways at and below the superior colliculus/pretectum and the cerebellum (see: De Haan et al. 2020; Tehovnik et al. 2021).

[7] The conscious and unconscious states can never be totally independent, since the neocortex constantly monitors the behavior of an animal looking for ways to optimize a response in terms accuracy and latency (Schiller and Tehovnik 2015), and this interaction explains the variability of response latency across a succession of trials.

[8] Lots of aimless movements are generated when learning a new task (Skinner 1938), and when building knowledge, one must dissociate the nonsense from facts to better solve problems. This initially takes energy but in time automaticity saves energy.

[9] When we (Edward J. Tehovnik and Peter H. Schiller) first reviewed this result for publication, we were mystified by the decline of neural responsivity with object familiarity, even though we accepted the paper based on its behavioral sophistication and the challenges of recording from such a large number of neurons (i.e., 476) using a single electrode.

Figure 14. (A) The bimodal distribution of express saccades and regular saccades made to a single target by a rhesus monkey. (B) Before and after a unilateral lesion of the superior colliculus for saccades generated to a target located contralateral to the lesion. (C) Before and after a unilateral lesion of the frontal and medial eye fields for saccades generated to a target located contralateral to the lesion. Data from figure 15-12 of Schiller and Tehovnik (2015).

Figure 15. Parallel oculomotor pathways in the monkey and the mouse. Posterior regions of the neocortex innervate the brain stem oculomotor generator by way of the superior colliculus, and anterior regions of the neocortex innervate the brain stem oculomotor generator directly. For the monkey the following regions are defined: V1, V2, V3, V4, LIP (lateral intraparietal area), MT (medial temporal cortex), MST (medial superior temporal cortex), sts (superior temporal sulcus), IT (infratemporal cortex), Cs (central sulcus), M1, M2, FEF (frontal eye field), MEF (medial eye field), OF (olfactory bulb), SC (superior colliculus), and brain stem, which houses the ocular generator. For the mouse: V1, PM (area posteromedial), AM (area anteromedial), A (area anterior), RL (area rostrolateral), AL (area anterolateral), LM (area lateromedial), LI (area lateral intermediate), PR (area postrhinal), P (area posterior), M1, M2, AMC (anteromedial cortex), OB (olfactory bulb), SC (superior colliculus), and brain stem containing the ocular generator. The posterior neocortex mediates ‘what’ function, and the superior colliculus mediates ‘where’ functions.

Figure 16. Performance (percent correct) is plotted (solid black curve) as a function of number of correct trials on a trial-and-error object-saccade-direction association task. A monkey was required to fixate a spot on a monitor for 0.6 seconds, which was followed by a 0.6 second presentation of an object at the fixation location. Afterwards, there was an imposed 2-3 second delay, followed by a trigger signal to generate a response to one of the four target locations to obtain a juice reward; the termination of the fixation spot was the trigger signal (see inset in top-right panel: OB represents object, and the four squares indicate the target locations of the task, and Figure 17, bottom summarizes the events of the task). Chance performance was 25% correctness, and the maximal performance was always greater than 95% correctness established within 20 correct trials. The performance shown is the aggregate performance. In each panel, the normalized (aggregate) unit response is represented by a dashed line. The representations are based on figures 10 and 11 of Chen and Wise (1995a) for the medial eye field, and the neurons were modulated by learning novel object-saccade associations (N = 101 of 476 neurons classified). Some cells modulated by learning were also found in the frontal eye fields (N = 14 of 221 neurons classified, Chen and Wise 1995b). In the lower right panel, the familiar objects induced a decline in the neural response over the 20 trials. The illustrations are based on data from figures 11 and 12 of Chen and Wise (1995a).

Figure 17. Performance (percent correct) is plotted (solid black curve) as a function of number of correct trials on a trial-and-error object-saccade-direction association task carried out by a monkey. The dashed curves represent normalized aggregate unit responses. The inset in the right panel shows the task. For other details see the caption of figure 16. The bottom panel summarizes the events of the task. The illustrations are based on data from figures 3C, 4C, 5C, and 10D of Chen and Wise (1995a).