How good a match is required for resonance and category learning to occur? The answer to this question clarifies how some of the most familiar cognitive symptoms of autism arise.

The matching criterion is set by a vigilance parameter ρ that is computed within the orienting system A (Figure 7; Carpenter and Grossberg, 1987, 1993). The size of the vigilance parameter determines the generality of the recognition categories that will be learned. If vigilance is high, then learning of a concrete or specific category occurs, such as learning to recognize a frontal view of a familiar face. If vigilance is low, then learning of an abstract or general category occurs, such as learning to recognize that everyone has a face. In general, vigilance is chosen as low as possible to conserve memory resources, without causing a reduction in predictive success. Because baseline vigilance level is initially set at the lowest level that has led to predictive success in the past, ART models try to learn the most general categories that are consistent with their experiences. This property may clarify the overgeneralization that occurs in young children (Brooks et al., 1999) until category refinement is achieved by subsequent learning (Tomasello and Herron, 1988).

When a given task requires a finer categorization, vigilance is raised. Vigilance can be automatically adjusted to learn either concrete or general information in response to predictive failures, or disconfirmations, within each environment. Such a predictive failure could occur, for example, if a viewer classifies an object as a dog, whereas it is really a fox. Within ART, such a predictive disconfirmation causes a memory search that automatically shifts attention to focus on a different combination of features that can successfully be used to learn and subsequently recognize that the object is, in fact, a fox.

One way that vigilance can change due to a predictive error is by a process of match tracking. Here, vigilance is increased in response to a predictive error by the minimum amount that is needed to drive a search for a more predictive category (Carpenter and Grossberg, 1987). Since lower vigilance allows learning of more general categories, match tracking learns predictive categories by sacrificing the minimum amount of category generality. It hereby realizes a kind of minimax learning that conjointly maximizes generalization while minimizing predictive error.

A great deal is now known about how vigilance is computed in the brain. For example, a sufficiently big mismatch due to a predictive disconfirmation can activate the nucleus basalis of Meynert which, in turn, can release acetylcholine (ACh) at cortical layer 5 cells. ACh can then trigger a search for a better matching category, even if the previous match was deemed sufficient. Many challenging psychological and neurobiological data about cortical regulation of category learning and memory that can be explained by this vigilance mechanism are described in Grossberg and Versace (2008), Palma et al. (2012a,b), and Grossberg (2017a). In particular, a catastrophic collapse of both tonic and phasic vigilance control can help to explain how the dynamics of learning, recognition, and cognition fail during Alzheimer’s disease, and why disorders such as Alzheimer’s disease and autism are often accompanied by abnormal sleep patterns (Grossberg, 2017a). Also modeled are how these ART dynamics can be incorporated into larger neural architectures that are capable of learning view-, size-, and position-invariant object categories and using them to search for desired objects in a cluttered scene (Cao et al., 2011; Grossberg et al., 2011; Chang et al., 2014; Grossberg, 2018).

High vigilance has been predicted to cause symptoms of hyperspecific category learning and attentional deficits in some individuals with autism (Grossberg and Seidman, 2006). This prediction has been successfully tested in psychophysical experiments showing that hyperspecific category learning occurs in high functioning individuals with autism (Church et al., 2010; Vladusich et al., 2010), thereby augmenting previous reports of problems with prototype learning in individuals with autism (e.g., Klinger and Dawson, 2001). It is also known that individuals with autism can exhibit abnormal cholinergic activity in the parietal and frontal cortices that correlates with nucleus basalis abnormalities (Perry et al., 2001), as well as neuron pathology (Kemper and Bauman, 1998) and morphological abnormalities (Riva et al., 2011), consistent with our account of how vigilance is controlled by the nucleus basalis via ACh release.

Hypervigilance can have multiple effects on learning and cognition. In particular, variations in social situations that might otherwise be categorized as familiar can lead to many resets and attention shifts in a hypervigilant individual, thereby preventing effective learning and performance in them. Section 3 will clarify how, when hypervigilance interacts with underaroused emotional depression in an individual with autism, highly aversive emotional responses may be triggered, whose avoidance may lead to coping strategies that include an insistence on sameness and circumscribed interests.

Bayesian models of autism include a concept of precision that may be compared and contrasted with the ART concept of vigilance. For example, the Lawson et al. (2014) article about their “aberrant prediction account of autism” states that “The discrepancy between the sensory input and descending predictions of that input is known as the prediction error. This prediction error reports what stimulus-associated information is ‘newsworthy’ in the sense that it was unpredicted and informative. This information is passed up the hierarchy to inform higher-level expectations, which subsequently generate better predictions and thereby resolve prediction errors. The influence of (top-down) prior beliefs, relative to (bottom-up) sensory evidence, is controlled by the precision, or confidence placed in prediction errors at each level of the hierarchy (Friston, 2008). A high sensory precision will increase the influence of ascending prediction errors by turning up the ‘volume’ of sensory channels in which we place more confidence…Crucially, if the predictive coding account on offer is true, precision itself has to be estimated, much like estimating a standard error in statistics, in terms of its expectation…” [italics ours].

Although every prediction theory needs to somehow cope with fine vs. coarse predictions, the above Bayesian account differs in multiple ways from ART in terms of both heuristics and mechanisms. For example, the ART Matching Rule does not a compute a “prediction error” that is “passed up the hierarchy to inform higher-level expectations, which subsequently generate better predictions and thereby resolve prediction errors.” Instead, the ART Matching Rule uses excitatory matching to generate resonant brain states that trigger learning, and big enough mismatches to drive a memory search to discover categories whose critical feature patterns, or prototypes, will learn to better represent the current input pattern, without requiring a hierarchy of “higher-level expectations.”

ART does not require “higher-level expectations” because it uses computationally complementary attentional and orienting parallel processing systems, which have detailed support from multiple kinds of experiments. Vigilance could not be defined, or trigger a memory search, without interactions between these attentional and orienting streams. Vigilance does not have “to be estimated, much like estimating a standard error in statistics, in terms of its expectation.” Rather, vigilance just determines when an input exemplar is too novel to be classified by a previously learned category—for multiple possible reasons, emotional, cognitive, cultural—and drives a search process that automatically discovers and learns a more predictive category. When vigilance control carries out match tracking, it automatically realizes a kind of minimax learning in response to any sufficiently big mismatch, without any explicit link from the expectation that caused the mismatch to vigilance change.

It here needs to be kept in mind that the prediction that is mismatched in the world is not the top–down expectation that is learned to dynamically stabilize the learning of the category itself, and these expectations can represent totally different things; e.g., motor outcomes vs. sensory categories. Indeed, ART has been derived from a thought experiment that shows how the need to overcome several kinds of uncertainty leads directly to ART mechanisms (Grossberg, 1980). Finally, these ART concepts and mechanisms have successfully explained and simulated many psychological and neurobiological data, and all the main ART predictions have been supported by such data. In contrast, a Bayesian account does not have a natural representation in terms of identified brain circuits and regions. Nor does it represent the real-time interactive dynamics whereby brains give rise to the emergent properties of observable behaviors.

The CogEM model (Figure 8) explains how object categories, in sensory cortical regions like ITa, and object-value categories, in cortical regions like orbitofrontal cortex, interact with value categories, in subcortical emotional centers like amygdala and hypothalamus. These brain regions are linked by a feedback loop which, when activated for sufficiently long time, can generate a cognitive-emotional resonance. Such a resonance can support conscious feelings while using conditioned reinforcer pathways (from sensory cortex to amygdala) and incentive motivation pathways (from amygdala to orbitofrontal cortex) to focus motivated attention upon valued object representations. These attended object representations can, in turn, release commands to perform actions compatible with these feelings. The next two sections say more about these several types of categories and the learned interactions between them.

Four different types of learned representations are included in the CogEM circuit of Figure 8: Invariant object categories respond selectively to objects that are seen from any perspective. As noted above, they occur in ITa, among other cortical regions. Value categories are sites of reinforcement learning that control different emotions and incentive motivational output signals. They occur in amygdala and hypothalamus. Object-value categories respond to converging signals from object and value categories. They occur in orbitofrontal cortex. Finally, motor representations (M) control motor actions. They occur in multiple brain regions, including motor cortex and cerebellum.

Three types of learning are shown in Figure 8 between these representations: Conditioned reinforcer learning strengthens the pathway from an invariant object category to a value category. Incentive motivational learning strengthens the pathway from a value category to an object-value category. Motor learning enables the performance of an act aimed at acquiring a valued goal object. A fourth kind of learning strengthens the connections between an invariant object category and its object-value category during memory consolidation. This last kind of learning will not be further explained here. It is included in the nSTART circuit (Figure 3) that augments CogEM to include both a type of adaptively timed learning, called spectral timing, that involves the hippocampus, and modulation of memory consolidation by brain-derived neurotrophic factor, or BDNF (Franklin and Grossberg, 2017). When all of these factors interact within nSTART, the model can explain and simulate the complex pattern of memory consolidation problems that arises if learning is followed by early vs. late ablations of amygdala, hippocampus, or orbitofrontal cortex, including symptoms of the famous amnesic patient HM (Milner et al., 1968). Section “Adaptively Timed Conditioning and Behavior and Its Breakdown during Fragile X” summarizes the relevance of breakdowns in spectral timing toward explaining symptoms of FXS.

Reinforcement learning, say classical conditioning (Pavlov, 1927/1960; Kamin, 1968, 1969), occurs within conditioned reinforcer pathways (Figure 8) that convert a CS into a conditioned reinforcer when its object category is activated sufficiently often just before the value category is activated by an US, or other previously conditioned reinforcer CSs. As a result of this kind of learning, a CS can subsequently activate a value category via this learned pathway. When this happens, the CS is said to be a conditioned reinforcer because it can cause many of the same reinforcing and emotional effects as a US.

During classical conditioning, incentive motivational learning also occurs from the activated value category to the object-value category that corresponds to the CS, Incentive motivational learning enables an active value category to prime, or modulate, the object-value categories of all CSs that have consistently been correlated with it. It is the kind of learning that enables you to think of favorite foods when you are hungry.

Motor, or habit, learning adaptively calibrates sensorimotor maps, vectors, and gains that are used for sensory-motor control, after which a CS can read-out correctly calibrated movements via its object-value category.

Although the above summary describes only classical conditioning, the CogEM model was, in fact, introduced to explain key data about operant conditioning (Grossberg, 1971). Many reinforcement learning and motivated attentional mechanisms exploit shared neural circuits, even though the experimental paradigms and behaviors that activate these circuits may differ.

The CogEM circuit in Figure 8 needs to have two successive sensory processing stages, an invariant object category stage in the temporal cortex, and an object-value category stage in orbitofrontal cortex, in order to ensure that the object-value category can release motivated behavior only if both sensory and motivational support for that behavior is provided as inputs to the object-value category. A polyvalent constraint on an object-value category prevents it from firing unless it simultaneously receives input from its invariant object category and from a value category. In other words, an object-value category can fire only when the action that it controls is valued at that time. Only when it fires can an object-value category trigger an action. After learning occurs, a conditioned reinforcer can satisfy the polyvalent constraint by sending a signal directly to its object-value category, and indirectly to the object-value category via the (conditioned reinforcer)-(incentive motivational) pathway.

Each value category in the amygdala/hypothalamus also obeys a polyvalent constraint because it also needs two converging inputs in order to fire: a reinforcing input from a US or conditioned reinforcer CS and a sufficiently large internal drive input (e.g., hunger, thirst). Each value category can only then generate large incentive motivational output signals to object-value categories.

Thus, both the value categories and the object-value categories obey polyvalent constraints: Due to these constraints, a reinforcing cue does not activate strong incentive motivation, and with it action, to satisfy a drive that is already satisfied.

Previous articles review some of the many psychological and neurobiological data that the CogEM model has explained and predicted, and how it compares with other models of cognitive-emotional dynamics; e.g., Grossberg (2013, 2017a,b, 2018). One particularly interesting comparison relates to the ability of the CogEM model to explain and predict clinical data. Damasio (1999) has derived from clinical data a heuristic version of the CogEM model, and used it to describe cognitive–emotional resonances that support “the feeling of what happens.” Each processing stage in Damasio’s model (see his Figure 6) corresponds to a processing stage in the CogEM circuit of Figure 8. In particular, the “map of object X” corresponds to the sensory cortical stage where invariant object categories are represented. The “map of the proto-self” becomes the value category and its multiple interactions. The “second-order map” becomes the object–value category. And the “map of object X enhanced” becomes the object category as it is attentively amplified by feedback from the object–value category. As this cognitive–emotional resonance develops through the excitatory feedback loop between object, value, and object–value categories, the attended object achieves emotional and motivational significance, and motivated decisions are made that can trigger context-appropriate actions toward valued goals.

CogEM hereby embodies, and anticipated, key concepts of the “somatic marker hypothesis” which proposes that decision-making depends upon emotion, while also providing a mechanistic neural explanation (e.g., Grossberg et al., 2008) of the different properties of amygdala and orbitofrontal cortex in making these decisions (Bechara et al., 1999; Baxter et al., 2000; Schoenbaum et al., 2003). In particular, the effects of amygdala or orbitofrontal lesions on subsequent behaviors are described and explained in Grossberg et al. (2008) and Grossberg (2018), including how such lesions influence the brain’s computation of an object’s “desirability” (Rudebeck et al., 2017).

Many terrestrial animals learn to time their behaviors by distinguishing expected disconfirmations (or non-occurrences) of reward from UNexpected disconfirmations (or non-occurrences) of reward (Grossberg and Schmajuk, 1989; Grossberg and Merrill, 1992, 1996). An expected non-occurrence is said to occur if a reward is expected roughly a fixed amount of time after a discriminative cue occurs in a given situation. The non-occurrence of the reward before that time is then not interpreted as a predictive failure. Such an expected non-occurrence does not lead to a reset of short-term memory, an attention shift to focus on other events, emotional frustration, and/or the release of exploratory behaviors to enable search for the desired goal object elsewhere. If, however, the reward does not occur at the expected time, and is thus an unexpected non-occurrence, then these cognitive, attentional, emotional, and motor consequences can occur to enable the animal to find the desired goal object elsewhere.

The START model explains how hippocampal activity can maintain motivated attention when an expected non-occurrence occurs, via a learned hippocampal-to-orbitofrontal incentive motivational pathway (Figure 3), while it also inhibits the orienting system A (Figure 9). This is the same orienting system that, left uninhibited, would otherwise cause a reset of short-term memory, a shift of attention, emotional frustration, and/or the release of exploratory behaviors as part of the ART category learning and memory search circuit (Figure 7C). How frustration can be triggered by an unexpected event is explained in Section 3.

An animal or human who cannot adaptively time its expectations and behaviors to distinguish expected vs. unexpected disconfirmations will fail to successfully learn many kinds of behaviors in social settings where timing one’s behaviors to appropriately respond to the behaviors of others is essential for social learning and success. Given the social cognitive problems of some individuals with autism, it is instructive that adaptively timed responses fail to occur in various individuals with autism (Sears et al., 1994; Szelag et al., 2004).

Contextually appropriate timing of motivated responses is, for example, often needed to share joint attention, which is often deficient in individuals with autism (Filipek et al., 2000), and to thereby be able to carry out successful imitation learning (Grossberg and Vladusich, 2010), or even to receive action-contingent reward. Moreover, socially unsuccessful behaviors due to bad timing can lead to large numbers of unexpected outcomes, and thus to persistent novelty-sensitive arousal bursts (Figure 7C). Section 3.1 will explain how such arousal bursts can, in turn, cause hypersensitive emotional reactions that may lead to coping strategies to prevent these reactions, including an insistence on sameness and circumscribed interests. The persistent failure to get reward may additionally contribute to the development of insufficiently aroused value categories, thereby exacerbating these hypersensitive emotional responses.

What is the neural mechanism that realizes adaptively timed learning? Adaptively timed learning is carried out by a neural mechanism that is called spectral timing (Grossberg and Schmajuk, 1989). Spectral timing enables the START, iSTART, and nSTART models to span an interstimulus interval (ISI), or temporal gap, of 100s of milliseconds, or even seconds, between the offset of a CS and the onset of an US during trace conditioning, or other learning experience with a delayed reward or punishment. Such a delay is orders of magnitude larger than the typical response rates of individual neurons. This learning mechanism is called spectral timing because it activates a “spectrum” of cells which respond at different, but overlapping, times. After this type of adaptively timed learning occurs, the population of these cells, acting together, can generate a population response that is maximal at, or near, the time when the US is expected (Grossberg and Schmajuk, 1989; Grossberg and Merrill, 1992, 1996). This kind of response was originally reported in neurophysiological experiments about adaptively timed conditioning in the hippocampus (Berger and Thompson, 1978; Nowak and Berger, 1992; Tieu et al., 1999).

Each cell in such a spectrum reaches its maximum activity at different times (Figure 10A). If the cell response peaks later, then its activity duration is broader in time (Figure 10A). This is also true for the adaptively timed population response. Figure 10D shows the population responses after learning with different interstimulus intervals, or ISIs. The increase of response variance with ISI is called a Weber law, or scalar timing, property (Gibbon, 1977). In addition to generating the Weber law, these model population responses also exhibit the familiar Inverted-U of learning as a function of ISI, with learning attenuated both at very small, and very large, ISIs. Within this range, learned responses are timed to match the statistics of the learning environment (e.g., Smith, 1968).

Recent neurophysiological data about “time cells” in the hippocampus have strongly supported the spectral timing model prediction that a spectrum of cells with different peak activity times obey a Weber law. Indeed, MacDonald et al. (2011) wrote: “…the mean peak firing rate for each time cell occurred at sequential moments, and the overlap among firing periods from even these small ensembles of time cells bridges the entire delay. Notably, the spread of the firing period for each neuron increased with the peak firing time…” (p. 3). MacDonald et al. (2011) have hereby provided direct neurophysiological support for the prediction that spectrally timed cells exist (“small ensembles of time cells”) and that these cells obey a Weber law (“spread of the firing period…increased with the peak firing time”).

The adaptively timed population response is generated by multiplying, or gating, each spectral cell activity by an adaptive weight, or long-term memory (LTM) trace (Figure 10B). Each of these LTM-gated cell activities is then added to compute the population response (Figures 10C,D). During conditioning, each adaptive weight is amplified or suppressed if its cell activity does, or does not, overlap times when the US occurs; namely, times close to the ISI between CS and US. Learning hereby selectively amplifies output signals from cells whose timing matches the ISI, at least partially (Figure 10B). Most cell activity intervals do not match the ISI perfectly. However, the population response that computes the sum of the LTM-gated signals from all the cells is well-timed, and typically peaks at or near the expected ISI of the US (Figures 10C,D). Spectrally timed learning hereby enables the START model to learn associations between events that are separated in time, notably during trace conditioning.

Spectral timing reconciles two potentially incompatible design constraints. On the one hand, the learned spectrum should peak at around the ISI in order to maximize the probability that the behavior is correctly timed. On the other hand, the orienting system should remain inhibited throughout the preceding time interval (Figure 9) to prevent an attention shift and maladaptive exploratory behavior before the expected reward occurs. These two constraints are both satisfied because of the Weber law property: Each curve in Figure 10D begins to grow at time zero, so it can inhibit the orienting system throughout the initial time interval, but it peaks around the ISI to maximize the probability of a correctly timed response.

As noted in Section 2.2 spectral timing has successfully modeled behavioral, neurophysiological, and anatomical data about several parts of the brain: the hippocampus to maintain motivated attention on prefrontal plans for an adaptively timed interval to enable completion of a goal-oriented action (Grossberg and Schmajuk, 1989; Grossberg and Merrill, 1992, 1996; cf., Friedman et al., 2000), the cerebellum to read out adaptively timed movements while motivated attention is maintained (Berger and Thompson, 1978; Ito, 1984; Fiala et al., 1996), and the basal ganglia substantia nigra pars compacta (SNc) to release dopamine bursts and dips. These bursts and dips regulate new associative learning in multiple brain regions in response to unexpectedly timed reward and non-reward (Schultz et al., 1992; Schultz, 1998; Brown et al., 1999, 2004; Gerfen, 2000; Goto and Grace, 2005).

In all of these cases, the metabotropic glutamate receptor (mGluR) system plays a critical role in enabling cell responses to bridge long time intervals. Fiala et al. (1996) have, for example, developed a detailed neural model of the underlying biochemistry of spectral timing in the cerebellum. Fiala et al. (1996) simulated how slow responses may be generated postsynaptically by mGluR-mediated phosphoinositide hydrolysis and calcium release from intracellular stores. These responses are capable of bridging the interstimulus interval (ISI) between the CS-activated parallel fibers that contact Purkinje cells, and the US-activated climbing fibers that deliver teaching signals to the Purkinje cells (Figure 4), thereby causing learned long-term depression, or LTD, at (parallel fiber)-(Purkinje cell) synapses.

Fragile X symptoms and the role of mGluR dynamics in causing them can now be better understood as a consequence of cognitive, emotional, and behavioral problems that can occur if the adaptively timed circuits that are needed for learning, and consolidating memories of, temporally delayed associations break down. If spectral timing circuits in hippocampus, cerebellum, and basal ganglia are all deficient, say due to inoperative or degraded mGluR dynamics, then all the kinds of data that were summarized in Section 2.3 have immediate mechanistic explanations.

For example, children with Fragile X can exhibit behavioral problems of severe inattention (Fryns et al., 1984; Baumgardner et al., 1995) and ADHD symptoms (Cornish et al., 2004) because their hippocampal adaptively timed circuits cannot maintain motivated attention long enough to successfully carry out many behaviors. The mouse model for FXS experiences severe impairment of trace conditioning (Zhao et al., 2005) for the same reason: its circuit for spectral timing in the hippocampus is not working. Finally, mGluR and FMR1 abnormalities in cerebral cortical and hippocampal synaptic processes can cause deficient cognitive, learning, and motor deficits in Fragile X patients (Huber et al., 2002; Koekkoek et al., 2005; Nosyreva and Huber, 2006; Guo et al., 2012; Vinueza Veloz et al., 2012) because of the several ways in which hippocampal, cerebellum, and basal ganglia spectrally timed circuits support these processes.

If there is a way to pharmacologically restore mGluR dynamics in otherwise intact hippocampal, cerebellar, and basal ganglia circuits, then that could ameliorate FXS symptoms by restoring adaptively timed learning. If not, then operant conditioning methods may be helpful that either differentially reward sustained attention for increasingly long time intervals, or punish orienting behaviors during these time intervals. The net effect will hopefully be the ability to maintain attention for increasingly long time intervals, until it is time to learn a contextually adaptive response.

The text will first discuss affective mechanisms that contribute to the insistence on sameness and circumscribed interests. In order to explain how these mechanisms work, the value categories in the amygdala/hypothalamus (Figure 8) need to be refined to incorporate circuits that control opponent emotional states. After summarizing some main properties of these opponent processes, the text can explain how they can become underaroused, When this happens, paradoxical symptoms of emotional unresponsiveness combined with emotional hypersensitivity can be explained.

Value categories in amygdala/hypothalamus contain ON cells and OFF cells that are organized in opponent processes that are called gated dipoles (Grossberg, 1972a,b, 1980, 1984; Grossberg and Seidman, 2006; Dranias et al., 2008; Grossberg et al., 2008). These ON and OFF cells can represent opponent emotional and motivational states, such as fear vs. relief, hunger vs. satiety, and so on. Gated dipoles help to explain many data about both classical and operant conditioning, including conditioned acquisition, extinction, learned escape and avoidance, attentional blocking and unblocking, partial reinforcement acquisition effect, gambling behaviors, self-punitive behaviors, and other behavioral properties that currently have no other mechanistic neural explanations. Neurophysiological data from the hypothalamus that match affective gated dipole properties have also been simulated (Dranias et al., 2008; Grossberg et al., 2008).

The simplest gated dipole circuit is depicted in Figure 11A (Grossberg, 1972b). It has non-recurrent, or feedforward, pathways. When gated dipoles augment the dynamics of the CogEM model in Figure 8, their hypothalamic ON and OFF channels deliver inputs to the amygdala which, in turn, provides incentive motivational signals to object-value categories in the orbitofrontal cortex, and thereby influences what actions are taken to achieve valued goals. Gated dipoles help to explain how changing reinforcement contingencies alter motivated behaviors because they respond to either sudden decreases in reinforcing inputs, or to unexpected events, with an antagonistic rebound that shuts off ongoing ON cell activity and transiently excites OFF cell activity. The transient OFF cell activation is the antagonistic rebound (Figure 11A).

For example, a sudden reduction of a fearful shock can cause a relief rebound. Likewise, the non-occurrence of an expected shock can cause a relief rebound (Masterson, 1970; Reynierse and Rizley, 1970; Denny, 1971). The unexpected non-occurrence of food can, in contrast, cause a frustrative rebound (Amsel, 1962, 1992). Thus, rebounds can occur from negative to positive affects, such as from fear to relief, or from positive to negative affects, such as from hunger to frustration. These antagonistic rebounds enable the brain to modify its reinforcement learning to quickly adapt to changing reinforcement contingencies. For example, if the sudden reduction of a fearful shock is due to a successful escape behavior, then the relief rebound can trigger new conditioned reinforcer learning and incentive motivational learning (Figure 8), using relief to motivate that escape behavior. In a similar way, the frustrative rebound that occurs after expected food does not occur can drive forgetting, or extinction, of motivational support for the consummatory actions that no longer lead to food.

Simple mechanisms, occurring in a prescribed order, enable gated dipoles to cause antagonistic rebounds either in response to changes in reinforcer amplitude, or to disconfirmations of cognitive expectations of reward. The reader who does not wish to immediately read the mechanistic explanation of how this happens can jump directly to the next section.

These mechanisms are: non-specific arousal (I in Figure 11A), cell activation (variables x_i with i = 1–6 in Figure 11A), activity-dependent habituative transmitters (variables z_i with i = 1 and 2 in Figure 11A), competition (pathways with plus and minus signs in Figure 11A), and output thresholds (which cause the final ON and OFF cell output signals). The antagonistic rebound in response to offset of a phasic input, such as a shock to the ON channel (variable J in Figure 11A), is the transient OFF-response (e.g., relief) at the output stage of the OFF channel. This rebound is energized by a tonically active input I that delivers arousal equally to both the ON and OFF gated dipole channels (Figure 11A).

The ON and OFF cell activities x₁ and x₂ in Figure 11A respond to the sum of tonic-plus-phasic ON input I+J, and the tonic OFF input I, respectively, before they generate output signals f(x₁) and f(x₂) to the next processing stage. Before they reach the next processing stage, these signals are multiplied, or gated, by the habituative transmitters z₁ and z₂, respectively. The gated output signals f(x₁)z₁ and f(x₂)z₂ excite the ON and OFF cell activities x₃ and x₄, respectively, at the next processing stage. The habituative transmitters transform the step-plus-baseline activity pattern x₁ in the ON channel into the overshoot-habituation-undershoot-habituation pattern at activity x₃. The baseline activity pattern x₂ in the OFF chancel is converted into the habituated baseline activity x₄.

Next, the opponent competition occurs across the ON and OFF channels. As a result, the habituated baseline activity x₄ in the OFF channel is subtracted from the ON activity x₃ to compute x₅. The overshoot and undershoot in x₅ are now shifted to be above and below the equilibrium activity zero, respectively. Then activity x₅ is thresholded by half-wave rectification to generate an ON output signal. This output signal has an initial overshoot of activation, after which it habituates. The undershoot is inhibited to zero by the output threshold. The signs of excitation and inhibition are reversed in the OFF channel, leading to activity x₆. Activity x₆ is simply the flipped, or mirror, image of x₅ with respect to the zero equilibrium activity. Thresholding x₆ inhibits to zero the flipped overshoot, while allowing the flipped undershoot to generate the OFF channel output. This is the transient antagonistic rebound. Thus, the antagonistic rebound is due to a combination of arousal, habituative transmitter gating, competition, and thresholding.

The non-recurrent gated dipole must be refined to realize additional properties that are important in the control of learning and behavior. In particular, feedforward interactions are not enough. A recurrent, or feedback, gated dipole circuit is needed to realize additional learning properties. The recurrent gated dipole in Figure 11B is called a READ circuit, for REcurrent Associative Dipole (Grossberg and Schmajuk, 1989), There is recurrent feedback in both the ON and OFF channels: Activity x₇ reactivates x₁ in the ON channel, while activity x₈ reactivates x₂ in the OFF channel. In addition, adaptive weights, or long-term memory (LTM) traces, w_k7 and w_k8 sample the ON and OFF channels, respectively, thereby allowing multiple objects and events, with sampling signals S_k, to learn to become conditioned reinforcers when they are associated with reinforcing events at the gated dipole.

A READ circuit can support several basic functional properties (Grossberg and Schmajuk, 1989): First, it can maintain steady motivation while a behavior is being performed, even during sufficiently small environmental distractions, but can rapidly switch to support a new behavior with a different motivation if the distraction is big enough. Second, it enables affective learning to remain sensitive to any number of reinforcing events throughout the lifespan; the LTM traces do not saturate. Third, it enables affective memories to be preserved for a long time, even years, until reward or punishment schedules change, or cognitive expectations are disconfirmed. They can then be quickly modified. Finally, these properties help to explain data about primary and secondary excitatory and inhibitory conditioning, among other important properties.

Arousal typically remains within an optimal range to ensure useful value category properties. Maintaining this optimal range during waking hours is a major achievement of the affective brain. Failure to do so is reflected in behavioral symptoms of several mental disorders, including autism. In particular, the activity of a gated dipole circuit exhibits an Inverted-U as a function of its tonic arousal level I (Figure 11C; Grossberg, 1984; Grossberg and Seidman, 2006): Gated dipole outputs in response to either abnormally small or abnormally large arousal inputs are depressed. Intermediate arousal input sizes generate a Golden Mean of responding at the middle of the Inverted-U. These gated dipole properties are a consequence of the same mechanisms that enable a gated dipole to trigger antagonistic rebounds and to thereby quickly adapt to changing reinforcement contingencies. In particular, the Inverted-U can be traced to how the state of habituation in the dipole’s transmitter gates (square synapses in Figures 11A,B) divide the effects of signals through the dipole. This division creates a Weber Law of dipole responsiveness.

In an underaroused gated dipole, the response threshold to inputs is abnormally high but, after input intensity exceeds this elevated threshold, further increments in input intensity lead to hypersensitive emotional responses (Figure 11C). This happens because the habituative transmitter that divides dipole responses in abnormally small. These properties help to explain why individuals with autism may show at best small affective responses to some inputs, but hypersensitive responses to inputs that exceed the elevated threshold. This property can interact with the hypervigilance of some individuals with autism to cause emotional outbursts in situations where individuals without autism would not respond in this way.

As noted in Section 2.4 hypervigilant individuals learn hyperconcrete recognition categories that are accompanied by a narrow focus of attention. Many events that would not be unexpected to a person with a broader and more flexible attentional focus are unexpected to an hypervigilant individual, thereby causing multiple arousal bursts when environments change even moderately (Figure 7C). These arousal bursts, in turn, can cause hypersensitive emotional reactions if they input to underaroused hypothalamic gated dipoles. The frequent hypersensitive emotional responses when environments change can be so aversive that individuals with autism may learn to avoid them by indulging in a reduced set of behaviors, including perseverative behaviors that seek to maintain a level of sameness which avoids the mismatches that would otherwise trigger hypersensitive emotional responses.

With this background in hand, one can begin to understand how an individual with autism may seek refuge in an insistence on sameness and circumscribed interests.

Because this mechanism for shaping a need for sameness and restricted interests involves value categories that regulate reinforcement learning and motivated attention (Figure 8), it can be modified by operant reinforcement contingencies that differentially reward variable behaviors (e.g., Miller and Neuringer, 2000; Lam et al., 2008), and thereby strengthen their conditioned reinforcer and incentive motivational pathways so that they can competitively inhibit pathways that support more restricted behaviors.

These operant manipulations may not, however, directly modify the basal ganglia circuit properties that can support RMBs of individuals with autism. How these behaviors may be caused is explained in Section 3.2.

A study of deer mice with induced stereotypy reported an imbalance in neuronal activity that was expressed in terms of cytochrome oxidase (CO) levels within the motor cortex, striatum, nucleus accumbens, thalamus, and hippocampus. In particular, Lewis et al. (2007) found that animals with high stereotypy rates had low CO levels in these brain regions, while animals with low stereotypy rates showed high CO levels.

Many additional studies on stereotypy are consistent with the proposal that imbalances in the direct and indirect pathway can support repetitive behaviors. For example, it is known that changes in the development of the striatum, where the direct and indirect pathways occur, are involved in repetitive behavior in autism (Langen et al., 2014). Moreover, drug-induced stereotypy manipulations in the SNr of the direct pathway and the sub-thalamic nucleus (STN) of the indirect pathway can cause repetitive behaviors. In particular, administering an intranigral GABA agonist causes stereotypy in rats (Scheel-Kruger et al., 1978), whereas administering a serotonergic (5-HT2) antagonist in the STN reduces stereotypy. These procedures altered either directly (intranigral GABA agonist administration) or indirectly (intra-STN 5HT2 antagonist administration) inhibitory GABAergic tone in thalamocortical relay neurons (Brunken and Jin, 1993). Manipulations that disinhibited thalamocortical projections induced stereotypy, whereas manipulations that inhibited thalamus reduced stereotypy (Barwick et al., 2000). For a similar reason, injecting opiate agonists into the substantia nigra leads to stereotypies in rats (Iwamoto and Way, 1977) by disinhibiting nigrostriatal dopaminergic projections, and thereby presumably elevating striatal dopamine release, as has also been shown in mice (Wood and Richard, 1982).

A more recent study observed that decreased indirect pathway activity occurs in animals that develop high rates of stereotypy (Tanimura et al., 2011). An alternative mechanism involving the balance of striosomal activity and matrix activity within the striatum has been suggested to impact the regulation of behavioral sequences, such that the relative enhancement of striosome-over-matrix activation can predict the amount of stereotypy that develops in the animals, where the ratio of activation was assessed using Immediate Early Gene (IEG) expression (Graybiel et al., 2000). The above experimental studies are thus compatible with the hypothesis that imbalanced basal ganglia circuitry can give rise to stereotypical behavior by allowing prolonged gate opening to release repetitive behaviors from recurrent networks further downstream.

The basal ganglia have also been predicted to regulate action sequences through interactions of the direct, indirect pathways and hyperdirect pathways (Gerfen and Bolam, 2010; Haynes and Haber, 2013; Jin et al., 2014). Figure 2 illustrates this interaction between the different pathways and how GO and STOP gating signals may be incorporated into this interaction. The direct pathway consists of GABAergic projections from the dorsal striatum to the globus pallidus (GPi) of the SNr, which in turn sends GABAergic neurons to the thalamus. The indirect pathway, on the other hand, consists of GABAergic projections to the external globus pallidus (GPe), which further inhibits the GPi (Graybiel, 2000; Gerfen and Bolam, 2010). Thus, activation of the striatum at the initiation of an action would generate a GO signal that would inhibit the GPi, further disinhibiting the thalamus and permitting the release of action sequences. Activating the dorsal striatum at the end of the action sequence would initiate a STOP signal in the indirect pathway, which disinhibits the GPi, thus counteracting the action of the GO signal.

The non-execution of a plan or action has been attributed in modeling studies to either the inability to activate a sufficiently strong GO signal, or an overactive STOP signal blocking the implementation of the plan, or some combination of these GO and STOP signals acting together. In particular, Brown et al. (2004) have supported these claims about GO and STOP basal ganglia dynamics with explanations and simulations using their TELOS neural model (Figure 6). TELOS simulates how monkeys learn five saccadic eye movement tasks (fixation, saccade, overlap, gap, and delayed) using interactions between the basal ganglia; prestriate, inferotemporal, parietal, and prefrontal cortices; frontal eye fields; and superior colliculus. After learning of all five movement tasks, TELOS was able to use the learned parameters to simulate the neurophysiologically recorded dynamics of 17 different types of identified neurons during these behaviors. Silver et al. (2011) extended TELOS to the lisTELOS model to simulate learning and performance of sequences of saccadic eye movements from a spatial working memory, guided by the temporally coordinated firing of three different basal ganglia loops, and to simulate challenging neurophysiological data about this kind of task.

In addition to modeling studies that clarify how the direct and indirect pathways interact, it is known that the subthalamic nucleus (STN), which comprises part of the cortico–STN–pallidal hyperdirect pathway, has glutamergic projections to both the GPi/SNr and GPe, and hence has the ability to influence both direct and indirect pathways (Aron and Poldrack, 1999; Charpier et al., 2010). Nambu et al. (2002) describe how an engaged motor program is initiated, executed, and terminated with the correct timing, while other competing programs are inhibited. In particular, just before a voluntary movement is initiated, a signal through the cortico–subthalamo–pallidal hyperdirect pathway inhibits large areas of the thalamus and cerebral cortex that are related to, and that would otherwise compete with, the selected motor program. Then a second signal through the cortico–striato–pallidal direct pathway disinhibits and releases the selected motor program. A third signal that engages the cortico–striato–(external pallido)–subthalamo–(internal pallidal) indirect pathway inhibits its targets.

The excitatory and inhibitory actions of the direct and indirect pathways are realized by different receptor types. In particular, the D1 dopamine receptor is expressed mainly by neurons in the direct pathway, while the D2 dopamine receptor is expressed mainly by neurons of the indirect pathway (Graybiel et al., 2000; Gerfen and Bolam, 2010). The D1 receptors have an excitatory effect on the direct pathway, while the D2 receptors have an inhibitory effect on the indirect pathway (Wichmann and DeLong, 1996), as can be seen in Figure 2.

Another important region is the nucleus accumbens (NAc), which is a part of the ventral striatum of the basal ganglia that is significant for reward processing (Figure 2; Kelley et al., 1997; Knutson et al., 2001). The NAc is also known to have GABAergic projections to the globus pallidus (Swanson and Cowan, 1975; Mogenson et al., 1980; Newman and Winan, 1980). An increase in the inhibitory activity from the NAc to the GPi could reduce the effect of the STOP signal, leading to perseverative behavior through the action of downstream recurrent circuits. In other words, an increase in inhibition of the GPi by the NAc could be one of the causes for an imbalance between direct and indirect pathways that can support perseverative behaviors by enabling a basal ganglia gate to remain open longer than is necessary for eliciting an individual behavioral response.

The following experimental data are relevant to this hypothesis. An imaging study using MRI and other imaging techniques studied the shape of the basal ganglia in boys with autism, comparing them to a control group. They observed an overgrowth of the nucleus accumbens in the form of an outward deformation, in the group consisting of boys with autism. They also observed that the volume of this overgrowth was positively correlated with greater social and communication deficits (Qiu et al., 2010). Mogenson et al. (1980) proposed a model for initiation of locomotor action comprising the caudate nucleus, ventral tegmental area, nucleus NAc, and the GP. These authors summarized experiments showing that the NAc region in mice receives inputs from both the hippocampus and the prefrontal cortex (PFC). They noted that the hippocampus mediates transmission through D1 medium spiny neurons (MSNs), while the PFC mediates transmission occurs through D2 MSNs. If this is the case, then an abnormality in one or both types of information processing could shift the balance of activation in the NAc.

A study of Neuroligin-3 (NL3) mutated mice, who exhibit autism spectrum disorder symptoms, led to the discovery that the D1 MSNs in the NAc show reduced synaptic inhibition compared to excitation (Rothwell et al., 2014). Since the NL3 mutation has been found to produce autism-like symptoms, it is a possible animal model for aspects of autism (Radyushkin et al., 2009). The altered balance of excitation and inhibition between the cortical and limbic pathways could result in over-activity of the NAc, inducing increased GABAergic inhibition of the GP. This increased inhibition would disinhibit the thalamus longer than is required, resulting in the basal ganglia gate being open for a prolonged time, thereby enabling repetition of the gated behavior. The fact that dopaminergic circuits in the basal ganglia may contribute to repetitive behaviors in autism is consistent with the efficacy of dopaminergic, serotonergic, and opiate drugs in diminishing repetitive behaviors in animal models (Lewis and Bodfish, 1998).

Such repetitive behaviors can also be involved in perpetuating a vicious cycle. For example, although a repetitive behavior may help to satisfy the need for sameness by focusing attention on the repeated behavior, by the same token, a repetitive behavior can prevent attention from being focused on reinforcing or socially important sensory cues, and can thereby prevent the kind of recognition learning and social cognitive learning that might help to overcome some of the social isolation that the repetitive behavior perpetuates (cf. Bodfish et al., 2000; Miller and Neuringer, 2000; Lam et al., 2008).