Status Updates From Computational Psychiatry: A...

Computational assays, such as those based on Bayesian approaches, could help diagnostic tests. Computational psychiatry could also help psychotherapy: psychotherapy being a learning process, it may benefit from the rich computational understanding of learning processes.

Data-driven and theory-driven approaches are not incompatible: theory-driven models can provide descriptions that efficiently summarize complex data, and these summaries can provide inputs for machine-learning algorithms. The combination of both methods has been found to outperform data-driven approaches alone.

The dopaminergic-hyperinnervation hypothesis explains why all medications with well-established efficacy for Tourette syndrome - anti-psychotics, low-doses of certain dopamine agonists like pergolide, ecopipam, VMAT2 inhibitors - reduce dopaminergic transmission.

Females learn better from rewards than males do - a finding that is consistent with higher striatal presynaptic dopamine synthesis capacity and possible higher striatal dopaminergic innervations, as assessed by dopamine transporter binding, in females relative to males.

Progress in realizing the clinical benefits of computational studies will require a broader range of study designs to be implemented in the future. Studies to find associations between computationally defined processes and treatment response. Studies to pinpoint causality to allow for cognitive and pharmacological interventions.

Phasic activity of the central norepinephrine system may contain an estimate of volatility or unexpected uncertainty, consistent with current theories on the broader role of norepinephrine, which is argued to increase the gain of sensory representations and thus increase their impact on behavior.

Learning about a volatile process is more efficiently achieved with a higher learning rate. Humans estimate the volatility of the process they are learning about and tune their learning rate to increase learning efficiency.

While finding the optimal approach to generalization and credit assignment is a core question tackled by the machine-learning literature, to date fear-conditioning studies in humans that examine generalization have tended to rely on summary statistics rather computational approaches.

Reinforcement learning has been described as especially promising in this regard and has indeed shown potential for classification of depression from purely behavioral data without the need for (subjective) questionnaires.

An important aim for computational psychiatry is the development of computational assays that can be used to separate patients into subgroups, generate treatment recommendations, and make predictions for the outcome of those treatments.

This suggests that model parameters obtained by fitting a behavioral task, such as probabilistic learning task, could be used to classify MDD to a high accuracy. The classification of diseases is an important goal of computational psychiatry.

A linear classifier can be used to distinguish between healthy and depressed participants after they had played a slot machine, based purely on the two values of individuals' parameters (sensitivity to reward and a prior belief about control).

Model-based fMRI can reveal differences in reward learning, even in the absence of behavioral effects.

A neural network that had overlearned on negative information could be retrained using positive information (akin to a cognitive behavioral therapy), which resulted in the normalization of network activity in response to negative information. The longer the network had 'ruminated,' the longer it took for the 'therapy' (i.e., retraining) to work, providing insights into the recovery from depression using CBT.

An important future challenge will be to link belief-updating parameters to those of spiking network models in order to understand how NMDAR function in reference to both pyramidal cells and inhibitory interneurons as well as neural 'noise' contributes to attractor instability, response stochasticity, and inference in general.

In schizophrenia there's a hierarchical imbalance in synaptic gain, as primary sensory areas have been shown to be 'hyperconnected', whereas higher regions are 'hypoconnected'. If this corresponds to a similar imbalance in the encoding of precision in a hierarchical model, then its effect would be to reduce the effect of priors on inference and cause larger belief updates in response to unexpected sensory evidence.

Incorrect calculations, leading to incorrect acquisition or calculation of long-run values, are another substantial source of problems. The automaticity of the Pavlovian pruning of the internal search used to calculate expected future values, such as when encountering losses, has been considered a point of vulnerability that could be relevant to mood disorders.

One way for the brain to inhibit exploration would be to dial down the subjective utility of outcomes. This would be pernicious, since failing to explore may entail failing to find out that the prior no longer pertains. Various such failings of the prior can lead to forms of depression, but they can readily extend to addiction and beyond.

Following seemingly random aversive events, a person could develop a prior that the world is not very controllable, with actions having highly stochastic consequences. This would come with the implication that there is little point exerting effort trying to explore it, since the information gained would not be expected to be exploitable.

A unitary mode of reinforcement of choices comes from outcomes being better than expected, rather than good.

Distal rewards are particularly important for quantities such as money, whose immediate utility is questionable. Again, decreasing marginal utility with increasing wealth has been suggested as underpinning phenomena such as risk aversion, that is, the reluctance to accept risky monetary prospects even when they involve an expected gain.

The evidence is so far consistent with the hypothesis that in schizophrenia the ability to represent and actively maintain contextual or task-goal information is disrupted. In future investigations, it will be important to more directly test the claims of the cascade model of hierarchical cognitive control that these deficits map appropriately along the rostro-caudal axis of PFS among individuals with schizophrenia.

Together, the algorithmic context-task-set (C-TS) and the neural network model make an important claim: that humans have a bias toward structured learning, even when it is costly, because such learning enables longer-term benefits in generalization and overall flexibility in novel situations.

Earlier models also utilized hierarchical frameworks to understand temporal abstraction in behavior, but the primary thrust of the cascade model and related variants has been to use reinforcement learning to subdivide temporally abstract complex action plan into simpler behaviors, an adaptive and efficient encoding strategy relevant for understanding structured abstract action representations.

Computational approaches can be utilized in both a reductionist and emergentist manner: deconstructing the mysterious intelligence of the homunculus into hopefully more understandable 'dumb' neural subcomponents, while at the same time making clear how complex control functions can emerge from the dynamic interaction among these multiple, simpler subcomponents of cognitive control.

Biophysically based circuit modeling has the potential to provide a method for simulating possible effects of treatments that act at the level of ion channels and receptors.

Despite the complexity of synaptic alterations in a disorder such as schizophrenia, the impact on cognitive function in neural circuits may be understandable in terms of their 'net effect' on effective parameters, such as E/I ratio, to which the circuit is preferentially sensitive.

An alignment between the clinical study, basic neurophysiology findings, and computational modeling allows stronger inferences and testing of hypotheses.

The standard psychophysical measurements from clinical populations cannot dissociate among distinct circuit-level alterations: elevated E/I ratio, lowered E/I ratio, or an upstream sensory-coding deficit.