In this chapter, a model of state inference and US prediction in the BLA- mPFC circuitry was presented to complement the CEA model on control of fear expression by tonic inhibition. While the two models are different in scope and methods, they constitute a coherent account of the neural circuitry of fear conditioning when combined. In this framework, the basolateral amygdala— in accord with the ITCs, prefrontal cortex and hippocampus—estimates the probability of impending US presentation. Note that this task is in general much more complex than mere associative learning, and it is presumably for this reason that a fairly complex network of structures is involved in the process. To develop this notion in further detail, I built on latent-variable models of conditioning which give a formalized account of structure learning. Learning structure, in these models, amounts to classifying experience into latent variables or states during learning, as well as learning CS-US contingencies for each state. Apart from explaining a number of behavioral effects, this framework echoes the notion of fear and extinction memory traces. The switching between states, or memory traces, has a neural substrate in the activity of neurons in the BA. Starting from this premise, the model ascribed subcomputations to the structures involved and the consistency with experimental results was demonstrated.
Subsequently, the CEA mediates fear expression based on the US-probability estimate it receives from its afferents. The CEA model in section 5 simulates this on a spiking neural network level and describes how modulation of tonic inhibition controls the responsiveness of the network to phasic stimulation. On a computational level, a key aspect of this model is that the control of responsive- ness should be governed by a number of factors, foremost US predictability, if the network is to serve its presumed function optimally. While GABA spillover is suggested in the model as a specific mechanism for estimating predictability by temporal integration of reward prediction errors, it is also conveivable that structures external to the CEA estimate US predictability and influence tonic inhibition, the more so, since uncertainty estimates are also needed in the model for other operations like state estimation in the BA.
Chapter 7
Conclusions and Outlook
Taken together, the two models presented in this work give a coherent account of how acquisition and extinction of fear responses, as well as the control of fear generalization, can be implemented in the neural circuitry. Particularly, extinction and fear generalization have important implications for the emergence of pathological anxiety. In this work, a model of probabilistic state-switching in the BA underlying extinction learning and mechanisms for controlling fear generalization in the CEA were analyzed and discussed using computational methods. These models reproduce known experimental findings and offer new insights into the mechanistic details and functional organization of the circuitry. Further, on a conceptual level, it was a principal goal of this work to make steps towards bridging the gap between high-level, computational models of fear conditioning and the implementational, neural network level. This combination is fruitful for constraining the models further—by both physiological constraints and functional considerations—and increases the potential for experimentally testable predictions. Correspondingly, this concluding chapter is devoted to outlining predictions and key hypotheses in more detail, addressing important open questions and possible expansions to the models, and finally providing an outlook on possible general directions for theoretical research on the neural basis of fear conditioning.
7.1
Predictions and Hypotheses
It is worth explicating the predictions that follow from the main hypotheses of the models in more detail at this point. The central hypothesis of the BLA model is that switching between different latent states is implemented in the basal part of the amygdala. This is inspired by and fully consistent with the experimental data and interpretation inHerry(2008). However, the interpretation in terms
88 CHAPTER 7. CONCLUSIONS AND OUTLOOK of latent variables and the details of the model presented previously allow for a number of arguably not too obvious predictions that follow from this hypothesis. Firstly, the model assumes two parallel pathways, LA and ITC, the involve- ment of which is controlled by the BA and mPFC. Under normal conditions, extinction learning is preceded by the activation of the alternative ITC pathway, in which a neural substrate of extinction memory is formed by synaptic plasticity. If state-switching is prevented by, e.g., pharmacological inactivation of the BA, however, some amount of extinction learning in the LA-pathway should still remain. This notion suggests that extinction that was acquired in this case is safe from renewal, because it posits actual unlearning of the original, LA-dependent fear memory trace. Hence, lower fear scores during the renewal test are expected when the BA is inactive during extinction learning and the extinction phase is long enough to still produce extinction learning.
Notably, state-switching also can have more subtle implications. Second- order conditioning and conditioned inhibition are two learning phenomena that happen in the same experimental procedure. In the first phase, one stimulus is conditioned by pairing with the US, while in the main phase of the experiment another stimulus is presented together with the previously conditioned stimulus. Initially, second-order conditioning takes place, i.e., the new stimulus also acquires a response, merely by pairing with the US. Subsequently, however, this new stimulus becomes a conditioned inhibitor. That means that when presented with a third stimulus that has been conditioned and elicits a response, the stimulus blocks the conditioned response (Yin,1994). Recently, it was proposed that this change from second-order conditioning to conditioned inhibition is associated with a transition to a more complex state in the animal’s model of the environment (Courville,2003). In the presented model, this would imply that the BA is involved in controlling the switch from second-order conditioning, which has been reported to be LA-dependent (Gewirtz,1997), to conditioned inhibition, which the model suggests would be mediated in the alternative ITC-pathway. Accordingly, BA-inactivation should enhance second-order conditioning at the expense of conditioned inhibition.
Another fairly subtle point relates to the processing of conditioned stimuli. The latent-variable models on which the BLA model is based infer a generative model of the environment, which means that they learn to infer the full probability distribution including the probability of conditioned stimuli. In our model, the LA performs discriminative learning, i.e., it does not learn about CS probabilities. Inference of CS statistics is mPFC-dependent in the model, and, correspondingly, effects that rely on the learning of CS statistics, like sensory preconditioning (see subsection 1.2.3), should be affected by lesions of the mPFC but not by
7.1. PREDICTIONS AND HYPOTHESES 89 Within the framework presented in this thesis, US prediction in the BLA is followed by a separate processing step in the CEA as a result of which freezing responses are initiated or not. Thus, the present account adheres to a model- based perspective of conditioning (Dayan, 2014) in that the decision to freeze is dissociated from estimating US prediction. This functional placement of the CEA together with the previously presented analysis of network dynamics allows for further testable predictions. For instance, it is assumed that the strength of mutual inhibition between the two CEA subpopulations is tuned such that the network is close to the bifurcation described in section 4.3. As a consequence, manipulations that increase synaptic efficacy only slightly in the entire network should have the effect of shutting down one population entirely. Conversely, decreasing the efficacy of GABAergic inhibition in the network should delay the acquisition of a response.
Moreover, it is conceivable that modulation of tonic inhibition and synaptic plasticity of BLA-CEA connections are mutually dependent. From a functional perspective, the combination of local—that is, neuron-specific—synaptic plastic- ity and the global—network-wide—modulation of tonic inhibition can have the effect of producing more reliable responses at the expense of discriminability of inputs. While tonic inhibition enhances network sensitivity for all inputs, synap- tic plasticity is input-specific but therefore also more susceptible to stochasticity in the input. Hence, noise-contaminated inputs can lead to variability in the synaptic weights, which can be detrimental to output reliability. However, if these two modes of plasticity are employed in combination, a good compromise between reliability and discrimination can be achieved, very similar to regulariza- tion for navigating the bias-variance-tradeoff in classification problems (Bishop,
2006). Assuming that function is optimized in such a way in the CEA network, one would expect that there exists a negative correlation between the magnitude of changes in synaptic strength and tonic inhibition during fear conditioning. That means, if there is stronger decreases in tonic inhibition in CEloff, there should be less synaptic plasticity. This follows also from assuming a reward prediction error as a driving force for changes in synaptic efficacy. If tonic inhibition is downregulated, network responses increase, leading to a smaller reward prediction error.
Finally, it is a central aspect of the high-level interpretation of CEA function that tonic inhibition is adjusted to uncertainty and US predictability. Normative analysis suggests that in situations of unpredictable threats, the animal is com- pelled to lower its freezing threshold by decreasing CEloff tonic ihibition. From this, it follows that higher decreases in CEloff tonic inhibition should be expected for animals that undergo partial conditioning or unsignaled US presentations. More broadly, taking into account that CEloff stimulation enhances anxiety
90 CHAPTER 7. CONCLUSIONS AND OUTLOOK (Botta,2015), we hypothesize that this adjustment of tonic inhibition in the CEA to uncertainty is the linking mechanism by which US unpredictability heightens anxiety.