Big Picture
Visual perception and visual cognitive functions such as attention and perceptual decision making are usually associated with the cerebral cortex. There is no doubt that the impressive visual abilities of primates depend in a vital way on the exquisite architecture and structured connectivity amongst cortical areas of the brain. But biological organisms were using brains for vision for hundreds of millions of years before the cerebral cortex evolved1.
Recent work has revealed subtle roles for subcortical structures that suggest new approaches are required to understand how "higher" cognitive functions arise as a result of coordinated cortical-subcortical interactions. It is now established that the superior colliculus (SC) is necessary for visual attention – a growing body of literature (including my own contributions) has established that perturbing SC neurons causally influences which visual information is selected and which is ignored2-5. Work focused on understanding how the prefrontal cortex (PFC) can influence sensory processing based on internal goals has revealed it is partly accomplished by an intimate functional interdependence between PFC, the basal ganglia (BG) and the thalamus6. Finally, further work on the BG has implicated its participation in perceptual decision making7 and I recently found that attention-related effects of perturbing SC may also involve SC-BG interactions8.
The findings highlighted above again point to the importance of interactions between older subcortical structures with newer cortical brain areas for a variety of visual cognitive faculties. Because of the core importance of the BG for reinforcement learning (RL) and the parallel between cortical sensory architecture and “deep neural network” models, there is also a tantalizing connection between the neurobiological ideas outlined here and the recent dramatic progress made in the field of artificial intelligence (AI) through the use of “deep RL” agents which fundamentally rely on interactions between RL and deep neural nets9. These exciting ideas open the door to a variety of experimental, theoretical, and computational projects, two of which I describe below.
Subcortical circuit interactions for learning to identify behaviorally relevant visual information
In well-known contexts, selection of behaviorally relevant visual information is automatic – if you stop at a local coffee shop in the morning, you can quickly determine whether they have the flavor of scone you prefer, how many staff are working (and thus how long your wait will likely be), and where to wait for your drink to be prepared. When you visit a coffee shop in a foreign country, one of the first challenges you confront is figuring out which visual information is relevant: what part of the menu even indicates whether they serve scones? Is there a designated place for drinks to be placed once they’re prepared? Who are the employees and who are the customers? Which bits of visual information are most behaviorally relevant?
Which brain areas and mechanisms support the experience-dependent transition from novel (in which the parsing of relevant and irrelevant is unknown) to well-known situations (in which visual details are automatically categorized as relevant and irrelevant)? My working hypothesis is that interactions amongst subcortical brain areas including the superior colliculus (SC) and substantia nigra pars compacta (SNc), and dopamine-dependent reinforcement learning (RL) mechanisms are involved in “learning to attend”. Neurons in the SC show clear attention-related modulation, SNc is one important location for finding dopamine (DA) neurons implicated in RL, and SC neurons are known to project to SNc, a pathway that has been implicated as important for signaling the occurrence of biologically relevant visual events. However, the possibility that dopamine-dependent RL mechanisms and this pathway may be important for the establishment of attention-related modulation in the SC or elsewhere has not been tested. This goal of this project is to test the hypothesis that learning to identify behaviorally relevant visual information activates DA-dependent RL in SNc and results in attention-related behavioral and neurophysiological correlates.
What defines a state?
The ability to recognize a particular set of environmental conditions as a context or state substantially simplifies the interpretation of sensory information and the choice of behaviors or actions. When you walk into a new kitchen for the first time, you can quickly and easily recognize the appliances, and you know that putting a pot of water on the stove and turning the burner on will result in boiling water. There’s no need to carefully inspect each object to discriminate its identity or experiment with possible modes of acting to achieve desired outcomes. This kind of contextual generalization is adaptively advantageous because it dramatically reduces the need to learn specific object identities or behavioral responses. It would be quite cumbersome if one had to, for example, re-learn how to boil water every time one tried to do so in a new kitchen.
Context or state-specific learning allows for generalization, but it begs the question of how states themselves are established and identified in the brain. What neuronal evidence is there that a particular setting is recognized as belonging to some categorical “state”? Are there particular parts of the brain that are dedicated to representing states? How is a new context or state established in the brain as a result of learning? The striatum of the basal ganglia is an area of the brain that is thought to have access to a state representation for the estimation of the value of stimuli or actions10. Indeed, in my recent work, I found that neurons in part of the striatum (the caudate nucleus) carried state information about the stimuli in a covert attention task. However, I also found that this state information was degraded during attention deficits induced by perturbing neurons in the superior colliculus (SC). These findings demonstrate the possibility of reading out state-related information from striatal neurons and offer the novel hypothesis that attention deficits resulting from perturbing SC neurons may be understood as misestimations of the value of stimuli or actions associated with those stimuli. The goal of this project is to examine how states are established and represented, and to test the hypothesis that state-value estimation plays a role in visual attention.
Bibliography
1. Krauzlis, R. J., Bogadhi, A. R., Herman, J. P. & Bollimunta, A. Selective attention without a neocortex. Cortex; a journal devoted to the study of the nervous system and behavior 102, 161–175 (2018).
2. Bogadhi, A. R., Katz, L. N., Bollimunta, A., Leopold, D. A. & Krauzlis, R. J. Midbrain activity shapes high-level visual properties in the primate temporal cortex. Neuron (2020) doi:10.1016/j.neuron.2020.11.023.
3. Krauzlis, R. J., Lovejoy, L. P. & Zénon, A. Superior Colliculus and Visual Spatial Attention. Neuroscience 36, 165–182 (2013).
4. Herman, J. P., Katz, L. N. & Krauzlis, R. J. Midbrain activity can explain perceptual decisions during an attention task. Nature Neuroscience 21, 1651–1655 (2018).
5. Jun, E. J. et al. Causal role for the primate superior colliculus in the computation of evidence for perceptual decisions. Nat Neurosci 1–11 (2021) doi:10.1038/s41593-021-00878-6.
6. Nakajima, M., Schmitt, L. I. & Halassa, M. M. Prefrontal Cortex Regulates Sensory Filtering through a Basal Ganglia-to-Thalamus Pathway. Neuron 103, 445-458.e10 (2019).
7. Doi, T., Fan, Y., Gold, J. I. & Ding, L. The caudate nucleus contributes causally to decisions that balance reward and uncertain visual information. Elife 9, e56694 (2020).
8. Herman, J. P., Arcizet, F. & Krauzlis, R. J. Attention-related modulation of caudate neurons depends on superior colliculus activity. Elife 9, e53998 (2020).
9. Botvinick, M., Wang, J. X., Dabney, W., Miller, K. J. & Kurth-Nelson, Z. Deep Reinforcement Learning and Its Neuroscientific Implications. Neuron 107, 603–616 (2020).
10. Rao, R. P. N. Decision making under uncertainty: a neural model based on partially observable markov decision processes. Front Comput Neurosc 4, 146 (2010).