Questions & explanations
1. Describe the fear circuitry in the brain and how it becomes dysfunctional in anxiety disorders.
The fear circuitry includes the amygdala, prefrontal cortex (PFC), and hippocampus. The amygdala detects threats and triggers fear responses. The PFC, especially the vmPFC, can inhibit the amygdala when the threat is not real. The hippocampus provides context, like remembering that a certain place is safe. In anxiety disorders, the amygdala is overactive, the PFC is underactive, and the hippocampus may not properly distinguish safe from dangerous contexts. This leads to fear in situations that are not actually threatening. For example, someone with social anxiety may feel intense fear in a normal conversation because their amygdala overreacts and the PFC fails to calm it.
2. How does glutamate dysfunction contribute to schizophrenia?
Glutamate is the main excitatory chemical in the brain. In schizophrenia, there is reduced activity at NMDA receptors, a type of glutamate receptor. This leads to problems with brain communication and can cause both positive and negative symptoms. For example, drugs that block NMDA receptors, like PCP, can produce symptoms similar to schizophrenia. Low NMDA activity on GABA neurons may lead to too much glutamate release, causing excitotoxicity and brain damage. This glutamate dysfunction is linked to dopamine overactivity, so treatments that target glutamate are being developed. Understanding glutamate helps explain why some patients don't respond to dopamine-based drugs.
3. What does a decoding algorithm do with neural activity?
A decoding algorithm takes patterns of neural activity, like firing rates of many neurons, and guesses what stimulus or behavior caused them. For example, it can predict which direction a monkey is reaching based on brain signals. Different algorithms use different math to make these guesses. Linear Discriminant Analysis (LDA) finds a line that best separates two conditions. Support Vector Machine (SVM) finds a boundary with maximum margin between categories. Population vector averages the preferred directions of active neurons to estimate a movement direction. Bayesian decoding uses probabilities to combine prior knowledge with neural data to make the most likely guess.
4. Give an example of how chronic pain differs from acute pain in terms of underlying mechanisms.
Acute pain is a direct response to tissue damage, like a cut or burn, and goes away when the injury heals. Chronic pain lasts beyond normal healing time (e.g., months or years) and often has no clear ongoing cause. In chronic pain, the nervous system itself is altered: there may be central sensitization, where spinal cord neurons are hyperexcitable, or peripheral sensitization, where nociceptors become more sensitive. Also, chronic pain often involves changes in brain areas that process emotion and memory, making the pain feel more distressing. Treatment for chronic pain often targets the nervous system (e.g., with drugs like gabapentin) rather than the original injury.
5. In an ideal observer analysis, what does it mean for a decision to be 'optimal'?
An optimal decision is one that maximizes expected gain or minimizes expected loss given the available information. The ideal observer uses a model that knows the true probabilities of different events and chooses the action with the highest expected utility. For example, in a visual detection task, the ideal observer would combine prior knowledge (e.g., how often a target appears) with sensory evidence (e.g., light intensity) to decide whether a target is present. This analysis provides a benchmark against which real animal or human performance can be compared. Deviations from optimality can reveal constraints or biases in the actual decision-making process.
6. Compare positive and negative symptoms of schizophrenia and their brain basis.
Positive symptoms are 'added' experiences like hallucinations (hearing voices) and delusions (false beliefs). They are linked to too much dopamine in the mesolimbic pathway. Negative symptoms are 'missing' experiences like lack of motivation, flat emotions, and social withdrawal. They are linked to too little dopamine in the prefrontal cortex (PFC) and reduced PFC activity. Positive symptoms often respond well to antipsychotic drugs that block dopamine. Negative symptoms are harder to treat and may require different approaches, like cognitive training or drugs that boost PFC function. Both types of symptoms can severely affect daily life.
7. Describe the process of cell type annotation after clustering scRNA-seq data.
After clustering, each cluster represents a group of cells with similar gene expression. To annotate cell types, scientists look for known marker genes that are highly expressed in specific cell types. For example, neurons might express Syn1, astrocytes express Gfap, and microglia express Cx3cr1. They compare the cluster's average expression to databases of known markers. If a cluster expresses markers of excitatory neurons, it is labeled as such. Sometimes, they use automated tools that compare to reference datasets. Annotation is confirmed by checking that the cluster's gene expression profile matches the expected biological function.
8. Explain how population decoding algorithms handle correlated noise.
Correlated noise means neurons' firing rates co-vary, e.g., due to shared input. Vector averaging ignores correlations, which can lead to suboptimal decoding. Maximum likelihood can incorporate a covariance matrix to weight evidence from correlated neurons less. For example, if two neurons are perfectly correlated, their joint response provides no more information than one alone. Bayesian decoding can also model correlations in the likelihood function. Ignoring correlations overestimates the information available. Properly accounting for correlations improves decoding accuracy, but estimating the covariance matrix requires many trials.
9. Give an example where an ideal observer analysis reveals that a subject is using a suboptimal strategy, and explain what the analysis tells us about the subject's decision process.
In a motion discrimination task, an ideal observer would combine all motion signals over time. If a subject's performance is worse than the ideal, but only for long stimulus durations, it suggests the subject may have a limited temporal integration window. The analysis can show that the subject's decision rule is to use only the first few hundred milliseconds of motion, ignoring later information. This reveals a constraint: the subject's brain cannot integrate evidence over long periods. The ideal observer benchmark thus highlights specific processing limitations, guiding further experiments to test hypotheses about neural mechanisms.
10. Compare divisive normalization with subtractive normalization.
Subtractive normalization subtracts a constant from the input, while divisive normalization divides. Subtractive normalization can produce negative responses, which are not physiological. Divisive normalization always yields positive responses and scales with input. For example, in contrast gain control, subtractive normalization would shift the response curve, but divisive normalization changes its slope. Divisive normalization is more robust and matches neural data better. It also implements a form of gain control that is multiplicative, not additive. In circuits, divisive normalization is often implemented via shunting inhibition.
11. Compare calcium imaging with GCaMP to electrophysiology for recording neural activity. What are the pros and cons?
Calcium imaging can record from many neurons simultaneously (hundreds) and identify cell types genetically, but it has low temporal resolution (milliseconds to seconds) and only indirectly measures spikes via calcium. Electrophysiology (e.g., patch clamp) records spikes directly with millisecond precision and can measure subthreshold activity, but it typically records from only one or a few cells at a time. Imaging is less invasive for long-term studies but requires a microscope; electrophysiology can be used in freely moving animals with wireless probes. The choice depends on whether population activity or precise timing is needed.
12. Why is divisive normalization considered a canonical computation?
Divisive normalization is found in many neural systems: visual cortex, auditory cortex, olfactory system, and even in decision-making circuits. It performs essential operations like gain control, contrast normalization, and attention modulation. The same mathematical form appears across species and modalities. This suggests it is a fundamental building block of neural computation. For example, in the visual system, it explains responses to complex stimuli like natural scenes. In decision-making, it can implement softmax selection. Its ubiquity makes it a 'canonical' computation, like a common algorithm reused in different contexts.