How deep is the brain?

Question 1

Why can the hierarchical view be problematic?

Answer 1

The more levels of hierarchy, the more synapses to pass through, the longer it takes –> the slowness of biological synaptic transmission matters

Question 2

Which brain areas project to the basal pontine nuclei and what does this mean for the hierarchy from the BPN point of view?

Answer 2

All cortical areas similarly projecct to BPN
–> from BPN’s perspective, there is no hierarchy, its all one big layer

Question 3

How is visual processing hierarchally built?

Answer 3

Following from V1 to the ventral pathway, more complex features are processed, but from this pathway on, V1 is just the first stage of visual processing

Question 4

What can subcortical projections from V1 do?

Answer 4

They can affect behavior

Question 5

What is the underrepresented anatomical evidence that there is a cortical hierarchy?

Answer 5

• The fact that each cortical area not only sends but also receives direct projections from subcortical areas
• The development of cortical hierarchy depends on the thalamus
• Extraordinary computational cpacity of each corticlal circuit, e.g. 1 pyramidal cell = 5-8 layer deep net.

Question 6

Which subcortical areas that project to the cortex?

Answer 6

• Thalamic nuclei (first order, higher-order, other)
• Claustrum
• Superior colliculus
• Neuromodulator nuclei

Question 7

What is the shallow brain hypothesis?

Answer 7

The brain has a shallow (fast) architecture interwoven with the cortical area hierarchy

–> The cortex is an array of massively parallel RNNs, each of which directly and reciprocally communicates with subcortical areas
- The computational capacity of hierarchical architecture is utilized
- Extraordinary computational capacity of cortical microcircuit permits fast yet powerful computations

Question 8

Which cells are the key player in the shallow architechture and why?

Answer 8

Layer 5 pyramidal cells as the key driver:
- They are the most active excitatory cells
- They directly receive signals from the thalamus and other subcortical areas
- Directly project to subcortical regions
–> create a disynaptic fast pathway

Question 9

How can shallow architectures compete with deep ones according to the McCulloch & Pitts model?

Answer 9

They see dendrites as passive receivers

Question 10

What does the McCulloch & Pitts model understimate?

Answer 10

It significantly understimates the computational capacity of real biological neurons

Question 11

What are the key computational benefits?

Answer 11

• Local learning: no need for backprpgagation of errors across layers
• Speed: synaptic transmission is slow –> the fewer synapses, the faster
• Compositionality & flexible combination of features

Question 12

Within the cortical hierarchy, different levels of information can be combined.
How does the brain select one or combine multiple cortical outputs?

Answer 12

Via the basal ganglia that connects to subcortical motor/premotor centres

Question 13

How is flexible combination of low and high-level information carried out?

Answer 13

In a subcortical area and not in the cortex

Question 14

Why is the basal ganglia important for combination within the hierarchy?

Answer 14

The BG projects to the thalamus and the brainstem. The projections are not independent of each other but are orchestrated to flexibly combine or regulate the massively parallel cortical outputs

Question 15

In what kind of tasks do shallow architechtures outperform deep ones?

Answer 15

In tasks that require compositionality and flexible combination of features