Expansion-Contraction Architecture and Olfaction Flashcards
(32 cards)
What are expansion-contraction architectures and why are they useful? (broad)
A generalised neural network structure. Seen in both biological and artificial systems.
Used to map complex and overlapping input data onto specific output functions while maintaning high discriminative power.
What are the overarching layers in an expansion/contraction architecture?
Sampling:
* Process of detecting stimuli (input)
* Defines which stimuli can be detected
(Data cleaning and whitening)
Expansion:
* Increases dimensionality of data
* Separates out stimuli combinations (forms a bank of possible stimuli combinations)
Contraction:
* Represents output functions of network
* Several expansion cells can feed into one output cell
What is the point of whitening in E-C architectures (e.g. olfaction)?
Process of removing irrelevant differences in inputs so that underlying patters can be determined:
E.g. in olfaction:
* Removes small variations in intensity and duration
* This information is not lost but processed separately
* Increases the differences between odorants leading to better discrimination
Why is olfactory information noisey?
There is redundancy in olfactory receptor binding:
* Binds strongly to a particular chemical structure
* Many similar structures will bind weakly
* Leads to noisey sampling since there may be many different molecules which cause same level of stimulation of a receptor.
How is whitening achieved in the vertebrate olfactory system?
Receptors of the same type feed into one olfactory glomerulus. Therefore a glomerulus could be active because there is a low concentration of a strongly binding molecule ot there is a high concentration of weakly binding molecule.
* Causes a problem since odorants are identified by the pattern of glomeruli active
Whitening:
* Periglomerular cells inhibit within a glomerulus – silences glomeruli with only weakly firing inputs (i.e. not very complementary) but allows strongly firing ones through.
* Superficial short axon between glomeruli and granule cells between output neurons: compare glomeruli to each other (remove weakest firing glomeruli)
* Means what reaches piriform cortex neurons are only the most strongly activated glomeruli - exaggerates differences between odorants to improve discrimination.
How is whitening achieved in the insect olfactory system?
Receptors of the same type feed into one olfactory glomerulus. Therefore a glomerulus could be active because there is a low concentration of a strongly binding molecule ot there is a high concentration of weakly binding molecule.
* Causes a problem since odorants are identified by the pattern of glomeruli active
Whitening:
* Broad lateral neurons (~PGN) inhibit within a glomerulus – silences glomeruli with only weakly firing inputs (i.e. not very complementary) but allows strongly firing ones through.
* Picky lateral neurons (~SSAN) between glomeruli and between output projection neurons: compare glomeruli to each other (remove weakest firing glomeruli)
* Means what reaches KCs are only the most strongly activated glomeruli - exaggerates differences between odorants to improve discrimination.
What determines the number of different olfactory receptors?
Olfactory detection neurons:
* ORNs in insects
* OSN in vertebrates
Receptors are ligand gated ion channels controlled by orco genes. These are recombined in a hypervariable region to produce many differet receptor binding sites.
What are the different types of neurons connecting glomeruli to expansion layers (in insects and vertebrates)
Insect:
* Uniglomerular projection neurons (~mitral and tufted) - for better odorant discrimination and normalisation between sides
* Multiglomerular projection neurons - fast, powerful response for evolutionarily important stimuli.
Vertebrate:
* Mitral cells (sustained response)
* Tufted cells (transient response)
* No direct equivalent to mPNs but some mitral cells seem to feed directly into hippocampus, amygdala, hypothalamus
How do mPN neurons work? Why are they important?
Muti-glomerular projection neurons sample from several glomeruli (i.e. activate if a particular combination of glomeruli are active).
* Project directly to lateral horn (can bypass mushroom body).
* Drive innate behaviours in a fast and powerful way (useful for molecules always bad e.g. bacteria toxin)
* Associations are engrained and not learned (cannot change association)
Is there an equivalent to mPNs in vertebrates? Why?
While there are no direct equivalents, some mitral cells do project from olfactory bulb to amygdala/hypothalamus:
* Seem to serve a similar purpose in encoding potent and innate behaivours (e.g. disgust reflex)
May be because vertebrates rely more on associative learning, less on reflexes:
* More space for better odor discrimination
* More behavioural plasticity
Describe the expansion layer of insect olfactory system. Why are all odorants not represented?
One cell in expansion layer = Kenyon cell receives input from a selection of glomeruli (via uPNs)
* When this matches the pattern activated by a particular odorant, the KC fires - i.e. each KC represents a particular odorant.
However, there are not enough KC cells to represent every combination:
* Too many combinations (Bell’s number)
* Anatomical constraints.
* The more KCs there are the better the discrimination (e.g. drosophila larvae 70 → 2000KCs)
Since the size of expansion layer is not limited by physical space in artificial neural networks, why is discrinimation still limited?
Running the system:
* The larger the size of the layer, the more computational power required to run it.
Forming the network:
* Evolution has pre-pruned less useful combinations off (biased the network) which can be adapted (odorant discrimination is not uniform across all smells, it is adapted for environment).
* Huge energy requirement and data input needed to do the same in an artificial neural network
Why is a contraction layer necessary in olfactory processing?
The number of identifiable odorants is significantly more than the diversity of behavioural outputs possible.
* E.g. 2000KCs in drosophila, a few innate behaviours, therefore contraction to 34MBONs
What is the evidence MBONs have a causal role in behavioural outputs?
Sufficiency: Aso et al
* Optogenetic stimulation can drive behaviour independent of olfactory stimulus
* Shown by transgenic flies with optogene in aversive gustatory neuron, move away from light side of petri dish without stimulus
Necessary:
* inactivation abolishes behaviour to specific stimulus
Why is lateral inhibition rarely seen in neural networks but common in biological?
Neural networks with lateral inhibition have been tried (e.g. Kohonen’s self-organising maps) but are not generally used.
* Difficult to scale and energy intensive in network development (network running is more efficient)
Suggest some systems where expansion-contraction networks are employed in biology:
- Olfaction (insect and vertebrate)
- Cerebellum
What are APL neurons and why are they necessary?
Anterior paired lateral neurons:
* Provide lateral inhibition at the KC level so that only the strongest firing KCs stimulate MBONs
* Prevents multiple MBONs from being fired simultaneously
What is the Proust phenomenon?
Powerful and vivid memories triggered by smells
How can the association of a smell be changed?
Changing the strength of synapses between expansion and contraction cells:
* E.g. KC-MBON connection can undergo targeted synaptic depression to reduce connection (or visa versa)
* Allows for associative learning
Provide an example of internal stimuli changing association between a smell and its behavioural response.
Hunger changes response to apple cider vinegar in drosophila (Vogt et al)
* Hunger causes release of 5-HT
* Suppresses mPN pathway for odour avoidance
* Evolutionarily sensible - some food better than no food.
What is the Braitenburg model of associative memory?
Employing Hebbian principles (‘cells that fire together, wire together’), complex associative memory can arise from simple connections.
Think about visual perception and memory (e.g. single coding neurons - paired coding neurons…)
What is Shultz’s prediction error coding theory? (1997). Give an example where this is employed in biology.
The prediction error = reward (expected) + surprise signal
* DA neurons are the effectors from this circuit, modulating the expected signal based on feedback
Example: Mushroom body of drosophila
* Strength of KC - MBON neurons which drive behavioural response to smell
* Strength of connnection is highly plastic
* DAN/OAN neurons release DA/octopamine in response to feedback neurons activation
* Learning to like a smell is effectively disliking it less
Give an examples of a command neuron
First discovered in crayfish
In a fly:
* The ‘moonwalker descending neuron’ - stimulates backward motion to aversive stimuli
* Lateral horn output neurons (general)
Maggot:
* Goro neuron stimulating rolling
* B1 (hunching), B2 (bending)
What defines a command neuron?
Necessity and sufficiency to cause a particular behaviour
Bridge between multi-sensory integration and motor execution modules. For decision making.
