The notes have not been proofread. Please do your research before you pick anything from this post. It is meant to be a rough sketch of everything that I heard and noted at the conference. Since quite a bit of this is new to me, it is bound to be inaccurate, unspecific, and possibly even incorrectly quoted.
Day 0 Keynote - "Birdsong"
- Adrienne Fairhall
- Birds learn their songs by trial and error.
- The Zebra Finch has a single song.
- STDP may require sustained depolarisation or bursting to occur.
- The structure of the basal ganglia is pretty conserved in all mammals.
- EI -> atractor -> stability
- Dopamine effect is U shaped in avalanche distribution in basal ganglia, therefore, both too much and too little will give negative results.
- Q: Why do you need variability for learning? (Structured variability)
- Q: How do we isolate the variability that was "good"?
Day 1 keynote - Wilson-Cowan equations
- Jack Cowan
- Wilson-Cowan equations.
- Attractor dynamics in neural systems.
- Exhibit various stable behaviours
- Oscillations before settle to a fixed point
- Stable forms.
- In the Vogels self organising model
- CITE: paper in press
- Near a phase transition, no need to look at details of single neurons - you're not missing anything by ignoring single neuron details.
Limits of scalability of cortical network models
- Sacha van Albada
- Mechanism at \(N \rightarrow \infty\) is not the same mechanism at finite size.
- Inappropriate scaling can also cause the network to become unstable -> for example, cause large osciallations.
- Asynchronous irregular state, therefore, Gaussian inputs assumed
- LIF is like the rate model with white noise added to outputs
- So, while scaling you have to maintain effective connectivity and also maintain mean activities.
- Important to simulate at natural scale to verify.
Complex synapses as efficient memory systems
- Markus Benna.
- Dense coding.
- Also use SNR.
- Synaptic weight distribution gets wider and wider - diffusion process.
- Good synaptic memory model:
- Work with tightly bounded weights
- Online learning
- High SNR.
- Not too complicated
- Long life time
- CITE: Amit and Fusi 1994
- Cascade model of complex synapse
- Need a balance of LTP/LTD - otherwise your distribution is squished against one of the boundaries.
Self-organisation of computation in neural systems by interaction between homoeostatic and synaptic plasticity
- Sakyasingha Dasgupta
- Cell assembly properties
- Pattern completion
- I-O association
- Persistent activity
- Synaptic scaling is about 50-100 times slower than synaptic plasticity process
A model for spatially periodic firing in the hippocampal formation based on interacting excitatory and inhibitory plasticity
- Simon Weber
- If inhibition is not precise enough, you get periodic firing.
- Model of grid and place cells