Modelling questions / aims¶
Starting point:
- What are the best strategies for localising sounds with a spiking neural network?
- Can we train networks end-to-end to perform as well or better than hand-crafted solutions?
- Zachary: I’m going to try a few things to improve performance
- Can we understand what these trained networks are doing in a high level way?
- Dan: I’m going to try out a few ideas for this.
- Do these networks do something similar to what has been proposed in the literature (labelled line models, hemispheric models, pattern match decoders) or something completely different?
- Lavinia: I’m interested into looking at this (how spiking neural networks can be used to model observed neural representations).
- Do different optimal models emerge in different parameter regimes (head size, signal to noise ratio, multiple sound sources)?
- Danish: I have some previous work on head size scaling which could be relevant here.
- How do the results depend on the neuron model and available dynamics? For example, does adaptation matter and in which conditions?
- Alberto: I’m interested into looking at this (neuron and synapse model, plasticity).
- Danish: I’m also interested in looking at this from the perspective of synaptic plasticity and auditory spatial working memory models.
- Dilay: I’m interested in using/adapting the filter-and-fire neuron model (recently proposed by Beniaguev et al., 2022) in the SNN.
Technical challenges¶
- Delay learning: can we train delays with surrogate gradient descent?
- Jakub: I’m interested in looking into this.
- Balázs: I’m looking into this using DCLS as in https://
arxiv .org /pdf /2306 .17670 .pdf
- Alternatives to surrogate gradient descent?
- Danish: I use a correlation learning model of synaptic plasticity which could in principle capture timing delays.
- Add a more realistic auditory periphery (cochlear filtering and more realistic spiking model).
- Danish: I have a simple vertebrate biophysical model of filtering in the auditory periphery which I could quickly test here.
- Number of time steps might be an issue for a more realistic model. May want to use dt=0.1ms and duration>=100ms so more than 1000 time steps.
- Add a more realistic sound localisation task (natural sounds, background noise, multiple sound sources, different frequencies).
- Tomas: experiment with this, starting with multi-frequency sounds
- Danish: I have looked into multi-frequency cochlear filtering for speech localization. It may be useful here.
- Divyansh: How well does the model perform at different frequencies (even if still just pure tones)
- Consider multiple architectures, potentially matching the auditory system.
- Alicja: I’ll look into that, beginning with checking the effect of adding more hidden layers.
- Mingxuan: I’ll experiment with this starting with altering the output neurons.
- Consider more complicated neuron models and perhaps make some of these neuron model parameters trainable; for what situations can that help?
- Danish: I use a correlation learning model of synaptic plasticity which could in principle learn timing and ITD information.
- Ido: Might be interesting to consider multi compartment biophysical neuron models. For starters, I’d be interested to try to add dendrites to the individual units (for example by using https://
github )..com /Poirazi -Lab /dendrify
- How to understand what the learned network does?
- Helena: This is just a warm up, and I am starting with a simple thresholding on the W1W2 plot to improve the contrast. I am going to push file named as SNN_sound_W1W2_threshold_plot.ipynb. I will probably come back to this later.
- Use regression instead of classification
- Mingxuan: I’d like to experiment with this idea.
- Which level of biological realisms does the learning need to comply to? Which observables are available at the synapse as input to a learning rule?
- Danish: Very interesting question that I would like to explore.
- Level of biological realism and impact on performance - Dale law.
- Jose: Looking into this
- Distribution of inhibitory and excitatory neurons using Dale’s law
- Sara: First exploratory and interesting results, warrant further investigation