|Chaytan||Biological Prediction - “Why Neurons Have Thousands of Synapses, a Theory of Sequence Memory in Neocortex”|
- With backprop, you can compute what you’re supposed to get from an error signal; there is no way in the brain to send a corrective to each neruon (maybe).
- Implication of backpropagation - every forward pass must also have a reciprocal neuron that goes the other way to propagate error. However, it is not true that every neuron in the brain has such capabilities (e.g. comes in self-updating ‘pairs’)
- Neocortex - responsible for ‘higher order’ intelligent thinking, planning, prediction, and consciousness.
- Composed of ~150k cortical columns, with thousands of neurons each.
- Within each column, there are six layers.
- A cortical column - mini-columnsa re small groups of pyramidal neurons that exist within one layer of each cortical column.
- Some of the mini-columns form ‘residual connections’ or standard linear connectives with other mini-columns.
- Hawkin’s claim: if the cortex has the same structure everywhere and learns, then there exists a general learning algorithm from this structure.
- Neocortex has pyramidal neurons: thousands of inputs from other neurons on synapses. Why so many synapses? When you get far from the body, the input becomes too weak to trigger action potentials.
Each neuron has thousands of dendrites
- Basal dendrites are far away and too weak to produce an action potential
- How does this distance play a role?
- Learning: input to prediction to next input to model adjustment.
- Predictive coding model: update less when you correctly predict the next input, update more when it was unexpected.
- Input → Prediction → Next Input → Adjust Model
- Local learning rules
- Continuous learning - you can’t just train in one environment and expect it to never train. Learning must be able to be continuously applied to new environmnets.
- Contextual information and prior experience - incorporates prior experience in the brain.
- Multiple simultaneous predictions - you can make multiple simultaneous predictions without necessarily being conscious of them.
- Robust to noise
- No mechanism for “ground truth”
- Brain is always aware of surrounding, but selectivey aware
- Dynamics of learning are changing (continuous)
What biological mechanisms meet these?
Key idea: predictions are encoded when an actively firing neuron connects to distal dendrites of another neuron that is predicted to come next.
Encode a prediction by making it easier to make the neurons we think will fire to fire when it receives an input it expects. This is a prediction. Local inhibition allows us to learn the things that we do not expect more than the things we do expect (fewer neurons firing - less learning).
Dendritic updates: if the next input was unexpected, the primed neurons do not fire faster. All the neurons encoding the unexpected input fire.
Only a few neurons fire at once in order to represent a given input. If a dense amount of neurons fire, then this pattern can be easily confused with other patterns. (Think - dense vs sparse networks, vanilla vs sparse autoencoders.)
- Importance of sparsity
- HTM Network. Numenta simulated the model using neural networks - it learned sequences of characters well and was robust to data perturbations. Very good at learning sequences.
- Can we add a depolarization mechanism to our networks?
- Can we add an inhibitory mechanism to encourage sparse representation of seen patterns?