3 ms·

So classical conditioning in humans requires special cells in the Cerebellum (Purkinje cells), which can even do single-cell learning. Which artificial neurons

So classical conditioning in humans requires special cells in the Cerebellum (Purkinje cells), which can even do single-cell learning. Which artificial neurons can't do, as only weights (artificial synapses) are updated. So how is classical conditioning actually implemented in artificial neural networks? I assume there is some minimum network which makes it work.

eurekin 3y ago
I always assumed ANN simply is a universal and learnable function approximator. That is, there is no direct equivalent of classical conditioning. Only data in, expected output pairs
- cubefox 3y ago
  There must be a minimal ANN architecture which implements classical conditioning. This architecture could be quite limited in what it can learn compared to ANNs in general. Similar to how feed-forward networks are limited compared to RNNs.
  - david-gpu 3y ago
    You can train single layer neural nets. Not very useful, but they do exist. There are certain ANN architectures that relied on, essentially, classical conditioning based on Hobbian learning rules and variants thereof. Kohonen self-organizing maps are an example of that. Not that such historical systems are popular today, though.
h0l0cube 3y ago
At a rough guess, each Purkinje cell is an MLP unto itself, and as the article states, this implies some orders of magnitude more computation for a brain simulation. I also heard something like 'a neuron is an MLP unto itself' on the Brain Inspired podcast. It's likely we've vastly underestimated the processing power of the brain.
im3w1l 3y ago
Well classical conditioning kind of only makes sense in the context of an agent that is receiving inputs and taking actions on them. Many neural networks don't solve problems of that type, and so have no need for classical conditioning. But when you do have such a problem conditioning is not very complicated. The normal algorithms and neural structures are designed to learn stuff like "when a given input happens a certain action must be taken" and thats all you really need for conditioning. How it actually does it? Well I guess with gradient descent it would work something like this: Every time there is a puff of air the network will be like "damn I should have blinked to avoid this" and so it makes its current internal statement a little more likely to lead to blinking. Gradually as it happens more times it will learn a strong association for the ringing bell or whatever. A small RNN could learn this.
- cubefox 3y ago
  Yeah. It's just not quite clear what the minimal example for such a network would be. I assume you have N inputs and one output. The output always active when input 1 is active, otherwise the output is inactive. So the other inputs are ignored. However, when one of those other inputs, x, tends to be temporally correlated with input 1, after a while x will generate an output upon activation even if input 1 isn't active. If x becomes decorrelated with input 1, x will again get ignored. Not sure what the simplest network architecture looks like that implements this behavior.
  - im3w1l 3y ago
    > The output always active when input 1 is active. Neural networks don't have instinctive behavior like that.
    - cubefox 3y ago
      Things can be hardwired
ChainOfFools 3y ago
> special cells in the Cerebellum (Purkinje cells), which can even do single-cell learning. As a neuroscience novice, I've always assumed that something about the gross model of the neuron, as far as I understand it, cannot be correct or is incomplete. Because I never understood why single cells aren't already performing single cell learning, given that there are always far more dendrites than axons. Since this characteristic turns each neuron into a lossy compression function, there has to be some process by which certain dendrites are considered 'more important' carriers of information than others, in order to make a tie-breaking decision about what to include in the compressed signal, and what to throw out, as the cell decedes whether or not to transmit an impulse (including whether or not to override prior inhibitory signals) back up the axon.
- cubefox 3y ago
  Well, not all incoming signals get the same weight for the outcoming signal, as the dendrites are e.g. more or less close to the part of the cell where the spikes are generated. But this computation is just analogous to the connection weights together with the activation function in artificial neural networks. That's not what enables classical conditioning in single Cerebellum cells.
- pests 3y ago
  I think it's possible. When you get down to the biology and chemistry the electrical action potential driving all this results in ion channels opening and closing on both sides to pass on the signal. I totally believe individual dendrite-axon connections can lose or gain strength over time based on an optimization of the efficiency of these ion channels. I think this is above and beyond a simple mathematical weight and the nueron structures around these dendrite-axon pathways change or express their genes differently over time.

eurekin 3y ago

I always assumed ANN simply is a universal and learnable function approximator. That is, there is no direct equivalent of classical conditioning. Only data in, expected output pairs

cubefox 3y ago
There must be a minimal ANN architecture which implements classical conditioning. This architecture could be quite limited in what it can learn compared to ANNs in general. Similar to how feed-forward networks are limited compared to RNNs.
- david-gpu 3y ago
  You can train single layer neural nets. Not very useful, but they do exist. There are certain ANN architectures that relied on, essentially, classical conditioning based on Hobbian learning rules and variants thereof. Kohonen self-organizing maps are an example of that. Not that such historical systems are popular today, though.

h0l0cube 3y ago

At a rough guess, each Purkinje cell is an MLP unto itself, and as the article states, this implies some orders of magnitude more computation for a brain simulation. I also heard something like 'a neuron is an MLP unto itself' on the Brain Inspired podcast. It's likely we've vastly underestimated the processing power of the brain.

im3w1l 3y ago

Well classical conditioning kind of only makes sense in the context of an agent that is receiving inputs and taking actions on them. Many neural networks don't solve problems of that type, and so have no need for classical conditioning. But when you do have such a problem conditioning is not very complicated. The normal algorithms and neural structures are designed to learn stuff like "when a given input happens a certain action must be taken" and thats all you really need for conditioning. How it actually does it? Well I guess with gradient descent it would work something like this: Every time there is a puff of air the network will be like "damn I should have blinked to avoid this" and so it makes its current internal statement a little more likely to lead to blinking. Gradually as it happens more times it will learn a strong association for the ringing bell or whatever. A small RNN could learn this.

cubefox 3y ago
Yeah. It's just not quite clear what the minimal example for such a network would be. I assume you have N inputs and one output. The output always active when input 1 is active, otherwise the output is inactive. So the other inputs are ignored. However, when one of those other inputs, x, tends to be temporally correlated with input 1, after a while x will generate an output upon activation even if input 1 isn't active. If x becomes decorrelated with input 1, x will again get ignored. Not sure what the simplest network architecture looks like that implements this behavior.
- im3w1l 3y ago
  > The output always active when input 1 is active. Neural networks don't have instinctive behavior like that.
  - cubefox 3y ago
    Things can be hardwired

ChainOfFools 3y ago

> special cells in the Cerebellum (Purkinje cells), which can even do single-cell learning. As a neuroscience novice, I've always assumed that something about the gross model of the neuron, as far as I understand it, cannot be correct or is incomplete. Because I never understood why single cells aren't already performing single cell learning, given that there are always far more dendrites than axons. Since this characteristic turns each neuron into a lossy compression function, there has to be some process by which certain dendrites are considered 'more important' carriers of information than others, in order to make a tie-breaking decision about what to include in the compressed signal, and what to throw out, as the cell decedes whether or not to transmit an impulse (including whether or not to override prior inhibitory signals) back up the axon.

cubefox 3y ago
Well, not all incoming signals get the same weight for the outcoming signal, as the dendrites are e.g. more or less close to the part of the cell where the spikes are generated. But this computation is just analogous to the connection weights together with the activation function in artificial neural networks. That's not what enables classical conditioning in single Cerebellum cells.
pests 3y ago
I think it's possible. When you get down to the biology and chemistry the electrical action potential driving all this results in ion channels opening and closing on both sides to pass on the signal. I totally believe individual dendrite-axon connections can lose or gain strength over time based on an optimization of the efficiency of these ion channels. I think this is above and beyond a simple mathematical weight and the nueron structures around these dendrite-axon pathways change or express their genes differently over time.