1
$\begingroup$

I was under an impression that artificial intelligence is modelled after organic intelligence. Under the context of organic intelligence, it seems that some individuals are capable of getting caught with the condition of learned helplessness towards their own judgments after exposing to repeated trauma or failures.

I'm curious if such "psychology" could ever influence an AI. Suppose that a node repeatedly fails at producing results relative to other nodes from its layer, would a node underweight itself similarly to how a pessimist underweights their own achievements?

How long does it take for a node to correct its bias, and is there a situation where they can never correct their bias after "learning helplessness" no matter the amount of fair and normally distributed data was input again to train the neural network? A similar context also applies to "learned arrogance" where nodes overweight themselves permanently after exposing to a series of hotstreaks.

$\endgroup$
1
  • 1
    $\begingroup$ “I was under an impression that artificial intelligence is modelled after organic intelligence.” This angle is usually promulgated by marketing departments or pop sci blogs. $\endgroup$ Jun 21, 2021 at 4:07

2 Answers 2

1
$\begingroup$

I'm not sure these analogies to psychology / the human brain are really that useful or accurate. It's not like a neural network is truly affected by psychological effects, it's simply several bits of matrix algebra chained together. However, you certainly can observe the following:

  • Dead "neurons": You can end up with weights in a neural network that make it so that the neuron never "activates" again and then no matter what training data you throw at your model, those particular neurons will never do anything, again. This could happen for many reasons, but fitting too aggressively to extreme data could be one scenario. Even if nothing much goes wrong during training, there will typically be some to which this has happened. However, unless you really get to this "dead" state, I believe you should normally be able to recover other behaviors, again given enough training data.
  • Overfitting: You can certainly end up with extremely overfit (to the training data) neural networks that would exhibit poor behaviour (e.g. bias etc.), where it could be really hard to recover a more generalizable model by training more (in contrast, when you are underfit, it is often more realistic to train more and still get something "good"). I'd imagine you could manage to get there by training on a subset of the data, if that is then still a local minimum of the model. You could also try, very hard, to find a really narrow optimum in the loss landscape (broader ones tend to be the ones you really want for good generalization), which might exhibit poor behaviours.

However, an non-overfit neural network should generally with more training data be able to get to a decent performance, again, unless it is somehow caught in a really unfortunate local minimum. Assuming that's not the case, then how long it takes to correct issues induced by early training will depend on a few things such as

  • In how bad a place did you end up? The closer we are too good weights and the clearer the path towards better weights is (ideally a steep gradient in an unchanging direction for all weights), the faster and easier this should be.
  • How high is your learning rate influences how fast weights change. Ideally, you get this just right: not so fast that the optimization goes off on a tangent, not so slow that it takes forever to improve.
  • How clear are the signals in the data? How high dimensional is the problem? Is the model inducing useful inductive biases for the problem (e.g. convolutional neural networks are quite a nice fit for vision problems)?
$\endgroup$
2
  • $\begingroup$ Thank you, it answers the basic sense of situations where neurons "underweight" or "overweight" themselves. That said, could a neuron ever get a signal that they may be overfitting certain variables too hard or that they are on the verge of becoming a dead/pruned neuron and perhaps go through a new adjustment? Seems like as soon as a neuron gets itself into either situation they are past the point of no return. I also don't know if this is even a good or a bad thing. $\endgroup$
    – David LE
    Jun 21, 2021 at 19:57
  • $\begingroup$ Well, there's the option of re-initializing dead neurons (see e.g. this paper: arxiv.org/abs/1812.00481). $\endgroup$
    – Björn
    Jun 21, 2021 at 20:20
1
$\begingroup$

AI is not necessary modeled after organic intelligence. There are subfields within AI and computational neuroscience or cognitive science, that are interested in creating biologically plausible algorihms, but this is more of an exception than a rule. In some fields, people create computational models of various behaviours or cognitive processes in order to study these processes, but not for the sake of improving AI. In this case it is possible that some group have computational models of learned helplessness.

"Learned helplessness" of a node is not the same as learned helplessness of an individual and I don't thing these effect would be related.

What would a not do with respect to other nodes depends on an learning algorithm. In a Hebbian style "wire together, fire together" learning, if a node fails to produce results relative to other nodes, it would get deweighted or maybe even pruned out of a network. In a more traditional gradient based optimization, a node will be adjusted based on it's error gradient regardless of what other nodes in a layer do. There might be a regularization parameter imposed on the layer which might take into account activations of other neurons, but this might also have an effect in either direction.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.