3 ms·

You may be joking, but I'm tired that everytime neural networks comes up this unfortunate conflation of different aspects of AI occurs; one based in accurate bi

You may be joking, but I'm tired that everytime neural networks comes up this unfortunate conflation of different aspects of AI occurs; one based in accurate biological modeling, the other based in tractable statistical methods. Whereas it is true that both are inspired by neural activation and that both can produce a complex, yet coherent mapping from input to output, this is like saying that birds and aircraft are similar because both fly and have wings.

quotemstr 8y ago
Yet both birds and aircraft can stall. (Granted, birds can get out of stalls much more easily than aircraft can.) While many analogies between natural neural systems and various ML arrangements are false, not all of them are.
- noobermin 8y ago
  The analogy is like "both a human and a baseball launcher can throw a ball, so they must be the same." People really need to let go of the terminology around machine learning that leads people to really believing that it's close to actual cognition.
  - pas 8y ago
    But what is cognition? People should reevaluate their views on how it works. The brain is very complex and a very big ball of highly specialized circuitry (and the software running on those and managing various aspects and parts of it and them). The similarities are at the same time very striking but the differences are just as drastic in number of layers and other size related parameters our various brain components have. Yes, AlphaGoZero is not going to learn to cook and sing and dance, but it beats our gaming component in a lot of areas. Of course a synapse is not really a memristor or a ReLU, but at the same time not that far off either. Similarly a biological neuron is not just a simple backpropagatin integrator like a perceptron, but it works very similarly. And just as we don't really know how all the representations work in an ANN, we don't exactly know the biological aspects of memory/learning/seeing work in our brains, yet we have made enormous progress on both. See all the demos/visualizations on how various layers encode in deep nets and look at the data about the brain's visual cortex and the V1 circuit, look at the gene spliced mouse studies where memory encoding is studied (and sometimes only one neuron encodes for a memory/face/concept - just as with deep nets).
- jhanschoo 8y ago
  Indeed, they have similar characteristics in certain respects, and sometimes ML is inspired by neurobiology. I suppose my exasperation is primarily directed at non-engineers/researchers hyping AI; to abuse the analogy, sometimes these futurist writers seem to think that just because both can fly, aircraft can one day lay eggs too.

quotemstr 8y ago

Yet both birds and aircraft can stall. (Granted, birds can get out of stalls much more easily than aircraft can.) While many analogies between natural neural systems and various ML arrangements are false, not all of them are.

noobermin 8y ago
The analogy is like "both a human and a baseball launcher can throw a ball, so they must be the same." People really need to let go of the terminology around machine learning that leads people to really believing that it's close to actual cognition.
- pas 8y ago
  But what is cognition? People should reevaluate their views on how it works. The brain is very complex and a very big ball of highly specialized circuitry (and the software running on those and managing various aspects and parts of it and them). The similarities are at the same time very striking but the differences are just as drastic in number of layers and other size related parameters our various brain components have. Yes, AlphaGoZero is not going to learn to cook and sing and dance, but it beats our gaming component in a lot of areas. Of course a synapse is not really a memristor or a ReLU, but at the same time not that far off either. Similarly a biological neuron is not just a simple backpropagatin integrator like a perceptron, but it works very similarly. And just as we don't really know how all the representations work in an ANN, we don't exactly know the biological aspects of memory/learning/seeing work in our brains, yet we have made enormous progress on both. See all the demos/visualizations on how various layers encode in deep nets and look at the data about the brain's visual cortex and the V1 circuit, look at the gene spliced mouse studies where memory encoding is studied (and sometimes only one neuron encodes for a memory/face/concept - just as with deep nets).
jhanschoo 8y ago
Indeed, they have similar characteristics in certain respects, and sometimes ML is inspired by neurobiology. I suppose my exasperation is primarily directed at non-engineers/researchers hyping AI; to abuse the analogy, sometimes these futurist writers seem to think that just because both can fly, aircraft can one day lay eggs too.