4
$\begingroup$

Today, we can measure EEG (or fMRI) in different individuals and use machine learning to predict their thinking. I want to know if the exact neural patterns (fMRI, EGG etc.) are still similar across different individuals? (or maybe I should use "cross subjects")

For example
I trained a model for Alice. When she sees a certain color, the model will infer the corresponding color based on her fMRi signals (or EGG, whatever). (You can try to use a number of samples, not only for Alice, but you must keep sure your model does work on all the "Alices")
Now, can I just use the model for Bob and also ensure some accuracy? Or I have to train a totally new model base on his fMRI measurements?
What about shape, position, phoneme, letter or even words?

If the exact neural patterns vary in from individual to individual, can we say "We just use the same method to recognize, but the specific neural links are learned independently after birth". Like a deep neural network: the frame and hyperparameters of the model are given, but even the same model with the same training task will have the totally different parameters?
Or we just use a very general way to "see" the world, all the fMRI/EEG results can be transferred to each individual (like for brain computer interface)?

$\endgroup$

1 Answer 1

1
$\begingroup$

It's much more complicated than you conceived even for a single subject since there're possibly numerous overlapping distributed neural networks at play located differentially via developmental and postnatal experiential selection by different environment for different subject even for some mundane tasks, an then sometimes neural circuits for the same task will get copied, consolidated and reorganized. See a reference supporting such neural selectionism and Darwinism as Fernando et al's Copying and Evolution of Neuronal Topology.

The theories of neural Darwinism [12] and neuronal selectionism [13], [14] propose that a primary repertoire of neuronal groups within the brain compete with each other for stimulus and reward resources. This results in selection of a secondary repertoire of behaviourally proficient groups [15]... Importantly, the loops between the medial temporal cortex (containing the hippocampus) and the neocortex have been implicated in memory consolidation and reconsolidation, processes that involve gradual reorganization of circuits... Consolidation has been supported by experimental evidence demonstrating that the anterior cingulated cortex is involved in the remote memory for contextual fear conditioning which is a hippocampus-dependent task [113], [114].

Even worse, for your specification to predict a subject's thinking (essentially mind reading) we need to pin down the unique propositional content of a certain neural circuit if any such translation map ever exits, but this seems impossible biologically from latest neuroscience results. For an example reference see Rosenberg's Eliminativism without Tears.

The paper begins by showing why, given the findings of neuroscience,brain states don’t have propositional content. It then examines a leading attempt to attribute content to brain states owing to their functional, i.e. evolutionary role, and show why it is best viewed as a modus tollens argument against brain states having content... Differences in the number, location, and wiring of individual neural circuits can only turn them from small sets of input/output systems into larger ones. It can’t turn them from one kind of thing—the stimulus/response wiring of a sea slug into an entirely different kind of thing, stored sentential content in the neurons.

Also in the same reference there's another "disjunctive indeterminacy problem" for any neural circuitry adaptively selected (against) from nature.

It’s now famous that there is no way any teleosemantic theory can tell whether the content of the relevant frog’s neural circuit is “Fly or black moving dot at x,y,z,t,” or “fly or bee bee at x,y,z,t.” or any of a zillion other disjunctive objects of thought, so long as none of these disjuncts has ever actually been presented to the fly. Whence the name, “disjunction problem.”... Here is Darwinian theory’s disjunction problem: the process Darwin discovered can’t tell the difference between these two genes or their traits until cross-over breaks the linkage between one gene, that is going to increase its frequency, and the other one, that is going to decrease its frequency. If they are never separated, it will remain blind to their differences forever... Natural selection will have an even harder time discriminating these two traits.

Having said above, there may be some structural similarity between different subject's neural circuitry for a same input stimuli hypothesized as structural resemblance theory according to the same reference.

In the 1980’s, experiments with macaque monkey, have isolated the structural resemblance between input vibrations the finger feels, measured in cycles per second, and representations of them in neural circuits, measured in action-potential spikes per second [Mountcastle, Steinmetz, and Romo, 1990]. This resemblance between two easily measured variables makes it unsurprising that they would be among the first such structural resemblances to be discovered... It is obvious how structural resemblance theory lends itself to theories of information as causal covariance [sensu Dretske, 1981], and theories that accord neutral circuitry the function of storing such information. The causally covariance of neural circuitry with any of its prior causes and future effects, will include inputs that make its outputs environmentally appropriate and so accord the neural circuitry adaptational functions. Note, it can do all this without these neural circuits having propositional content.


Reference:

Chrisantha Fernando, K. K. Karishma, and Eörs Szathmáry. "Copying and Evolution of Neuronal Topology". PLoS One. 2008; 3(11): e3775.doi: 10.1371/journal.pone.0003775.

Rosenberg, Alex. "Eliminativism without Tears". S2CID 17491588.

$\endgroup$
2
  • 1
    $\begingroup$ Thanks for the well referenced answer. While citing text is fine when properly referenced, please keep it to a minimum. At the very least extract the most relevant information only and always try to formulate answers in your on words. As of now you've copied large swaths of text that no one will likely read. Folks want answers, not textbooks. $\endgroup$
    – AliceD
    Nov 15, 2022 at 7:54
  • $\begingroup$ @AliceD thanks for your comment. I've shortened referenced texts and hopefully this helps. $\endgroup$
    – cinch
    Nov 15, 2022 at 20:18

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.