> Strikingly, the brain atlases were similar for all the participants, suggesting that their brains organised the meanings of words in the same way. The scientists only scanned five men and two women, however. All are native English speakers, and two are authors of the study published in Nature. It is highly possible that people from different backgrounds and cultures will have different semantic brain atlases.
Would love to see this study performed on a much larger, diverse, group, to see what similarities and differences there are. Also, same person at different ages.
This would definitely be very interesting in regards to language study. Repeating the experiment with people which are actively in the progress of learning a new language every few weeks could give some remarkable insight in how the new words are being mapped onto the brain.
Interesting to compare this semantic map with the semantic maps that Mikolov's Word2vec processes produce and Word2vec's resultant semantic vector math (King - Man + Woman ~= Queen).
Perhaps Hinton's Vectors of Thought do have a wetware analogy.
In terms of (Artificial) Neural Networks it implies their internal representation is functional not merely symbolic.
Soumith has shown this applies to visual modalities and smile vectors or wearing sunglasses vectors are similarly present[1].
This types of vector algebra has some use in translating between languages and modalities.
Karpathy[2] has demonstrated an internal vector of a descriptive sentence can 'plugged into' an image recognising network and used to find vectors representing pictures.
Perhaps a sentence vector could even generate a picture.
What Neural Nets do internally is still mysterious.
(A Highly speculative example) If DeepGo's internal vectors could translate to the modality of plain English we might glimpse its 'mind' at work.
What this means for brains is anyones guess, I find inspiration in Geoff Hinton's guesses[3].
The brain evolved, and neural nets are trained, to optimize for high-density, "low-cost" information storage. So it would make sense if there is a structural correlation between the two in terms of how information is represented. I suspect that more biologically realistic neural nets would lead to more correlation.
11 comments
[ 29.9 ms ] story [ 987 ms ] thread> Strikingly, the brain atlases were similar for all the participants, suggesting that their brains organised the meanings of words in the same way. The scientists only scanned five men and two women, however. All are native English speakers, and two are authors of the study published in Nature. It is highly possible that people from different backgrounds and cultures will have different semantic brain atlases.
Would love to see this study performed on a much larger, diverse, group, to see what similarities and differences there are. Also, same person at different ages.
Perhaps Hinton's Vectors of Thought do have a wetware analogy.
Soumith has shown this applies to visual modalities and smile vectors or wearing sunglasses vectors are similarly present[1].
This types of vector algebra has some use in translating between languages and modalities.
Karpathy[2] has demonstrated an internal vector of a descriptive sentence can 'plugged into' an image recognising network and used to find vectors representing pictures.
Perhaps a sentence vector could even generate a picture.
What Neural Nets do internally is still mysterious.
(A Highly speculative example) If DeepGo's internal vectors could translate to the modality of plain English we might glimpse its 'mind' at work.
What this means for brains is anyones guess, I find inspiration in Geoff Hinton's guesses[3].
[1] - searching
[2] https://www.youtube.com/watch?v=ZkY7fAoaNcg
[3] http://www.computing.co.uk/ctg/news/2409871/document-search-...
Gonna read and watch your links and do some thinking of my own, thanks.