Understanding information processing in human brain by interpreting machine learning models. A data-driven approach to computational neuroscience
Kuupäev
2020-07-09
Autorid
Ajakirja pealkiri
Ajakirja ISSN
Köite pealkiri
Kirjastaja
Abstrakt
Modelleerimine on inimkonna põline viis keerulistest nähtustest arusaamiseks. Planeetide liikumise mudel, gravitatsiooni mudel ja osakestefüüsika standardmudel on näited selle lähenemise edukusest. Neuroteaduses on olemas kaks viisi mudelite loomiseks: traditsiooniline hüpoteesipõhine lähenemine, mille puhul kõigepealt mudel sõnastatakse ja alles siis valideeritakse andmete peal; ja uuem andmepõhine lähenemine, mis toetub masinõpele, et sõnastada mudeleid automaatselt.
Hüpoteesipõhine viis annab täieliku mõistmise sellest, kuidas mudel töötab, aga nõuab aega, kuna iga hüpotees peab olema sõnastatud ja valideeritud käsitsi. Andmepõhine lähenemine toetub ainult andmetele ja arvutuslikele ressurssidele mudelite otsimisel, aga ei seleta kuidas täpselt mudel jõuab oma tulemusteni. Me väidame, et neuroandmestike suur hulk ja nende mahu kiire kasv nõuab andmepõhise lähenemise laiemat kasutuselevõttu neuroteaduses, nihkes uurija rolli mudelite tööprintsiipide tõlgendamisele.
Doktoritöö koosneb kolmest näitest neuroteaduse teadmisi avastamisest masinõppe tõlgendamismeetodeid kasutades. Esimeses uuringus tõlgendatava mudeli abiga me kirjeldame millised ajas muutuvad sageduskomponendid iseloomustavad inimese ajusignaali visuaalsete objektide tuvastamise ülesande puhul. Teises uuringus võrdleme omavahel signaale inimese aju ventraalses piirkonnas ja konvolutsiooniliste tehisnärvivõrkude aktivatsioone erinevates kihtides. Säärane võrdlus võimaldas meil kinnitada hüpoteesi, et mõlemad süsteemid kasutavad hierarhilist struktuuri. Viimane näide kasutab topoloogiat säilitavat mõõtmelisuse vähendamise ja visualiseerimise meetodit, et näha, millised ajusignaalid ja mõtteseisundid on üksteisele sarnased.
Viimased tulemused masinõppes ja tehisintellektis näitasid et mõned mehhanismid meie ajus on sarnased mehhanismidega, milleni jõuavad õppimise käigus masinõppe algoritmid. Oma tööga me rõhutame masinõppe mudelite tõlgendamise tähtsust selliste mehhanismide avastamiseks.
Building a model of a complex phenomenon is an ancient way of gaining knowledge and understanding of the reality around us. Models of planetary motion, gravity, particle physics are examples of this approach. In neuroscience, there are two ways of coming up with explanations of reality: a traditional hypothesis-driven approach, where a model is first formulated and then tested using the data, and a more recent data-driven approach, that relies on machine learning to generate models automatically. Hypothesis-driven approach provides full understanding of the model, but is time-consuming as each model has to be conceived and tested manually. Data-driven approach requires only the data and computational resources to sift through potential models, saving time, but leaving the resulting model itself to be a black box. Given the growing amount of neural data, we argue in favor of a more widespread adoption of the data-driven approach, reallocating part of the human effort from manual modeling. The thesis is based on three examples of how interpretation of machine-learned models leads to neuroscientific insights on three different levels of neural organization. Our first interpretable model is used to characterize neural dynamics of localized neural activity during the task of visual perceptual categorization. Next, we compare the activity of human visual system with the activity of a convolutional neural network, revealing explanations about the functional organization of human visual cortex. Lastly, we use dimensionality reduction and visualization techniques to understand relative organization of mental concepts within a subject's mental state space and apply it in the context of brain-computer interfaces. Recent results in neuroscience and AI show similarities between the mechanisms of both systems. This fact endorses the relevance of our approach: interpreting the mechanisms employed by machine learning models can shed light on the mechanisms employed by our brain
Building a model of a complex phenomenon is an ancient way of gaining knowledge and understanding of the reality around us. Models of planetary motion, gravity, particle physics are examples of this approach. In neuroscience, there are two ways of coming up with explanations of reality: a traditional hypothesis-driven approach, where a model is first formulated and then tested using the data, and a more recent data-driven approach, that relies on machine learning to generate models automatically. Hypothesis-driven approach provides full understanding of the model, but is time-consuming as each model has to be conceived and tested manually. Data-driven approach requires only the data and computational resources to sift through potential models, saving time, but leaving the resulting model itself to be a black box. Given the growing amount of neural data, we argue in favor of a more widespread adoption of the data-driven approach, reallocating part of the human effort from manual modeling. The thesis is based on three examples of how interpretation of machine-learned models leads to neuroscientific insights on three different levels of neural organization. Our first interpretable model is used to characterize neural dynamics of localized neural activity during the task of visual perceptual categorization. Next, we compare the activity of human visual system with the activity of a convolutional neural network, revealing explanations about the functional organization of human visual cortex. Lastly, we use dimensionality reduction and visualization techniques to understand relative organization of mental concepts within a subject's mental state space and apply it in the context of brain-computer interfaces. Recent results in neuroscience and AI show similarities between the mechanisms of both systems. This fact endorses the relevance of our approach: interpreting the mechanisms employed by machine learning models can shed light on the mechanisms employed by our brain
Kirjeldus
Märksõnad
automatic learning, neurosciences