深度睡眠外文文献 Deep dreaming, aberrant salience and psychosis_ Connecting the dots by artificial neural networks_第1页
深度睡眠外文文献 Deep dreaming, aberrant salience and psychosis_ Connecting the dots by artificial neural networks_第2页
深度睡眠外文文献 Deep dreaming, aberrant salience and psychosis_ Connecting the dots by artificial neural networks_第3页
深度睡眠外文文献 Deep dreaming, aberrant salience and psychosis_ Connecting the dots by artificial neural networks_第4页
全文预览已结束

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

Deep dreaming aberrant salience and psychosis Connecting the dots by artifi cial neural networks Matcheri S Keshavan a b Mukund Sudarshanc a BethIsrael Deaconess Medical Center United States b Harvard Medical School United States c Cornell University United States a b s t r a c ta r t i c l ei n f o Article history Received 23 October 2016 Received in revised form 8 January 2017 Accepted 8 January 2017 Available online 24 January 2017 Whysomeindividuals whenpresentedwithunstructuredsensoryinputs developalteredperceptionsnotbased inreality isnotwellunderstood Machinelearningapproaches canpotentiallyhelpus understandhow thebrain normally interprets sensory inputs Artifi cial neural networks ANN progressively extract higher and higher level features of sensory input and identify the nature of an object based on a priori information However some ANNs which use algorithms such as the deep dreaming developed by Google allow the network to over emphasize some objects it thinks it recognizes in those areas and iteratively enhance such outputs lead ingtorepresentationsthatappearfartherandfartherfrom reality Wesuggestthatsuch deepdreaming ANNs may model aberrant salience a mechanism suggested for pathogenesis of psychosis Such models can generate testable predictions for psychosis 2017 Elsevier B V All rights reserved Keywords Psychosis Deep dreaming Artifi cial neural networks Schizophrenia 1 Introduction The connections between creativity madness and dreams is well known Benson and Park 2013 but defy explanation Given the fact that all of these originate from the brain the logical place to connect these dots is by looking at how our neural networks interact with each other to produce our experiences and beliefs This line of thinking led me and Mukund Sudarshan a computer science undergraduate in my lab to glean some interesting yet speculative insights using princi ples of machine learning Mukund applied the Google s deep dream ing algorithm an artifi cial neural network program Szegedy et al 2014 toalandscapepaintingofmine Fig 1 top Tantalizingly there alistic picture of lakes and mountains transformed to a Dali esque dream like picture with hidden animals lamps and other obscure ob jects Fig 1 bottom Given that I did not ask for this particular surreal ismeffectbutthecomputergenerateditonitsown wewonderedwhat might have happened behind the black box of this computer based artifi cial neural network ANN program 2 How artifi cial neural networks learn Current thinking in the fi eld of machine learning might help us un derstand how the brain normally interprets sensory inputs Artifi cial neural networks ANN are seen as a series of sieves with progressively small holes stacked on one another these sieves progressively strain higher and higher level more and more fi ne grained features of sen sory input Fig 2 In a neural network that has been trained to identify a cat for example thelower input layers interpret basic or coarse fea tures like edges or corners like the outline of a face and the interme diate layers look for overall shapes The fi nal layers then combine the detailstoassembletheseintoan answer sothattheimagebestdepicts a cat Thus after training each layer progressively extracts higher and higher level features of the image until the output layer essentially makes a decision on what the image shows The fi gure below is a con ceptual representation of a neural network Essentially the image is processed in overlapping pieces The fi rst layer analyzes each piece of the image convolution layer The subsequent layer pooling layer pools together the knowledge gained from pieces that are close to one another and creates a more dense representation the smaller the square in the image the more dense the representation This process repeats until we have a dense representation that captures the absolute essence of the image rather than minute details This fi nal representa tion is used to determine what the image actually contains In this case the image contains the realistic representation of a cat 3 Deep dreaming and aberrant salience Typically wefeedaninputintothenetworkandaskittoidentifythe image based on whatever it has seen before If all is well the neural network outputs the correct label for the image However this deep Schizophrenia Research 188 2017 178 181 Corresponding author at Room 610 Massachusetts Mental Health Center 75 Fenwood Road Boston MA 02115 United States E mail address mkeshava bidmc harvard edu M S Keshavan http dx doi org 10 1016 j schres 2017 01 020 0920 9964 2017 Elsevier B V All rights reserved Contents lists available at ScienceDirect Schizophrenia Research journal homepage dreaming neural network allows the network to take parts of the image and over emphasize some objects it thinks it recognizes in those areas Whatever this neural network thinks it sees is then shown in the image For example if the neural network thinks the ear on a cat face is a butterfl y wing it allows it to draw a butterfl y wing on top of the eyebrow This modifi ed image is fed back into the neural network and this process repeats After numerous iterations we end up with the painting on the right with unexpected hidden images that were gradually brought out in the image Thus the network has wan deredofffarfrom reality bypayingtoomuchattentiontolessrelevant areas We see what the neural network thinks exists in the inputted image By observing the output of each step in this iterative process we can see the transition of the neural network s dream from a normal imageintooneoffantasy Suchaphenomenonisreminiscentofthepro gressivechangesintheartisticdepictionofcatsbythefamous19thcen tury English painter Louis Wain McGennis 1999 4 Is psychosis related to deep dreaming neural networks The above model can potentially help us understand psychotic phe nomena We suggest that an aberrant enhancement of salience of in coming information in selected layers of ANNs may create unexpected and internally generated informational representations these are not just based on external inputs and thus not matched to reality akin to what is seen in psychosis Indeed Kapur 2003 has proposed that psychosis may be related to aberrantassignmentofsaliencetotheelementsofone sexperience Ata neural network level it has been suggested that such aberrant salience may be related to an abnormal neural activity in the ventral striatum and a failure of feedback regulation by prefrontal structures in individ uals with schizophrenia and those at risk for this illness Roiser et al 2013 Thus our response to sensory inputs is regulated by feedback systems whereby higher brain regions may modulate assignment of Fig 1 Application of a deep dreaming algorithm to a painting by the fi rst author Top before applying the algorithm bottom after applying the algorithm 179M S Keshavan M Sudarshan Schizophrenia Research 188 2017 178 181 salience and normally prevent a runaway assignment of salience A derailment of this feedback loop may result in psychosis In a non psy choticnetwork theinternalrepresentationoftheinputisnotdrastically affectedand therefore if this internal representationis fed backinto the neural network the classifi cation will most likely not change This is shown in the fi gure below Clearly investigating the functioning of artifi cial neural networks can potentially yield insights about the mysterious mechanisms that underlie psychosis It remains unclear why ANNs after being fed large amounts of unstructured data generate certain unexpected dream like over representations of previously learned features via a positive feedback loop One possible explanation is in the aberrant neural net work the signal to noise ratio is lowered due to the increased impor tance placed on background noise This is validated by a recent resting state fMRI study by us Hager et al 2016 which found that probands tend to exhibit increased brain randomness when compared to healthy controls Similarly even when we input completely random noise into the neural network it still manages to produce hallucinations by an notating objects it thinks it sees that clearly do not exist However these computer models offer the possibility of adjusting the network parameters that can allow testing specifi c predictions First functional neuroimaging data and analyses of functional connectivity can poten tially help make predictions about creativity or psychosis like neural representations by looking at patterns of neuronal activations when attempting to recognize patterns Second we can introduce regulation components to modulate deep dreams in a deep dreaming network model reducingthepositivefeedbackandintroducingareciprocalneg ativefeedbackloop likewhatexistsinthefrontolimbicpathwaysofthe humannervoussystem couldservetoreducepsychosis likewhatcog nitive behavior therapy does Goldin et al 2013 and removing such negative feedback could serve to accentuate psychosis or anxiety in the neural network Finally the therapeutic implications generated by these ANN models could then be tested in the clinic As an example neuromodulating interventions like TMS and TDCS could potentially inhibit subcortical neural circuit activity which otherwise generate hal lucinations and delusions While attemptingtodevelopa neuralnetwork model of schizophre nia itisimportanttobearinmindthatauditoryhallucinationsaremore characteristic of schizophrenia than are visual hallucinations Indeed the proposed model of aberrant ANNs might better explain drug in duced hallucinations such as those caused by LSD and psilocybin We chose to use visual hallucinations only as an example However we can also consider thoughts just like visual images as inputs and inter pretations of thoughts as the response Normal ANNs might generate consistent interpretations while an aberrant ANN might amplify some parts of thought inputs leading to delusions 5 Aberrant ANNs dreams and creativity Can ANNs help us understand the mysteries of why we dream and can the applications of ANNs offer us clues as to what underlies creativ ity Todd 1989 While a detailed exposition of these questions is be yond the scope of this paper we will briefl y touch upon them here Mamelak and Hobson 1989 have suggested that the bizarre content of dream content may be due to uninhibited neural network responses to pedunculopontine occipital input due to failure of aminergic modu lationbytheforebrain Ithasindeedbeensuggestedthatdreamcontent may represent the brain s effort to elaborate on recent memories via meaningful and novel associations with remote memories which are emotionally salient Llewellyn 2013 The neural underpinnings of someformsofcreativitymaybesimilar aswehaveseenwiththeexam ple of Louis Wain s cats In an fMRI study it was shown that metaphor ical interpretations of word pairs is determined by salience of the linguistic expressions rather than their literal meanings and this is de termined by selective activations in right hemispheric regions such as thesuperior temporal sulcus Mashal et al 2007 Thus theneural net work alterations that underlie aberrant salience in psychosis may also explain the relation between creativity and psychosis Interestingly Fig 2 Diagram illustrating how artifi cial neural networks may lead to either normal realistic representations dreams or psychotic states 180M S Keshavan M Sudarshan Schizophrenia Research 188 2017 178 181 when comparing healthy controls to individuals with schizophrenia those in the schizophrenia group enhanced mental imagery manipula tion Benson and Park 2013 Insummary studiesofartifi cialneuralnetworkcanilluminatemany aspects of brain function in health and disease as well as dreams and creativity Increasing understanding of the brain s circuit dynamics andconnectomicalterationsinneuropsychiatricdisordersmakestudies of neural network models of these illnesses increasingly timely Funding source None Confl icts None Acknowledgements None References Benson T L Park S 2013 Exceptional visuospatial imagery in schizophrenia implica tions for madness and creativity Front Hum Neurosci 7 756 1 11 Goldin P R Ziv M Jazaieri H Hahn K Heimberg R Gross J J 2013 Impactof cognitive behavioral therapy for social anxiety disorder on the neural dynamics of cognitive re appraisal of negative self beliefs randomized clinical trial JAMA Psychiatry 70 10 1048 1056 Oct 2013 Hager B Yang A C Brady R Meda S Cleme

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论