Neuroadaptive Bayesian Optimization Implications for the Cognitive - PowerPoint PPT Presentation

Neuroadaptive Bayesian Optimization Implications for the Cognitive Sciences Romy Lorenz Postdoctoral Research Fellow Cognitive, Clinical and Computational Neuroimaging Lab Imperial College London NIPS Workshop 2017

Overview Overview 1. Motivation 2. The framework 3. Validation study 4. Application 1: Human brain mapping 5. Application 2: Brain stimulation 6. Ongoing work 7. Implications & Discussion NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Overview Overview 1. Motivation 2. The framework 3. Validation study 4. Application 1: Human brain mapping 5. Application 2: Brain stimulation 6. Ongoing work 7. Implications & Discussion NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Reproducibility crisis Survey by Nature 2016 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Reproducibility crisis Disciplines most profoundly affected: Ioannidis et et al. TiCS 2014, Head et al. PLOS Biol . 2015, Szucs & Ioannidis PLOS Biol. • Biomedical sciences 2017,…Ioannidis et et al. TiCS 2014, Head et al. PLOS Biol . 2015, Szucs & Ioannidis PLOS Biol. • Psychology 2017,… • Cognitive Sciences Ioannidis et et al. TiCS 2014, Head et al. PLOS Biol . 2015, Szucs & Ioannidis PLOS Biol. 2017,… Survey by Nature 2016 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Reproducibility crisis in Cognitive Sciences • Cognitive biases - IKEA-effect - Texas sharp-shooter effect • Bad research practices - P-hacking - HARKing - File-drawer effect • Limitations of methodology - Underpowered studies - “Narrow” experimental designs NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Reproducibility crisis in Cognitive Sciences • Cognitive biases - IKEA-effect - Texas sharp-shooter effect • Bad research practices - P-hacking - HARKing - File-drawer effect • Limitations of methodology - Underpowered studies - “Narrow” experimental designs NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Aims of cognitive neuroscience Research questions Standard approach What are the fundamental aspects of cognition? select task What are the fundamental roles of distinct networks in the brain? > 20 test How can cognitive processes be modulated or enhanced? group-level inference broad narrow NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Aims of cognitive neuroscience Human-brain mapping Standard approach • Over-specified inferences about functional-anatomical mappings • Inflated test statistics select task (Westfall et al. Wellcome Open Research 2017) Biomarker discovery > 20 • Which exact task conditions will be test sensitive to certain patient group? (Sprooten et al. Human Brain Mapping 2017) group-level Non-invasive brain stimulation inference • Many free parameters, confusion surrounding efficacy broad narrow NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Overview Overview 1. Motivation 2. The framework 3. Validation study 4. Application 1: Human brain mapping 5. Application 2: Brain stimulation 6. Ongoing work 7. Implications & Discussion NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

The framework The framework Lorenz et al. Trends in Cognitive Sciences 2017 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

“The Automatic Neuroscientist” The framework Lorenz et al. NeuroImage 2016 | Neuroadaptive Bayesian optimization. (1) The subject is presented with an experimenta NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Overview Overview 1. Motivation 2. The framework 3. Validation study 4. Application 1: Human brain mapping 5. Application 2: Brain stimulation 6. Ongoing work 7. Implications & Discussion NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Target brain state lateral occipital cortex activity é superior temporal cortex activity ê masks derived from Braga et al. NeuroImage 2013 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Experiment space Optimal stimuli for target brain state NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Bayesian optimization Bayesian optimization Gaussian Propose new process (GP) “update stimuli knowledge” regression combination “close the loop” Rasmussen & Williams 2006 Brochu et al. arXiv 2010 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Bayesian optimization Bayesian optimization Gaussian process (GP) “update knowledge” regression squared exponential kernel: ! ! , ! = ! ! exp − ! − ! ! choice of ! 2 ! ! ! covariance function audio-visual stimulus ! , ! ∈ ℝ ! ! ! ! ∈ ℝ ! variance of covariance kernel length of covariance kernel ! ∈ ℝ ! Rasmussen & Williams 2006 Brochu et al. arXiv 2010 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Bayesian optimization Bayesian optimization Gaussian Propose new process (GP) stimuli Expected improvement acquisition function: regression combination !" ! = ! ! − ! ! !"# q ! + !"# ( ! ) ! ( ! ) ! m(x) : predicted mean predicted variance var(x) : choice of choice of maximum predicted value f max : acquisition function covariance function cumulative distribution function q() : probability density function (squared exponential, p() : linear, periodic kernel) “close the ! = ! ! − ! ! !"# ! loop” !"# ( ! ) Rasmussen & Williams 2006 Brochu et al. arXiv 2010 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Results 5 Lorenz et al. NeuroImage 2016 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Overview Overview 1. Motivation 2. The framework 3. Validation study 4. Application 1: Human brain mapping 5. Application 2: Brain stimulation 6. Ongoing work 7. Implications & Discussion NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Motivation Maths task Go/No-Go task N-back task Divided attention task Stroop task frontoparietal networks (FPNs) Duncan & Owen TiNS 2000 Fedorenko et al. PNAS 2013 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Motivation Maths task Go/No-Go task N-back task Divided attention task Stroop task frontoparietal networks (FPNs) Hampshire et al. Neuron 2012 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Motivation Maths task Go/No-Go task N-back task Divided attention task Stroop task Standard fMRI approach • Limited generalizability • Limited reproducibility Lorenz et al. Trends in Cognitive Sciences 2017 Westfall et al. Wellcome Open Research 2017 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Searching across cognitive tasks NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Task space based on meta-analysis Study 1 Study 2 Study 3 Stage 1 Stage 2 Yeo et al. Cerebral Cortex 2015 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Find optimal tasks Study 1 Study 2 Study 3 Stage 1 Stage 2 Tower of London & Deductive Reasoning tasks maximally dissociate FPNs NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Zoom in task space and fine-tune tasks Study 1 Study 2 Study 3 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Find optimal task parameters Study 1 Study 2 Study 3 NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Find unique functional profile Study 1 Study 2 Study 3 ventral FPN > 3 other FPNs dorsal FPN > 3 other FPNs Tower of London, Deductive Reasoning, Encoding Go/No-Go, Divided Auditory Attention, & Wisconsin Card Sorting tasks Passive Listening & Reading tasks NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Summary • High inter-subject reliability • Functional profile across many tasks is unique to each FPN • Set of optimal tasks only partially corresponds to meta- analysis and previous functional labels • Neurally-derived cognitive taxonomy needed • Powerful synergy between neuroadaptive Bayesian optimization and meta-analyses Lorenz et al. under revision (bioRxiv: 128678 ) NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Overview Overview 1. Motivation 2. The framework 3. Validation study 4. Application 1: Human brain mapping 5. Application 2: Brain stimulation 6. Ongoing work 7. Implications & Discussion NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Transcranial alternating current stimulation (tACS) § Status Quo - Ad hoc definition of frequency Ines Violante and phase - Cohort testing § Limitation 1. How to choose frequency and phase? 2. Stimulation parameters may vary due to anatomy or pathology NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz

Concurrent real-time fMRI/tACS Ines Violante Lorenz et al. PRNI 2016 Lorenz et al. in preparation NIPS Workshop 2017 Neuroadaptive Bayesian optimization Romy Lorenz