Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
References
Funding
Rodent Art Credit: Copyright (c) 2015 Etienne Ackermann
[A],[B],[C]
Synchronous firing of neurons in CA3
150-250 Hz Oscillations in CA1
Hippocampus is the center for learning and memory
What Are Sharp-Wave Ripples?
We apply machine learning techniques towards improvinga realtime closed loop sharp-wave ripple (SWR) disruptionsystem.
Memory consolidation and recallSequential reactivation based on a previousexperience
Why Do We Care?Never having interrogated SWR generation, it is likelycurrent disruption experiments do not entirely preventaccess to specific memories. Thus, it is impossible todetermine whether or not SWRs are responsible for memoryformation, or simply are a large contributing factor.Additionally, accurate prediction of SWR events may notonly allow for improve SWR detection, but also may enableselective, premature disruption of specific memories. Thismethodology, then, may be effective in terms of preventingintense flashbacks that are symptomatic of Post-TraumaticStress Disorder (PTSD). Additionally, recent literature hasemphasized the role SWRs play in memory enhancement,indicating SWR access may lead to a better understandingof memory.
Accounting for Anomalous Nature of Sharp-Wave Ripples Purposeful PreProcessing
Boosting Cascade Learning
Spectral Domain Conversion
1112 1567 288
008039
15 0 21
True L
abel
Non-event
Non-event
Pre-SWR
Pre-SWR
SWR
SWR
Predicted Label
0
1500
1200
900
600
300
2967 0
00119
36 0
0
0
0
True L
abel
Non-event
Non-event
Pre-SWR
Pre-SWR
SWR
SWR
Predicted Label
0
1500
1200
900
600
300How can we account for this imbalance?
Realistically, approximately 1-5% of dataset
High accuracy labeling every sample as a non-event
Ripples are relatively rare in neural data
Custom loss functionMultiple classifiers: Boosting vs. Cascading
Segment input data
Train multiple weak classifiers
Maximize performance by alteringweights
Increase the weight ofmislabeled data
Naïve network performance
1 classifier for all input data
Aim for 100% TP, ~50% FP
Maximize performance viaadditional training
Pass mislabeled data to trainnext classifier
Each layer produces lessmislabeled data
Each classifier specialized forsome feature
Sliding window across data
Rodent electrophysiological recordings from sleepand active states
14 bins in a window, 12 samples per bin~134.4 ms (sampling rate 1250 Hz)
Continuous wavelet transformStandardized within frequency band over entireperiod of activity/inactivity
Label each bin (every 12 samples)
Online detection/disruption improvement isthe ultimate goal
Access to ripple generation/pre-SWR processes
In vivo validation
Preventative stimulation
Correlations between SWRsand other neural activity as aformal way to quantify detectionand disruption
To the Buzsáki lab at NYU, forsharing multishank recording datafor the purpose of this project.
[A] Colgin, Laura Lee. "Rhythms of the hippocampal network."Nature Reviews Neuroscience (2016).[B] Carr, Margaret F., Shantanu P. Jadhav, and Loren M. Frank."Hippocampal replay in the awake state: a potential substratefor memory....[C] Buzsáki, György, and Fernando Lopes da Silva. "Highfrequency oscillations in the intact brain." Progress inneurobiology 98.3 (2012): 241-249.[D] Jadhav, Shantanu P., et al. "Awake hippocampal sharp-waveripples support spatial memory." Science 336.6087 (2012):1454-1458.[E] Maingret, Nicolas, et al. "Hippocampal-cortical couplingmediates memory consolidation during sleep." NatureNeuroscience (2016).[F] Sethi, Ankit, and Caleb Kemere. "Real time algorithms forsharp wave ripple detection." Engineering in Medicine andBiology Society (EMBC).....
This work is funded by HFSP Young Investigators (RGY0088),NSF CAREER (CBET-1351692), NSF BRAIN EAGER (IOS-1550994)and the Rice Undergraduate Scholars Program.
weight1
weight2weightn
1C C2 Cn
Combine votes fromall the classifiers
1C C2 Cn
Combine votes fromall the classifiers
[E]
Future Works
Decoupling frequency band activitycorresponding to SWRs
Special Thanks
Realtime closed-loop detection and disruption systems haveproven difficult to implement, as the transient nature of SWR events causes the detection latency of a system tocontribute significantly towards its accuracy and application.State-of-the-art algorithms implement a threshold crossingapproach, which have been shown to have relatively hightrue positive and low false positive rates, both online andoffline.However, online applications have a latency ofapproximately 35-60 ms, indicating that the first 40-60% ofa SWR event may proceed uninterrogated. We aim tocircumvent this issue by searching for predictive indicatorsof SWR events, with the goal of predicting 20-30 ms beforea SWR event.
Objective
Figure adapted from Colgin Nature Reviews Neuroscience 2016
Pla
ce C
ell
Spik
es
Sharp
-Wave R
ipple
Active Exploration Inactivity
Machine Learning Methods for Sharp-Wave Ripple PredictionAriel Feldman, Shayok Dutta, Caleb Kemere
1
Department of Electrical and Computer Engineering, Rice University, 4Department of Neuroscience, Baylor College of Medicine3
2Department of Computer Science, Rice University, Department of Psychological Sciences, Rice University,
1,2 3 3,4
Frequency
Band (
Hz)
50
150
350
250
Time (s)
3.0 4.0 5.0 6.0 7.0
Average CA3 Activity During SWR
3356.9 3357.0 3357.1 3357.2
Standardized SWR in CA1