Classifying neurons using virtual reality...
1
HF PER POR EC 1. Introduction and data 2. Grid vs. {border, non-grid} 3. Gain manipulations 5. Summary and discussion 6. Future directions References Classifying neurons using virtual reality recordings Group 530: Malcolm Campbell and Mark Plitt {malcolmc, mplitt}@stanford.edu Open Field (OF) Virtual Reality (VR) 400 cm linear track 400 cm • Neurons in the mammalian medial entorhinal cortex (MEC) respond to the current location of the animal, forming a cognitive map of its environment [1]. • These neurons can be classified based on their pat- tern of activity relative to the environment as “grid,” “border,” and other cell types (A). • To leverage new recording technologies, we wish to record these neurons while the animal’s head is fixed in place. To achieve this, we trained mice to run along a virtual hallway (VR, B). • We would like to know whether classical MEC cell types are separable based on VR recordings alone (C). A 0 max firing rate B 0 25 0 20 0 15 0 30 0 4 0 20 0 15 0 15 0 6 0 10 0 20 0 25 0 10 0 2.5 0 25 0 20 Hz Hz Hz Hz Hz trial trial trial trial trial Hz Hz Hz trial trial trial Hz Hz Hz Hz trial trial trial trial Hz Hz Hz Hz trial trial trial trial 0 cm 0 cm 400 400 0 cm 0 cm 400 400 0 cm 0 cm 400 400 0 400 0 cm 400 cm 0 cm 400 0 cm 0 cm 400 400 0 400 0 cm 0 cm 400 400 cm 0 cm 0 cm 400 400 Grid Cells Border Cells Irregular Spatial Cells Non-Spatial Cells C Grid Border Irregular Spatial Non-Spatial 90 cm A new cell class: “Drifty bursting cells” • So far, it was not possible to separate MEC cell types based on features of the firing rate map alone, even with data augmentation • Adding features derived from cells’ responses to gain manipulations improved performance • This could reflect the fact that different cell types derive their spatial responses from different inputs • 80-85% accuracy appears to be the limit of perfor- mance using this feature • We identified by eye and hand-labeled a new cell class which we call “drifty bursting cells” • Using 7 hand-crafted features, we could separate drifty bursting cells (~30% of the population) from other cells with ~75% accuracy • Dynamic time-warping could in theory identify cells whose firing patterns drift over time, but on first pass this technique tended to over-fit our data POR • Unbalanced classes (grid, border, other): How much of a problem? • Penalize false-positives more highly in cost function • Refine definition of “drifty bursting cells” • How well can we predict response to gain manipula- tions? • Convolutional approaches to identify spatially-invari- ant features of firing rate maps • Identify and eliminate failure modes of dynamic time-warping • Use dynamic time-warping to identify neurons that drift, warp, or bifurcate over time in particular ways -8 -6 -4 -2 0 2 4 PC1 -6 -4 -2 0 2 4 6 PC2 Other cells Drifty bursters • Mean firing rate (1x1) • Firing rate map (200 x 1) • Magnitude of FFT of firing rate map (100 x 1) • Average cross-correlation between single trial rate maps (21 x 1) • Location of peak average trial cross correlation (1x1) 1) Mean firing rate (r) 2) Median inter-spike interval (i) 3) Burstiness (1/(r·i)) 4) Stability (s): Correlation between firing rate map computed in first and second half of recording 5) Trial-to-trial stability (t): Mean correla- tion between firing rate maps comput- ed from adjacent trials 6) Stability ratio (t/s) 7) Field size (f ): Width of autocorrelation Features (n = 7) • Hand labeled 1,244 recor- dings (375 drifty bursters) 1. Baseline features (n = 323) 2. Data-augmentation 3. Dynamic time-warping 4. Drifty bursting cells Classifier performance (accuracy, LOOCV) Train Test L2- Logistic Linear SVM RBF- SVM GDA 74.0 74.0 87.5 78.7 73.7 73.8 74.7 78.4 Drifty bursters vs. other cells Classifier performance (accuracy, LOOCV) Train Test L2- Logistic Linear SVM RBF- SVM GDA 77.6 78.1 90.0 1 73.9 72.5 71.0 53.6 Train Test 80.7 84.4 92.7 1 77.1 81.4 81.4 52.9 Grid vs. border, gain decrease Grid vs. border, gain increase Classifier performance (accuracy, LOOCV) Train Test L2- Logistic Linear SVM RBF- SVM GDA 54.4 60.4 84.4 100 49.0 53.7 54.2 54.7 Train Test 52.2 61.0 90.7 99.0 52.3 58.0 58.0 54.4 Grid vs. non-grid (downsampled) Grid vs. border Original firing rate map Augmented features (rotations and reflections • Idea: Get classifier to generalize features over loca- tion by feeding it rotated and reflected copies of the original data Smoothed raw data Aligned data • Idea: Re-align cells that drift over time to extract underlying structure [2] Smoothed raw data Aligned data Success Failure Classifier performance (accuracy, LOOCV): Grid vs. non-grid Train Test L2- Logistic Linear SVM RBF- SVM GDA 97.5 97.5 97.5 97.5 50.3 50 50 50 400 virtual cm, 400 real cm Baseline (A): 1x ≥15 trials 400 virtual cm, 800 or 267 real cm Gain change (B): 0.5x or 1.5x 5 or 10 trials Baseline (A’): 1x ≥15 trials) 400 virtual cm, 400 real cm -100 -50 50 100 -0.2 0 0.2 0.4 0.6 grid (40) border (33) 0 gain decrease lag (cm) A-B correlation Features: Cross-correlation of A and B period firing rate maps (101 x 1) 0 10 20 30 40 50 60 training set size 0.4 0.5 0.6 0.7 0.8 0.9 1 accuracy Grid vs. border, gain increase (linear SVM) train test 1) Rowland, D.C., Roudi, Y., Moser, M.B. & Moser, E.I. Ten years of grid cells. Annu Rev Neurosci 39, 19-40 (2016). 2) Cuturi, M., and Blondel, M. Soft-DTW: a Differentia- ble Loss Function for Time-Series. arXiv:1703.01541v1, 5 Mar 2017. 0 30 0 400 0 30 0 400 0 15 0 400 0 10 0 400 trial trial trial trial Hz Hz Hz Hz Dataset: 781 cells (96 grid, 97 border, 590 other)
Transcript of Classifying neurons using virtual reality...
HF PERPOR
EC
1. Introduction and data 2. Grid vs. {border, non-grid} 3. Gain manipulations 5. Summary and discussion
6. Future directions
References
Classifying neurons using virtual reality recordingsGroup 530: Malcolm Campbell and Mark Plitt {malcolmc, mplitt}@stanford.edu
Open Field (OF)
Virtual Reality (VR)400 cm linear track
400 cm
• Neurons in the mammalian medial entorhinal cortex (MEC) respond to the current location of the animal, forming a cognitive map of its environment [1].
• These neurons can be classi�ed based on their pat-tern of activity relative to the environment as “grid,” “border,” and other cell types (A).
• To leverage new recording technologies, we wish to record these neurons while the animal’s head is �xed in place. To achieve this, we trained mice to run along a virtual hallway (VR, B).
• We would like to know whether classical MEC cell types are separable based on VR recordings alone (C).
A
0
max
�rin
g ra
te
B
0
25
0
20
0
15
0
30
0
4
0
20
0
15
0
15
0
6
0
10
0
20
0
25
0
10
0
2.5
0
25
0
20
Hz
Hz
Hz
Hz
Hz
tria
ltr
ial
tria
l
tria
l
tria
l
Hz
Hz
Hz
tria
l
tria
l
tria
l
Hz
Hz
Hz
Hz
tria
l
tria
l
tria
l
tria
l
Hz
Hz
Hz
Hz
tria
l
tria
l
tria
l
tria
l
0 cm 0 cm 400400 0 cm 0 cm 400400
0 cm 0 cm 400400 0 400 0 cm 400cm
0 cm 400 0 cm 0 cm 4004000 400
0 cm 0 cm 400400
cm
0 cm 0 cm 400400
Grid Cells Border Cells
Irregular Spatial Cells Non-Spatial Cells
C
Grid Border Irregular Spatial Non-Spatial
90 c
m
A new cell class: “Drifty bursting cells”
• So far, it was not possible to separate MEC cell types based on features of the �ring rate map alone, even with data augmentation
• Adding features derived from cells’ responses to gain manipulations improved performance
• This could re�ect the fact that di�erent cell types derive their spatial responses from di�erent inputs
• 80-85% accuracy appears to be the limit of perfor-mance using this feature
• We identi�ed by eye and hand-labeled a new cell class which we call “drifty bursting cells”
• Using 7 hand-crafted features, we could separate drifty bursting cells (~30% of the population) from other cells with ~75% accuracy
• Dynamic time-warping could in theory identify cells whose �ring patterns drift over time, but on �rst pass this technique tended to over-�t our data
POR
• Unbalanced classes (grid, border, other): How much of a problem?
• Penalize false-positives more highly in cost function
• Re�ne de�nition of “drifty bursting cells”
• How well can we predict response to gain manipula-tions?
• Convolutional approaches to identify spatially-invari-ant features of �ring rate maps
• Identify and eliminate failure modes of dynamic time-warping
• Use dynamic time-warping to identify neurons that drift, warp, or bifurcate over time in particular ways
-8 -6 -4 -2 0 2 4PC1
-6
-4
-2
0
2
4
6
PC2
Other cellsDrifty bursters
• Mean �ring rate (1x1)• Firing rate map (200 x 1)• Magnitude of FFT of �ring rate map (100 x 1)• Average cross-correlation between single trial rate maps (21 x 1)• Location of peak average trial cross correlation (1x1)
1) Mean �ring rate (r)2) Median inter-spike interval (i)3) Burstiness (1/(r·i))4) Stability (s): Correlation between �ring rate map computed in �rst and second half of recording5) Trial-to-trial stability (t): Mean correla- tion between �ring rate maps comput- ed from adjacent trials6) Stability ratio (t/s)7) Field size (f ): Width of autocorrelation
Features (n = 7)
• Hand labeled 1,244 recor-dings (375 drifty bursters)
1. Baseline features (n = 323)
2. Data-augmentation
3. Dynamic time-warping
4. Drifty bursting cells
Classi�er performance(accuracy, LOOCV)
Train
Test
L2-Logistic
LinearSVM
RBF-SVM
GDA
74.0 74.0 87.5 78.7
73.7 73.8 74.7 78.4
Drifty bursters vs. other cells
Classi�er performance (accuracy, LOOCV)
Train
Test
L2-Logistic
LinearSVM
RBF-SVM
GDA
77.6 78.1 90.0 1
73.9 72.5 71.0 53.6
Train
Test
80.7 84.4 92.7 1
77.1 81.4 81.4 52.9
Grid vs. border, gain decrease
Grid vs. border, gain increase
Classi�er performance (accuracy, LOOCV)
Train
Test
L2-Logistic
LinearSVM
RBF-SVM
GDA
54.4 60.4 84.4 100
49.0 53.7 54.2 54.7
Train
Test
52.2 61.0 90.7 99.0
52.3 58.0 58.0 54.4
Grid vs. non-grid (downsampled)
Grid vs. border
Original �ring rate mapAugmented features
(rotations and re�ections
• Idea: Get classi�er to generalize features over loca-tion by feeding it rotated and re�ected copies of the original data
Smoothed raw data Aligned data
• Idea: Re-align cells that drift over time to extract underlying structure [2]
Smoothed raw data Aligned data
Success Failure
Classi�er performance(accuracy, LOOCV):
Grid vs. non-gridTrain
Test
L2-Logistic
LinearSVM
RBF-SVM
GDA
97.5 97.5 97.5 97.5
50.3 50 50 50
400 virtual cm, 400 real cm
Baseline (A): 1x≥15 trials
400 virtual cm, 800 or 267 real cm
Gain change (B): 0.5x or 1.5x5 or 10 trials
Baseline (A’): 1x≥15 trials) 400 virtual cm, 400 real cm -100 -50 50 100
-0.2
0
0.2
0.4
0.6grid (40)border (33)
0
gain decrease
lag (cm)
A-B
cor
rela
tion
Features: Cross-correlation of A and B period �ring rate maps (101 x 1)
0 10 20 30 40 50 60training set size
0.4
0.5
0.6
0.7
0.8
0.9
1
accu
racy
Grid vs. border, gain increase(linear SVM)
traintest
1) Rowland, D.C., Roudi, Y., Moser, M.B. & Moser, E.I. Ten years of grid cells. Annu Rev Neurosci 39, 19-40 (2016).
2) Cuturi, M., and Blondel, M. Soft-DTW: a Differentia-ble Loss Function for Time-Series. arXiv:1703.01541v1, 5 Mar 2017.
0
30
0 400
0
30
0 400
0
15
0 400
0
10
0 400tr
ial
tria
l
tria
l
tria
lH
zHz
HzHz
Dataset: 781 cells (96 grid, 97 border, 590 other)
