Amp-FM Linear-Non-Linear Model
A Rat Model of Speech: Cortical Encoding of Vocalizations Isaac M. Carruthers1,2, Ryan G. Natan1,3, Maria Neimark Geffen1 1. Department of Otorhinolaryngology, 2. Physics Graduate Program, 3. Neuroscience Graduate Program, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA
5 0
FM responsiveness is correlated with vocalization responsiveness
Firing Rate (Hz)
10 0 .32
1 3.2 10 32 100 Best Frequency (kHz)
80 70 60 50 40 30 20 10
Firing rate (Hz)
80
40
1
10 Frequency (kHz)
Cell 2 5 4 3 2 1 0
1
0
2
3
4
5
6
7
25
5 0
Purified vocalization #
0
1 2 3 FM Sweep Responsiveness
FM sweep responsiveness
−4
0
−200
450
0 −0.2 −0.4
150
250 Time (ms)
350
−200
450
Temporal Modulation (Hz)
Firing rate is invariant to USV transformations
Delay(ms)
0
40
Frequency (kHz)
0.6
Delay(ms)
0.4
40
2 1
0.8 0.6 0.4 0.2
0.2
0.4
0.6
0.8
1
Prediction Accuracy for Original USVs
20 0
3
0
0
−1
4
1
0
0
5
0
0.8
40
6
0
0.2
0.4
0.6
0.8
1
Prediction Accuracy for Original USVs
1
Linear Prediction
CONCLUSIONS A1 neurons respond selectively to a subset of USVs
3 Time (s)
4
0 1 2 3 Response d−prime
Acknowledgements Thanks to: - Dr. Diego A. Laplagne for providing the vocalization recordings, - Dr. Yale Cohen and Dr. Mark Aizenberg for helpful discussions, - Liana Cheung, Danielle Mohabir, Lisa Liu and Laetitia Mwilambwe-Tshilobo for technical assistance, - Ana Calabrese and Sarah Woolley for providing an earlier version of code for the generalized linear model Funding provided by: - UPenn Complex Scene Perception IGERT grant to IMC - Buroughs Welcome Career at the Scientific Interface award to MNG - Pennsylvania Lions Hearing Research grant to MNG - Klingenstein Award for the Neurosciences to MNG - University of Pennsylvania Center for Collaborative Neuroscience Pilot Grant to MNG
0.5
Vocalization Responsiveness
2
0
0
Vocalization Selectivity
1
Firing rate (Hz)
30 1
25 20 15 10 5 0
Measured response Amp-FM LN prediction
r = 0.49, p = 1e−07
16 64
0
0.2
1
4
Probability
FM sweep rate (Octave/s)
30% neurons respond to at least one USV Responsive neurons respond to 20% of USVs
−16
FM tuning index is correlated with vocalization responsiveness
200
0.2
0.4
0.2
10
8
−64
0
20 15
350
0.1
0
0
r = 0.67, p = 5e−15
30 Vocalization Responsiveness
20
Response Strength
30
0
0.4
ed
Purified vocalization #
−200
ed rs ve Re d te ila d D se es pr m Co al in
8
200
−0.4
d se
7
250
r ve Re
6
4
10 Best Frequency
150
20
d
5
3
10
0
0 −0.2
40
te ila
4
−3
x 10
25
Time (s)
5 0
200
0.2
−0.4
60
D
3
4
24
s es pr
2
2 3 Spectral Overlap
23
0.2
al
1
1
10
0
0.4
−0.2
20
fit to original fit to transformed
m Co
0
0
25
0
15
200
rig
0.5
Best frequency of recorded units Number of neurons
10
24
Firing rate (Hz)
1
15
450
0
Lower prediction accuracy for transformed vocalizations
Frequency modulation (FM)
20
350
40
50
0 23
Amplitude
1.5
20
250
60
Amplitude
r = 0.37, p = 1e−10
25 Vocalization Responsiveness
2
25 Vocalization Responsiveness
Cell 1
r = 0.01, p = 0.81
REVERSED
in
Bregma -4.92 mm
Different cells respond to different subsets of the vocalizations. We want to explain this effect in terms of the properties of the neurons.
20 150
−200
450
40
50
rig
AuV
Time
DILATED
350
60
50
Amp-FM Linear-Non-Linear Model
Spectral tuning is correlated with vocalization responsiveness
0
Time (s)
64
100 50
25
O
Au1
Amplitude (dB)
1 2 4 8 16 32 Best Frequency (kHz)
24
250
O
AuD
A traditional example is a "frequency tuning curve" like the one ro the right. We play a long series of tones at different volumes and frequencies, and this gives us a picture of how much a cell tends to respond to different tones.
23
0
Cell 2
−0.4
150
Mean firing rate (Hz)
0
0 −0.2
20
Prediction Accuracy for Dialted USVs
50
Frequency (kHz)
100
COMPRESSED
Frequency (kHz)
1
0.2
40
50
Frequency (kHz)
Cell 1
0
Mean prediction accuracy
100
ORIGINAL
0.4
60
Prediction Accuracy for Compressed USVs
40
Response Strength
Localization of the electrode tract in A1
Networks of neurons pass information via what are called "action potentials", which are very fast pulses of electrical discharge. By recording these action potentials, or "spikes", we can identify what sorts of stimuli ellicit the most response from a particular neuron
Another way we can examine the responses of cells is to expose them to tones that sweep upward or downward in frequency. This tells us whether cells prefer tones that sweep upward or downward, or tones that sweep faster or slower. We can use this sort of information to predict how cells will respond to other sounds.
Vocalization Selectivity
Frequency (kHz) Firing Rate (Hz) Trial #
80
Trial #
CELLS IN AUDITORY CORTEX
Transformed vocalizations
1
1
- Do cells in the auditory cortex respond selectively to vocalizations? - Can we predict how a cell will respond to different vocalizations? - How Do the responses compare if we distort the vocalizations?
ENCODING OF TRANSFORMED VOCALIZATIONS
Sound Wave
not significant
100
QUESTIONS
RELIABLE MODEL OF NEURAL RESPONSES TO VOCALIZATIONS
SpectralModulation(1/kHz)
RESPONSES TO RAT VOCALIZATIONS
Rats produce complex vocalizations in communicating with each other. Over 14 types of distinct calls can be distinguished in their repertoire (Clarke et al., 2009). Neurons in the primary auditory cortex respond selectively to con-specific vocalizations (Wang et al. 1995, Liu and Schreiner, 2007, Petkov et al, 2008, Huetz et al., 2009, Chandrasekaran et al., 2010). However, the precise - mecha nisms of how complex vocalizations are encoded in the auditory pathway are not well understood. To learn how the auditory cortex encodes information about rat vocalizations, we presented a library of recorded and purified vocalizations to awake rodents, recorded neural activity in the auditory cortex and constructed a mathematical model that allowed us to predict A1 responses to novel - vocaliza tions.
In order to record the responses of neurons in the auditory cortex, we implanted several rats with tetrode electrodes. Tetrodes are bundles of four wires, twisted tightly so that we can isolate the activity of many different cells by looking at how the signal compares across the four channels.
Spectral tuning is not correlated with vocalization selectivity
FM
INTRODUCTION
0
0.1 0.2 0.3 FM Rate Tuning Index
0.4
20
The responses are accurately predicted by a reduced LN model, based on integration of frequency modulation and amplitude
10 0
Responses to USVs are correlated with responsiveness- to fre quency modulation and spectral tuning
23
24
Time (s)
25
Responses are predicted less accurately to temporally transformed vocalizations suggesting a differential encoding mechanism specific to the temporal statistics of original USVs
5 0
FM responsiveness is correlated with vocalization responsiveness
Firing Rate (Hz)
10 0 .32
1 3.2 10 32 100 Best Frequency (kHz)
80 70 60 50 40 30 20 10
Firing rate (Hz)
80
40
1
10 Frequency (kHz)
Cell 2 5 4 3 2 1 0
1
0
2
3
4
5
6
7
25
5 0
Purified vocalization #
0
1 2 3 FM Sweep Responsiveness
FM sweep responsiveness
−4
0
−200
450
0 −0.2 −0.4
150
250 Time (ms)
350
−200
450
Temporal Modulation (Hz)
Firing rate is invariant to USV transformations
Delay(ms)
0
40
Frequency (kHz)
0.6
Delay(ms)
0.4
40
2 1
0.8 0.6 0.4 0.2
0.2
0.4
0.6
0.8
1
Prediction Accuracy for Original USVs
20 0
3
0
0
−1
4
1
0
0
5
0
0.8
40
6
0
0.2
0.4
0.6
0.8
1
Prediction Accuracy for Original USVs
1
Linear Prediction
CONCLUSIONS A1 neurons respond selectively to a subset of USVs
3 Time (s)
4
0 1 2 3 Response d−prime
Acknowledgements Thanks to: - Dr. Diego A. Laplagne for providing the vocalization recordings, - Dr. Yale Cohen and Dr. Mark Aizenberg for helpful discussions, - Liana Cheung, Danielle Mohabir, Lisa Liu and Laetitia Mwilambwe-Tshilobo for technical assistance, - Ana Calabrese and Sarah Woolley for providing an earlier version of code for the generalized linear model Funding provided by: - UPenn Complex Scene Perception IGERT grant to IMC - Buroughs Welcome Career at the Scientific Interface award to MNG - Pennsylvania Lions Hearing Research grant to MNG - Klingenstein Award for the Neurosciences to MNG - University of Pennsylvania Center for Collaborative Neuroscience Pilot Grant to MNG
0.5
Vocalization Responsiveness
2
0
0
Vocalization Selectivity
1
Firing rate (Hz)
30 1
25 20 15 10 5 0
Measured response Amp-FM LN prediction
r = 0.49, p = 1e−07
16 64
0
0.2
1
4
Probability
FM sweep rate (Octave/s)
30% neurons respond to at least one USV Responsive neurons respond to 20% of USVs
−16
FM tuning index is correlated with vocalization responsiveness
200
0.2
0.4
0.2
10
8
−64
0
20 15
350
0.1
0
0
r = 0.67, p = 5e−15
30 Vocalization Responsiveness
20
Response Strength
30
0
0.4
ed
Purified vocalization #
−200
ed rs ve Re d te ila d D se es pr m Co al in
8
200
−0.4
d se
7
250
r ve Re
6
4
10 Best Frequency
150
20
d
5
3
10
0
0 −0.2
40
te ila
4
−3
x 10
25
Time (s)
5 0
200
0.2
−0.4
60
D
3
4
24
s es pr
2
2 3 Spectral Overlap
23
0.2
al
1
1
10
0
0.4
−0.2
20
fit to original fit to transformed
m Co
0
0
25
0
15
200
rig
0.5
Best frequency of recorded units Number of neurons
10
24
Firing rate (Hz)
1
15
450
0
Lower prediction accuracy for transformed vocalizations
Frequency modulation (FM)
20
350
40
50
0 23
Amplitude
1.5
20
250
60
Amplitude
r = 0.37, p = 1e−10
25 Vocalization Responsiveness
2
25 Vocalization Responsiveness
Cell 1
r = 0.01, p = 0.81
REVERSED
in
Bregma -4.92 mm
Different cells respond to different subsets of the vocalizations. We want to explain this effect in terms of the properties of the neurons.
20 150
−200
450
40
50
rig
AuV
Time
DILATED
350
60
50
Amp-FM Linear-Non-Linear Model
Spectral tuning is correlated with vocalization responsiveness
0
Time (s)
64
100 50
25
O
Au1
Amplitude (dB)
1 2 4 8 16 32 Best Frequency (kHz)
24
250
O
AuD
A traditional example is a "frequency tuning curve" like the one ro the right. We play a long series of tones at different volumes and frequencies, and this gives us a picture of how much a cell tends to respond to different tones.
23
0
Cell 2
−0.4
150
Mean firing rate (Hz)
0
0 −0.2
20
Prediction Accuracy for Dialted USVs
50
Frequency (kHz)
100
COMPRESSED
Frequency (kHz)
1
0.2
40
50
Frequency (kHz)
Cell 1
0
Mean prediction accuracy
100
ORIGINAL
0.4
60
Prediction Accuracy for Compressed USVs
40
Response Strength
Localization of the electrode tract in A1
Networks of neurons pass information via what are called "action potentials", which are very fast pulses of electrical discharge. By recording these action potentials, or "spikes", we can identify what sorts of stimuli ellicit the most response from a particular neuron
Another way we can examine the responses of cells is to expose them to tones that sweep upward or downward in frequency. This tells us whether cells prefer tones that sweep upward or downward, or tones that sweep faster or slower. We can use this sort of information to predict how cells will respond to other sounds.
Vocalization Selectivity
Frequency (kHz) Firing Rate (Hz) Trial #
80
Trial #
CELLS IN AUDITORY CORTEX
Transformed vocalizations
1
1
- Do cells in the auditory cortex respond selectively to vocalizations? - Can we predict how a cell will respond to different vocalizations? - How Do the responses compare if we distort the vocalizations?
ENCODING OF TRANSFORMED VOCALIZATIONS
Sound Wave
not significant
100
QUESTIONS
RELIABLE MODEL OF NEURAL RESPONSES TO VOCALIZATIONS
SpectralModulation(1/kHz)
RESPONSES TO RAT VOCALIZATIONS
Rats produce complex vocalizations in communicating with each other. Over 14 types of distinct calls can be distinguished in their repertoire (Clarke et al., 2009). Neurons in the primary auditory cortex respond selectively to con-specific vocalizations (Wang et al. 1995, Liu and Schreiner, 2007, Petkov et al, 2008, Huetz et al., 2009, Chandrasekaran et al., 2010). However, the precise - mecha nisms of how complex vocalizations are encoded in the auditory pathway are not well understood. To learn how the auditory cortex encodes information about rat vocalizations, we presented a library of recorded and purified vocalizations to awake rodents, recorded neural activity in the auditory cortex and constructed a mathematical model that allowed us to predict A1 responses to novel - vocaliza tions.
In order to record the responses of neurons in the auditory cortex, we implanted several rats with tetrode electrodes. Tetrodes are bundles of four wires, twisted tightly so that we can isolate the activity of many different cells by looking at how the signal compares across the four channels.
Spectral tuning is not correlated with vocalization selectivity
FM
INTRODUCTION
0
0.1 0.2 0.3 FM Rate Tuning Index
0.4
20
The responses are accurately predicted by a reduced LN model, based on integration of frequency modulation and amplitude
10 0
Responses to USVs are correlated with responsiveness- to fre quency modulation and spectral tuning
23
24
Time (s)
25
Responses are predicted less accurately to temporally transformed vocalizations suggesting a differential encoding mechanism specific to the temporal statistics of original USVs
