Encoding of Naturalistic Stimuli by Local Field Spectra in Networks of Excitatory and Inhibitory Neur.pptx

A summary of the findings and methods presented in the paper

1. Encoding of Naturalistic Stimuli by Local Field Spectra in Networks of Excitatory and Inhibitory Neurons Alberto Mazzoni, Stefano Panzeri, Nikos K. Logothetis, Nicolas Brunel

2. Abstract

3. Abstract Recordings of local field potentials (LFPs) show that the sensory cortex displays rhythmic activity and fluctuations over a wide range of frequencies and amplitudes. But its role in encoding sensory information is largely unknown. Aim: simulate a sparsely connected network of excitatory and inhibitory neurons modeling a local cortical population to determine how the LFPs generated encode information about input stimuli. This was done to understand the rules of translation between the structure of sensory stimuli and the fluctuations of cortical responses. Two types of stimuli considered: simple static and periodic stimuli naturalistic input stimuli based on electrophysiological recordings from the thalamus of anesthetized monkeys watching natural movie scenes.

4. Abstract The simulated network produced stimulus-related LFP changes that were in striking agreement with the LFPs obtained from the primary visual cortex. T he network encoded static input spike rates into gamma-range oscillations generated by inhibitory–excitatory neural interactions and encoded slow dynamic features of the input into slow LFP fluctuations mediated by stimulus–neural interactions. The model cortical network processed dynamic stimuli with naturalistic temporal structure, using low and high response frequencies as independent communication channels - in agreement with reports from visual cortex responses to naturalistic movies. 1 2 3 Result Summary

5. Introduction

6. Introduction Presentation of sensory stimuli elicits oscillations in EEG and LFP recordings which span a broad frequency spectrum (0-100+ Hz) Gamma band (30–100 Hz) oscillations are robustly triggered and modulated by sensory stimuli. Some behaviorally relevant stimuli (stimuli with rhythmic, complex, or naturalistic dynamics) elicit cortical oscillations at specific frequencies in the low-frequency (10–20 Hz) range. The prominent presence of oscillations in sensory systems raises two important questions: How are these oscillations generated? - The mechanism? Why are they generated? - Function?

7. Introduction Purely chemical synapses -> oscillatory synchrony can emerge through: mutual inhibitory interactions due to a feedback loop between excitatory and inhibitory neurons Firing frequency depends on synaptic time scales and the ratio of excitation to inhibition Population of cells -> fire rhythmically at high frequencies Single cells -> fire stochastically at a much lower rate The oscillatory regime strongly depends on external inputs: For weak external inputs -> asynchronous network , with small damped oscillations due to finite size effects As the inputs increase -> network becomes more synchronized, amplitude of the oscillation increases Mechanism of Oscillation Generation

8. Introduction How, even in the same sensory area, different types of stimuli encode information in different frequency bands How the combination of fluctuations generated by stimulus-neural interactions and the oscillations generated by neural-neural interactions shapes the network dynamics and sensory information encoding What is the computational advantage of the cortical encoding of stimuli by a variety of frequencies Function of Oscillations

9. Introduction Stimulus-related changes of low frequency cortical fluctuations o rigin : stimulus-neural interaction encode information about slowly varying features in the sensory/thalamic input Stimulus related changes of high-frequency cortical oscillations origin: neural-neural interactions carry information about sensory features that provoke thalamic responses that differ only in terms of their total spike rate Simulated a network of excitatory and inhibitory neurons modeling a local population in primary visual cortex Determined how LFPs and spiking activity generated by the network encode information about simple or complex inputs, the latter simulating sensory-related thalamic signals Hypothesis H ypothesis Testing

10. Results

11. Results Experiment Leaky integrate-and-fire neurons used to model a cortical network Synaptic currents had time courses resembling AMPA currents (excitatory) and GABA currents (inhibitory) Interneurons received stronger inputs than pyramidal neurons LFPs are monitored due to irregularity in individual neuron data Three types of inputs (each superimposed with noise) Constant in time and vary only in rate Periodic inputs of different frequency and amplitude Complex broadband inputs with a statistics similar to that of geniculate neurons responding to naturalistic movies

12. Results Constant input at different rates Average firing and LFPs Increasing the signal rate led to an increased average firing and LFPs in the network and population spikes occurred in a synchronous fashion A pronounced population oscillations in the gamma band

13. Results Constant input at different rates The trial-averaged power spectra of LFPs Showed the highest power at low frequencies, with a local peak in the gamma range Power modulation Low below 30Hz, peak at ~70Hz, plateau at >100Hz Total firing rate Similar power spectra to LFPs

14. Results Constant input at different rates Shannon Information I(S; Rf) Low below 30Hz, peak at ~70Hz, plateau at >100Hz Information Redundancy Information is redundant if 2 frequencies Are tuned in the same way to the stimulus features Share correlated sources of noise redundancy is highly positive for >50Hz

15. Results Constant input at different rates Signal and Noise Correlation Signal correlation is very high for 30-100Hz -> they are tuned to the same stimuli Noise was uncorrelated Redundancy is due to signal correlation

16. Results Periodic inputs of different frequency and amplitude Trial-averaged LFP power spectra (per frequency) Peak at exactly the input frequency Very small modulation in gamma band Trial-averaged LFP power spectra (per amplitude) Peak at all input frequencies with higher height for higher amplitude

17. Results Periodic inputs of different frequency and amplitude Information(Frequency, Amplitude) Peak at exactly the input frequencies Peak at input frequency only for Information(Frequency). No peak for Information(Amplitude) Circular variance of the phase difference between the signals and the band-passed LFP Entrainment at input frequency Stronger for lower frequencies

18. Results Complex broadband inputs Information of LGN MUA (input) Almost all the information was carried by the power at frequencies below 5 Hz and in the average spike rate Information of power of LFP at different frequencies (output) High information in ranges of 1-5 Hz and 50-80 Hz, corresponding to information about the low frequency modulations of the signal and its rate

19. Results Complex broadband inputs Joint Information Peak when one frequency is low (3 Hz) and other is in gamma range (70 Hz) Redundancy One low and one gamma frequency gives ~0 redundancy Two low frequencies also have ~0 redundancy Two gamma frequencies have av = 0.08 bits redundancy

20. Results Complex broadband inputs Signal Correlation High among frequencies in gamma range, as they are all modulated in a similar way by the naturalistic stimuli Low and gamma frequencies had low signal correlation Noise Correlation Negligible everywhere Thus low and gamma frequencies are completely decoupled and add independent information about the stimulus

21. Results Complex broadband inputs We change average rate by varying baseline (B) LFP at low frequencies are not affected, high modulation in gamma range Slight increase in information Any change in B decreased the agreement of the information content of the spectrum of simulated and recorded LFPs

22. Results Complex broadband inputs Change spectral content and make it constant (= average signal rate) LFP decreases at low frequencies Thus signal oscillations determine only a narrow band of the signal output Information falls slightly below significance level for low frequencies

23. Results Changing Synaptic Strengths Decrease in GABA currents Increased the power for all frequencies (less in low frequency range) No change in information Increase in AMPA currents Increased the power for all frequencies (more in gamma range) Decrease in information at high frequencies Only quantitative, no qualitative changes (in inhibitory-dominated regime)

24. Discussion

25. Discussion Advances provided by this study: Q uantified the information content of the fluctuations generated by the network and determined which LFP frequencies convey most information. Found explicit coding rules between features of the stimulus dynamics and LFP frequency which are compatible with several neurophysiological reports. Demonstrated that these coding rules lead to low and high LFP frequencies acting as largely independent information channels, consistent with experimental data.

26. Advances with Respect to Previous Modeling Work Generalized the results to characterize the network dynamics to slowly-varying periodic and naturalistic stimuli (instead of only constant stimulation). Previous studies quantified the network output only as the total firing rate, whereas this study quantified its output also in terms of simulated LFPs. Findings included: Modulation of the LFP gamma band when using stimuli eliciting firing rate modulations, the entrainment of low frequency LFPs to stimulus oscillations and the way the two phenomena contribute to the information content of the whole LFP spectrum. If we study LFP from 3D models of networks of compartmental model neurons, we could see which combination of average AMPA/GABA currents best represents the ‘true’ LFP.

27. Dependence of LFP Frequency on Stimulus First coding rule: Gamma-range LFPs carry information about sensory features, that provoke thalamic responses, that differ only in terms of their total spike rate. Second coding rule: Stimulus-related changes of low-frequency cortical fluctuations encode information about slow dynamic features in the sensory or thalamic input that vary at the considered frequency. The double peak of information at low frequencies and in the gamma range can be explained by this rule, since a natural movie contains both temporal frequency changes at low frequencies and objects and features capable of eliciting firing rate changes

28. Correlation between Stimulus Selectivity of Different Frequency Bands Model reproduced findings that low frequency LFPs and gamma-range LFPs act as independent information channels and was able to provide an explanation: They reflect two independent input features (slow frequency variation of input rate and total input spike count). Allows cortical network to transmit more information by multiplexing over timescales. The independence of information carried by low and high frequency LFPs might be relevant to the development of brain machine interfaces, as it suggests that simultaneous decoding of different LFP bands may provide more information. The model’s lack of noise correlation in the 12-24 Hz range aligns with the hypothesis that this region is related mainly to stimulus-independent neuromodulation, which the model doesn’t account for.

29. Functional Characterization of LFP Bands Many reports differ in their setting of boundaries of each band and which frequency range they investigate. If bands are partitioned in an information theoretically optimal manner, the optimal gamma range partitioning would remain roughly the same as gamma range coding happens robustly whenever the network receives input rate modulations. However, partitioning lower frequency ranges in the same manner may provide boundaries that are dependent on the stimulus dynamics and not intrinsically in the network.

30. Experimental Predictions Arising from the Model The model hypothesizes that low frequency LFPs follow the dynamics of the stimulus in some capacity. This hypothesis can be tested by using faster stimuli than natural movies and studying how this affects the informative LFP-frequency range: If exact positions of low frequency information peaks match with input spectrum, low frequency band modulations are mostly reproducing the input spectrum Similarly, Gamma-band information should align with the rate range of the input, which can be modulated in input stimuli with image contrast In the presence of GABA antagonists, there is an increase in the power of the gamma band and only a minute decrease of the associated information

31. Methods

32. Methods Model Simulated network composed of 5000 neurons (4000 pyramidal neurons and 1000 interneurons) described by LIF dynamics. Connection probability between any 2 neurons - 0.2. Each neuron, k, described by membrane potential - Auxiliary variables, x Ak and x Gk defined to obtain dynamic equations for AMPA and GABA currents -

33. Methods Model Dynamic variability of x Ak depends on time of spikes received from pyramidal neurons and external inputs. x Gk variance with time depends on time of spikes received from interneurons. Modifying parameters of the model (latency or decay/rise time constants) can change both, location and shape of peak in LFP power spectrum and information vs. frequency curve.

34. Methods External Inputs Each neuron receives an external excitatory synaptic input (assumed to come from the thalamus in model). Provided as random Poisson spike trains, with time varying rate given by - Each simulation repeated 20 times with same signal and independently generated noise for each trial.

35. Methods Signal Constant signals used (v ranging from 1.2 to 2.6 spikes/ms) - Periodic signals used (with varying ω and A) - Naturalistic signals - Built from single electrode multi unit activity (MUA) recording from lateral geniculate nucleus (LGN) of anesthetized monkey watching natural movie scenes.

36. Methods Noise 2 sources of noise in model - n(t), stochastic variable generated according to an Ornstein-Uhlenbeck process, η(t) : a Gaussian white noise Due to different neurons receiving independent realizations of a Poisson process, with same rate.

37. Methods Measure of Information Information, I(S, R f ), between power R f at frequency f and the stimuli S - P(s) : Probability of presentation of stimulus s. P(r f ) : Probability of frequency f to have power r f over all trials and stimuli Information that joint knowledge of powers R f1 and R f2 conveys - Information redundancy - Red(f 1 , f 2 )