Download Invariant selectivity of auditory neurons due to predictive coding

Invariant selectivity of auditory neurons due to predictive coding
Izzet B. Yildiz1, Brian Fischer 2, Sophie Deneve1
1
Introduction
Results
Network effect on a single neuron
Learning predictive fields from natural stimuli
What do auditory neurons represent?
•
•
Group for Neural Theory, LNC, DEC, Ecole Normale Superieure, Paris, France.
2 Seattle University, Department of Mathematics, Seattle, WA, USA.
Spectro-temporal receptive field (STRF) is the interpretation of auditory neurons as linear filters 1.
We propose that auditory neurons are predictors rather than filters of their input and we hypothesize
that they have a "true selectivity" independent of stimulus context 2.
•
•
•
•
How does a model neuron’s STRF change with network activity?
Online Expectation-Maximization algorithm
36-neuron case
2-neuron case
We propose a dynamic Bayesian inference model and train it on a large database of natural speech to
predict this invariant selectivity, i.e. predictive fields (PFs).
The model can account for nonlinear contextual effects such as two-tone and forward suppression.
The model neurons adapt rapidly to new input statistics (such as behaviorally relevant tones).
The model can be used to explain neuronal and behavioral responses.
We use decoding filters obtained from neuronal data 4 as an approximation
to predictive fields and employ the model to predict neuronal responses
detectors
Stimuli: TIMIT
speech database
(642 sentences)
A predictive coding model for auditory neurons
•
Modeling of auditory neuronal data
encoding
decoding
frequency channels
frequency channels
Predictive
fields are wider
than STRFs.
1
Encoding filters (STRF)
Background
narrow
Explaining Away: Dealing with overlapping stimuli
stimulus
neuronal responses
Decoding filters (predictive fields)
wider
2
STRF is not a
good predictor
of model neuron
1
New suppressive areas have emerged
2
Response became more precise in time
Rapid adaptation to stimulus statistics
Reconstruction of speech by model responses
Before
After
Before
average in time to get predictive fields
After
Model fit to neural data
frequency(kHz)
frequency(kHz)
Original
8
2kHz tone is partial evidence
for both X and Y
40 sec of tone
stimulation at 0.5 kHz
20 sec of tone
stimulation at 0.35 kHz
0.1
Model
500
1000
time (ms)
• Feature detectors predict
the sensory stimuli using
the generative model.
• Predictive fields tuned to speech rapidly adapt to tone
stimulation which causes specific changes in their STRFs.
Discussion
• Input from the receptors
(predictive fields, PF) are
divided 3 by the prediction
from other neurons
(receptive fields, RF).
A normative approach to auditory coding
Modeling of complex neuronal responses
• We derive an optimal inference model where auditory neurons
• Decoding filters obtained from auditory data are qualitatively
predict their input.
• Each neuron is selective to specific invariant features (PFs).
• The dynamics is given by
a Hidden Markov Model.
A dynamic Bayesian
inference model
8
frequency(kHz)
Highly overlapping
natural stimuli
Neuron x explains away the
2kHz + 1kHz combination and
suppresses y
Reconstruction
0.1
prior
x
likelihood
similar to predictive fields trained on speech data.
• This predicts that auditory responses are shaped as much by other
• We learn these features by training the model on speech data.
neurons’ responses and selectivity as by the stimulus itself.
• PFs are strongly overlapping and highly structured as opposed
• This accounts for the strong inhibitory lobes in STRFs and high
to STRFs which are narrow band-pass filters.
temporal precision of neural responses.
• Fit of a model neuron (blue) to a single neuron data (black): cc ≈ 0.43.
• Much better fit (blue vs. red) when there are more neurons in the network:
The network introduces sharper responses through inhibition.
References
1.
2.
3.
4.
Aertsen, A., and Johannesma, P. Biological cybernetics 1981, 42, 133-143.
Lochmann T., Ernst UA., Deneve S. The Journal of Neuroscience 2012, 32(12):4179-4195.
Deneve S: Bayesian spiking neurons I: Inference. Neural computation 2008, 20(1):91-117.
Mesgarani N., David SV., Fritz JB. and Shamma SA. J Neurophysiol . 2009 ,102:3329-3339.
Acknowledgment
We thank Nima Mesgarani for sharing his auditory data and Matthew Chalk for sharing
his code and helpful discussions.
This work was funded by:
ERC "Predispike" ERC-2012-StG_312227,
James MacDonnell Foundation award, "human cognition“ and ANR-10-LABX-0087 IEC