Wellcome Trust Centre for Neuroimaging University College London Fifth Conference on Mismatch Negativity (MMN) and its Clinical and Scientific Applications April 4-7, 2009 | Budapest, Hungary A-0115 The MMN and perception K Friston Wellcome Trust Centre for Neuroimaging University College London This talk will try to frame the MMN in terms of perceptual learning and inference. We start with a statistical formulation of Helmholtz’s ideas about neural energy to furnish a model of perception that can explain a remarkable range of neurobiological facts. Using constructs from statistical physics it can be shown that the problems of inferring what caused our sensory inputs and learning causal regularities in the sensorium can be resolved using exactly the same principles. Furthermore, inference and learning can proceed in a biologically plausible fashion. The ensuing scheme rests on empirical Bayes and hierarchical models of how sensory information is generated. The use of hierarchical models enables the brain to construct prior expectations in a dynamic and context-sensitive fashion. This scheme provides a principled way to understand many aspects of the brain’s organization and responses. Using this framework we can simulate evoked responses to repeated stimuli and examine the within and between-trial changes (reflecting inference and learning respectively). These simulations suggest that learning in lower levels of the sensory hierarchy may explain early effects (c.f., N1 enhancement), whereas changes in higher levels may underlie later MMN components proper.

Overview Inference and learning under the free energy principle Hierarchical Bayesian inference Bird songs (inference) Perceptual categorisation Prediction and omission Bird songs (learning) Repetition suppression The mismatch negativity

The free-energy principle Sensation External states Internal states environment Action agent - m Action to minimise a bound on surprise Perception to optimise the bound

Neuronal implementation Perception and inference Learning and memory Activity-dependent plasticity Synaptic activity Synaptic efficacy Functional specialization Attentional gain Enabling of plasticity Synaptic adaptation Attention and salience

Forward prediction error Neuronal hierarchies Forward prediction error SG L4 IG Backward predictions Synaptic efficacy Synaptic adaptation Synaptic activity

Overview Inference and learning under the free energy principle Hierarchical Bayesian inference Bird songs (inference) Perceptual categorisation Prediction and omission Bird songs (learning) Repetition suppression The mismatch negativity

Synthetic song-birds Syrinx Higher vocal centre Sonogram 0.2 0.4 0.6 0.8 2000 2500 3000 3500 4000 4500 5000 time (seconds)

Forward prediction error Hierarchical recognition 10 20 30 40 50 60 -5 5 15 prediction and error time Backward predictions 10 20 30 40 50 60 -10 -5 5 15 causes time (bins) percept 0.2 0.4 0.6 0.8 2000 2500 3000 3500 4000 4500 5000 time (seconds) Forward prediction error 10 20 30 40 50 60 -5 5 15 hidden states time

Categorizing sequences 90% confidence regions Frequency (Hz) Song A 0.2 0.4 0.6 0.8 2000 3000 4000 5000 time (seconds) Song B Song C 0.2 0.4 0.6 0.8 1 -20 -10 10 20 30 40 50 Categorizing sequences 90% confidence regions Song A Song B Song C time (seconds) 10 15 20 25 30 35 1 1.5 2 2.5 3 3.5 Song A Song B Song C

Overview Inference and learning under the free energy principle Hierarchical Bayesian inference Bird songs (inference) Perceptual categorisation Prediction and omission Bird songs (learning) Repetition suppression The mismatch negativity

Sequences of sequences Neuronal hierarchy Syrinx sonogram Frequency (KHz) 0.5 1 1.5 Time (sec)

Hierarchical perception 20 40 60 80 100 120 -20 -10 10 30 50 prediction and error Hierarchical perception 5000 percept 4500 4000 Prediction error encoded by superficial pyramidal cells that generate ERPs 3500 3000 2500 2000 0.5 1 1.5 time (seconds) time 5000 stimulus 4500 4000 3500 3000 2500 20 40 60 80 100 120 -10 10 30 50 causes time 20 40 60 80 100 120 -20 -10 10 30 50 hidden states time hidden states 2000 0.5 1 1.5 50 time (seconds) 40 30 20 10 -10 -20 20 40 60 80 100 120 time

omission and violation of predictions Stimulus but no percept Percept but no stimulus

Overview Inference and learning under the free energy principle Hierarchical Bayesian inference Bird songs (inference) Perceptual categorisation Prediction and omission Bird songs (learning) Repetition suppression The mismatch negativity

Repetition suppression and the MMN The MMN is an enhanced negativity seen in response to any change (deviant) compared to the standard response. Main effect of faces Suppression of inferotemporal responses to repeated faces Henson et al 2000

Hierarchical learning Synaptic adaptation Synaptic efficacy 10 20 30 40 50 60 -0.2 -0.1 0.1 0.2 0.3 0.4 0.5 causes time Time (sec) Frequency (Hz) 0.1 0.2 0.3 0.4 2000 2500 3000 3500 4000 4500 5000 10 20 30 40 50 60 -5 5 15 25 35 prediction and error time 10 20 30 40 50 60 -2 -1 1 2 hidden states time

Simulating ERPs to repeated chirps hidden states percept prediction error 5 5000 10 Simulating ERPs to repeated chirps 4000 Frequency (Hz) LFP (micro-volts) 3000 -10 -5 2000 0.2 0.4 0.1 0.2 0.3 100 200 300 5 10 4000 Frequency (Hz) LFP (micro-volts) -10 -5 2000 0.2 0.4 0.1 0.2 0.3 100 200 300 5 10 4000 Frequency (Hz) LFP (micro-volts) Perceptual inference: suppressing error over peristimulus time Perceptual learning: suppression over repetitions -10 -5 2000 0.2 0.4 0.1 0.2 0.3 100 200 300 5 4000 10 Frequency (Hz) LFP (micro-volts) -10 -5 2000 0.2 0.4 0.1 0.2 0.3 100 200 300 5 10 4000 Frequency (Hz) LFP (micro-volts) -10 -5 2000 0.2 0.4 0.1 0.2 0.3 100 200 300 5 4000 10 Frequency (Hz) LFP (micro-volts) -10 -5 2000 0.2 0.4 0.1 0.2 0.3 100 200 300 Time (sec) Time (sec) peristimulus time (ms)

Synthetic MMN Synaptic efficacy Synaptic gain First presentation 3 6 Synthetic MMN 2.5 5 2 4 changes in parameters 1.5 hyperparameters 3 1 2 0.5 1 1 2 3 4 5 1 2 3 4 5 presentation presentation 10 10 10 10 10 primary level prediction error First presentation (before learning) -10 -10 -10 -10 -10 Last presentation (after learning) 100 200 300 100 200 300 100 200 300 100 200 300 100 200 300 0.2 0.2 0.2 0.2 0.2 secondary level -0.2 -0.2 -0.2 -0.2 -0.2 -0.4 -0.4 -0.4 -0.4 -0.4 100 200 300 100 200 300 100 200 300 100 200 300 100 200 300 primary level (N1/P1) secondary level (MMN) 20 0.4 15 0.2 10 5 Difference waveform Difference waveform -0.2 -5 -0.4 -10 -15 -0.6 100 200 300 400 100 200 300 400 peristimulus time (ms) peristimulus time (ms)

Intrinsic connections Synaptic efficacy Synaptic gain Extrinsic connections Intrinsic connections 3 6 200 200 180 180 2.5 5 160 160 2 4 140 140 120 120 changes in parameters 1.5 hyperparameters 3 100 100 80 80 1 2 60 60 0.5 1 40 40 20 20 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 presentation presentation presentation presentation A1 STG subcortical input repetition effects Synthetic and real ERPs

Summary A free energy principle can account for several aspects of action and perception The architecture of cortical systems speak to hierarchical generative models Estimation of hierarchical dynamic models corresponds to a generalised deconvolution of inputs to disclose their causes This deconvolution can be implemented in a neuronally plausible fashion by constructing a dynamic system that self-organises when exposed to inputs to suppress its free-energy