A proposal for the function of canonical microcircuits André Bastos July 5th, 2012 Free Energy Workshop Outline • Review of canonical (cortical) microcircuitry (CMC) • Role of feedback connections – Driving or modulatory? – Excitatory or inhibitory? • Recapitulation of free energy principle – Derive the predictive coding CMC • Empirical vs. predictive coding CMC • Frequency dissociations in the CMC What does a CMC need to do, in principle? • Amplify weak inputs from thalamus or other cortical areas – LGN provides only 4% of all synapses in V1 granular layer • Maintain a balance of excitation and inhibition • Select meaningful signals from a huge number of inputs (on average 10,000 synapses onto a single PY cell) • Segregate outputs from and inputs to a cortical column A first proposal on the CMC • Amplify thalamic inputs through recurrent connections • Maintain a balance of exc./inh. • Segregate super/deep Douglas and Martin, 1991 Quantitative study of C2 barrel cortex Lefort et al., 2009 Information flow summarized Lefort et al., 2009 Spread of feedforward activity through the CMC L1 Extrastriate (V2) L2/3 4A/B 4Ca/B L5 L6 LGN Pulvinar LGN www.brainmaps.org Drivers vs. modulators The corticogeniculate feedback connection displays modulatory synaptic characteristics. This suggested that corticocortical feedback is also modulatory… Sherman and Guillery, 1998, 2011 The “straw man” • Feedforward connections are driving – V1 projects monosynaptically to V2, V3, V3a, V4, and MT – In all cases, when V1 is reversibly inactivated, neural activity in the recipient areas is strongly reduced or silenced (Girard and Bullier, 1989; Girard et al., 1991a, 1991b, 1992, Schmid et al., 2009) • Feedback connections are modulatory – Synaptic characterization of Layer 6 -> LGN feedback • Longstanding proposal: corticocortical feedback connections are also modulatory (not an unreasonable assumption) At least some feedback connections are not just modulatory… Feedforward connections A1->A2 Feedback connections A2->A1 De Pasquale and Sherman, 2011, Covic and Sherman, 2011 Feedback: inhibitory or excitatory? • On theoretical grounds, we would predict inhibitory – Higher-order areas predict activity of lower areas. When activity is predictable it evokes a weaker response due to inhibition induced by higher areas • Neuroimaging studies (repetition suppression, fMRI, MMN) suggests inhibitory role for feedback • Electrophysiology with cooling studies are mixed Inhibitory corticogeniculate and intrinsic feedback Silence V1 Stimulate V1 dLGN Olsen et al., 2012 Corticocortical feedback targets L1 Shipp, 2007 Inhibitory “hot spot” in L1 Meyer et al., 2011 L1 cells are functionally active and inhibit PY cells in L2/3 and L5/6 Shlosberg et al., 2006 Spread of feedback activity through the CMC L1 L2/3 Higher-order cortex 4A/B 4Ca/B L5 L6 LGN www.brainmaps.org Anatomical and functional constraints ??? canonical microcircuit for predictive coding ??? Predictive coding constraints The Free Energy Principle, summarized • Biological systems are homoeostatic – They minimise the entropy of their states • Entropy is the average of surprise over time – Biological systems must minimise the surprise associated with their sensory states at each point in time • In statistics, surprise is the negative logarithm of Bayesian model evidence – The brain must continually maximise the Bayesian evidence for its generative model of sensory inputs • Maximising Bayesian model evidence corresponds to Bayesian filtering of sensory inputs – This is also known as predictive coding What generative model does the brain use??? Hierarchical Dynamical Causal Models Advantage: Extremely general models that grandfather most parametric models in statistics and machine learning (e.g., PCA/ICA/State-space models) Output đĻ = đ đĨ, đŖ + đ§ Observation noise Hidden states đĨ = đ đĨ, đŖ + đ¤ State noise Inputs Friston, 2008 Sensations are caused by a complex world with deep hierarchical structure input v2 Level 0 Level 1 x2 v1 đŖ1 = đ đĨ 2, đŖ 2 + đ§ 2 (cause) (state) (cause) Level 0 x1 đĨ1 = đ đĨ1, đŖ1 + đ¤ 2 (state) s đ = … (sensation) A simple example: visual occlusion A simple example: visual occlusion Hierarchical causes on sensory data input v2 x2 v1 x1 s đĨ 2 = đ đĨ 2, đŖ 2 + đ¤ 2 đŖ1 = đ đĨ 2, đŖ 2 + đ§ 2 (cause) (state) (cause) (state) (sensation) Perception entails model inversion Hierarchical generative model īˇv(3) v (2) īˇx(2) x (2) īˇv(2) v (1) īˇ x(1) x (1) īˇ v(1) Recognition Dynamics Expectations: ī x( i ) īŊ D ī x( i ) ī īļ xīĨ ( i ) ī ī¸ ( i ) (1) vx(0) Prediction errors: Hierarchical generation īv( i ) īŊ D īv( i ) ī īļ vīĨ ( i ) ī ī¸ ( i ) ī ī¸ v( i īĢ1) ī¸ v( i ) īŊ ī (vi ) ( īv( i ī1) ī g ( i ) ( ī x( i ) , īv( i ) )) ī¸ x( i ) īŊ ī (xi ) (D ī x( i ) ī f ( i ) ( ī x( i ) , īv( i ) )) Mind meets matter… Hierarchical generative model Hierarchical predictive coding Bottom-up prediction errors īˇv(3) īˇx(2) īˇv(2) īˇ x(1) īˇ v(1) ī¸ v(1) ī¸ x(1) ī¸ v(2) ī¸ x(2) ī¸ v(3) v (2) x (2) v (1) x (1) (1) vx(0) ī v(0) ī x(1) ī v(1) ī x(2) ī v(2) Hierarchical generation Top-down predictions đ =đŖ (0) (0) = đđŖ Recognition Dynamics Canonical microcircuit for predictive coding Forward prediction error ī (vi īĢ1) ī¸ Expectations: Prediction errors: īv( i ) īŊ D īv( i ) ī īļ vīĨ ( i ) ī ī¸ ( i ) ī ī¸ v( i īĢ1) ī x( i ) īŊ D ī x( i ) ī īļ xīĨ ( i ) ī ī¸ ( i ) ī¸ v( i ) īŊ ī (vi ) ( īv( i ī1) ī g ( i ) ( ī x( i ) , īv( i ) )) ( i īĢ1) v g (i īĢ1) ī v( i ) ī x( i ) ī¸ v( i ) ī¸ x( i ) Forward prediction error ī¸ ( v ,i ) ī¸ x( i ) īŊ ī (xi ) (D ī x( i ) ī f ( i ) ( ī x( i ) , īv( i ) )) ī v( i ) f (i ) ī x( i ) g (i ) Backward predictions Backward predictions Canonical microcircuit from anatomy Canonical microcircuit from predictive coding Forward prediction error ī (vi īĢ1) ī¸v(i īĢ1) Forward prediction error ī¸ ( v ,i ) g (i īĢ1) ī v( i ) ī x( i ) ī¸ v( i ) ī¸ x( i ) ī v( i ) f (i ) ī x( i ) g (i ) Haeusler and Maass (2006) Backward predictions Bastos et al., (in review) Backward predictions Spectral asymmetries between superficial and deep cells Rate of change of units encoding expectation (send feedback) Prediction error units (send feedforward messages) 0.3 1 īˇ2 0.25 0.2 ī¸v(i īĢ1) (īˇ ) 2 superficial ī¸v(i īĢ1) 0.15 0.1 ī v( i ) ī x( i ) 0.05 0 0 x 10 20 40 60 80 frequency (Hz) 100 120 -4 ī¸ x( i ) ī¸ v( i ) 2 Fourier transform īv(i ) (īˇ ) īŊ 2 2 1 (i īĢ1) ī¸ ( īˇ ) v īˇ2 ī v( i ) ī 1 0 0 20 40 60 80 frequency (Hz) īĸ ī§ deep (i ) x 100 120 Different oscillatory modes for different layers V1 V2 V4 Buffalo, Fries, et al., (2011) Unpublished data We apologize, but cannot share this slide at this point Unpublished data We apologize, but cannot share this slide at this point Unpublished data We apologize, but cannot share this slide at this point Unpublished data We apologize, but cannot share this slide at this point Integration of top-down and bottom-up through oscillatory modes? alpha/beta gamma ??? ??? (đ+1) đđŖ = prediction error ΠđŖ (đ+1) precision (đđ state đ −đ đ+đ ) higher-level prediction Integration of top-down and bottom-up streams Forward prediction error ī (vi īĢ1) ī¸ Forward prediction error ī¸ ( v ,i ) ( i īĢ1) v g (i īĢ1) Backward predictions ī v( i ) ī x( i ) ī¸ v( i ) ī¸ x( i ) ī v( i ) f (i ) ī x( i ) g (i ) (đ+1) đđŖ = prediction error ΠđŖ (đ+1) precision Backward predictions (đđŖ đ state −đ đ+1 ) higher-level prediction Canonical microcircuits and DCM V1 (primary visual cortex) ī¸ (s) ī¸ (v ) V4 (extrastriate visual area) Feedback connections Feedforward connections Intrinsic connections ī¸ ( x) ī (v) ī (v) ī ( x) ī¸ (v ) ī ( x) local fluctuations local fluctuations Unpublished data We apologize, but cannot share this slide at this point Unpublished data We apologize, but cannot share this slide at this point Conclusions • Repeating aspects of cortical circuitry suggest a “canonical microcircuit” exists to perform generic tasks that are invariant across cortex • Traditional roles for feedback pathways are being challenged by newer data • Predictive coding offers a clear hypothesis for the role of feedback and feedforward pathways • Predicts spectral asymmetries which may be important for how areas communicate • In short: the function of CMCs may be to implement predictive coding in the brain • These predictions might soon be testable with more biologically informed (CMC) DCMs Acknowledgements • Julien Vezoli • Conrado Bosman, Jan-Mathijs Schoffelen, Robert Oostenveld • Martin Usrey, Ron Mangun • Pascal Fries • Rosalyn Moran, Vladimir Litvak • Karl Friston Behaviors of a realistic model for oscillations • Laminar segregation and independence of gamma and beta rhythms Roopun 2008 Where do HDMs come from? Friston 2008