Reinagel lectures 2006 Take home message about LGN 1. Lateral geniculate nucleus transmits information from retina to cortex 2. It is not known what computation if any occurs in the LGN 3. For white noise stimuli, responses are precise and reliable 4. PRECISION is the trial to trial jitter in spike TIMING (order 1msec) feed forward inhibition may be the mechanism of precise timing 5. RELIABILITY is the trial to trial variability in spike NUMBER (subpoisson) refractoriness may be the mechanism of reliable spike count 6. BURSTING in the LGN is a distinct biophysical phenomenon, of unknown importance. The *right* question to ask is whether the bursting state is visually primed and whether priming itself encodes information 7. We now have a visually behaving rodent prep to address all these questions Take home message about efficient coding 1. Natural scenes are full of spatial and temporal correlations 2. This suggests WHY center-surround RF's are GOOD: redundancy reduction 2. Test: LGN responses to natural scenes are decorrelated (whitened) 3. More generally: are natural scenes optimal stimuli? is this even the right question? www-biology.ucsd.edu/labs/reinagel/ Lateral Geniculate Nucleus Retina LGN Cortex Ramon y Cajal Hubel 1960 (alert cat) Hubel & Wiesel 1961 (anesthetized) What happens in the LGN? • spiking inputs • intrinsic properties • local circuits • cortical feedback Gating? Attention? Binding? Prediction testing? Nothing? Retina LGN Cortex Reinagel & Reid 2000 Repeat Luminance LGN response to purely temporal stimuli Descriptive questions: • how precise is the timing? • how reliable is the number? • are there internal patterns? In each case: • visual information? • mechanism of encoding? • mechanism of decoding? Reinagel & Reid, 2000 PSTH peaks are milliseconds wide Reinagel & Reid, 2002 Temporal patterns conserved across animals A PSTH 0.2 0.1 -0.1 0.6 0.9 0.2 0.2 0.3 -0.9 -0.3 -0.1 Norm'd t 0.5 0 1500 1600 1700 1800 1900 2000 Time (ms) Reinagel & Reid 2002 Mutual Information (bits/s) Temporal precision of visual information a b c 4 8 de 100 50 0 0.5 1 2 16 32 64 128 Precision of spike times used (ms) Theory of Shannon, 1948 Method of Strong et al., 1998 Result of Reinagel & Reid, 2000 Mechanisms Underlying Precise Timing Pouille & Scanzian 2001 Deterministic Mean 4 Variance 0 Poisson Mean 4 Variance 4 Spike Count: Trial to Trial Variability Measure of variability Variance in Spike # Mean Spike # Random (Poisson) =1 Deterministic =0 LGN vs. Poisson Model PSTH PSTH LGN Variability << Poisson Fano Factor 1 Poisson 0.75 0.5 0.25 0 0 100 LGN 500 bin size T (msec) 1000 Variability increases from retina to cortex Fano Factor at ~ 40 Hz 1 0 RGC LGN V1 Kara, Reinagel & Reid, 2000 When firing rate is high, variability is low LGN RGC 200 V1 Firing Rate 0 FF 1 0 0 500 Time (ms) Kara, Reinagel & Reid, 2000 Refractoriness Regularizes? PSTH Poisson model Poisson with Refractory Period probability Estimating refractoriness from data model Method: Berry & Meister 1998 data 0 10 20 30 40 50 ISI 60 70 80 90 100 Recovery Function recovery function 1 absolute and relative refractoriness 0.5 0 0 5 10 15 20 25 time since last spike (ms) 30 35 Free Firing Rate firing rate (sp/s) 500 free 400 300 200 100 0 observed 0 500 time (ms) 1000 Refractory models for all cell types Fano Factor LGN RGC 2 V1 1 0 200 0 0 Time (ms) Kara, Reinagel & Reid, 2000 Variability increases from retina to cortex Fano Factor at ~ 40 Hz 1 0 RGC LGN V1 Kara, Reinagel & Reid, 2000 Refractoriness decreases from retina to cortex Recovery Function 1.0 V1 0.5 RGC 0.0 0 10 20 30 Time (ms) Kara, Reinagel & Reid, 2000 Summary of Reliability • Spike count has sub-Poisson variability • High FR High Reliability • Refractoriness completely explains • Noise is low, but doubling each synapse - firing rate is decreasing - refractoriness is decreasing Thalamic Bursts (It) Hubel and Wiesel (1961) Jahnsen and Llinas (1984) Bursting in the LGN • dominate during sleep, when vision is suppressed • not rhythmic or synchronous in anesthetized animals • frequent under anesthesia, when vision is absent • synapses prefer bursts • almost never seen in alert animals, when vision is happening • visual in anesthetized animals • do occur in alert animals, and rare signals can be important • cool computational ideas ERGO ERGO Bursts are irrelevant to vision Bursts are crucial to vision Visual inputs trigger bursts Optimal Guess of Stimulus Before a burst 0.2 Before a tonic spike 0.1 0 -0.1 -0.6 -0.4 -0.2 0 Time before spike (s) Coding Efficiency Bits/event 1.5 1 0.5 0 Burst Tonic 0.2 0.15 0.1 0.05 0 Burst Tonic Reinagel, Godwin, Sherman & Koch 1999 AP times Bursts: Triggering vs. Priming * •• • LT-Ca++ channel state Trigger synaptic input Ca++ spike observable • •• • • active inactive time Bursts in LGN are distinct code words Denning & Reinagel 2005 Alitto, Weyand & Usrey 2005 Lesica & Stanley 2004 Summary: Bursting • LGN neurons have 2 states • Visual inputs trigger responses in both states • Visual inputs also control the state BUT All this is under anesthesia What about alert? - Stimulus ensemble matters - Behavioral state may also - Triggering and priming What happens in the LGN? • spiking inputs • intrinsic properties • local circuits • cortical feedback Directions • Do bursts occur and are they visual in alert animals? • Function of cortical feedback to the LGN? • Does precision in the LGN matter for perception? An awake behaving rodent prep for vision Thanks to collaborators at CSHL Flister, Meier , Conway & Reinagel (unpub) Bursts in LGN in the awake, behaving rat Flister, Meier & Reinagel (unpub) [break] Center Surround Opponent RFs Kuffler 1958 Natural scenes are spatially correlated Spatial correlations in unnatural images Spatial correlation in natural images Natural Image 1 Correlation Power spectrum 2 10 0.8 0.6 0 10 0.4 0.2 -40 -2 -20 0 20 distance 40 10 0 10 102 cycles/degree (cf. Field 1987; Tadmore & Tolhurst; Ruderman & Bialek; van Hateren) Natural Image Correlation Power spectrum 4 1 10 0.8 2 10 0.6 0.4 0 10 0.2 0 -2 -20 0 20 10 -2 10 0 10 2 10 4 1 10 0.8 2 10 0.6 0.4 0 10 0.2 0 -2 -20 0 20 Distance (pixels) 10 -2 10 0 10 2 10 Spatial frequency (cf. Barlow 1961) luminance Natural temporal stimulus 0 10 -1 10 -2 10 0 1 2 3 4 time (s) Correlation Power Spectrum 2 1 10 0.8 0 10 0.6 -2 10 0.4 -4 10 0.2 0 -1 -6 0 Distance (sec) 1 10 -1 10 0 10 1 10 2 10 3 10 Temporal frequency (Hz) (cf. Dong & Atick 1995; van Hateren 1997) luminance 0 10 -1 10 -2 10 0 1 2 3 4 time (s) Correlation Power Spectrum 1 0 0.8 10 0.6 0.4 -5 10 0.2 0 -2 -1 0 1 distance (sec) 2 -1 10 0 10 1 10 2 10 3 10 Temporal frequency (Hz) (cf. Dan Atick & Reid 1996) Barlow 1961 Redundancy Reduction Hypothesis + + Sensory neurons decorrelate natural inputs to reduce redundancy Dan, Atick & Reid 1996 Dan, Atick & Reid 1996 Whitening in the fly Van Hateren 1997 Summary: Redundancy Reduction • Shannon 1948: Optimal codes lack redundancy • Kuffler 1958: Center-surround receptive fields in retina Hubel 1960: Center-surround RFs in LGN • Barlow 1961: Center-surround RFs reduce redundancy for natural scenes • Dan, Atick & Reid 1996: Responses in LGN are less redundant for natural scenes Bullfrog Auditory Neuron: Natural Stimulus is ‘Optimal’ B 300 C 7 6 Information (bits/spk) 250 0.8 5 200 0.6 4 150 3 100 0 0.4 2 50 0.2 1 white natural 0 1 Efficiency (bits/bit) Information (bits/s) A white natural 0 white natural Rieke, Bodnar & Bialek 1995 Cat LGN Neuron: Opposite result B 60 40 20 0 white natural C 3 Information (bits/spk) Information (bits/s) 80 2.5 2 1.5 1 0.5 0 0.6 Efficiency (bits/bit) A 0.5 0.4 0.3 0.2 0.1 white natural 0 white natural Analog LED Stimuli Reinagel & Reid, in prep. LGN Responses to full field temporal stimuli 105 Natural visual stimulus Power 104 103 Random visual stimulus 102 101 10 0 101 102 temporal frequency (Hz) 103 (replication of Dan et al 1996)