A new path to understanding vision

Transcription

1 A new path to understanding vision from the perspective of the primary visual cortex Frontal brain areas Visual cortices Primary visual cortex (V1) Li Zhaoping Retina

2 A new path to understanding vision Traditional paths to understanding vision (1) Low level vision, mid-level vision, high-level vision (2) David Marr: primal sketch, 2.5 d sketch, 3-d model. Li Zhaoping July 16, 2018, presented at APCV 2018, HangZhou, China

3 Talk outline (1) The functional role of the primary visual cortex (V1) (2) In light of V1 s role a new plan to understanding vision (3) A first example study in this new plan

4 1953, Stephen Kuffler Center-surround feature detectors in retina

5 1960s-70s, Hubel and Wiesel Bar/edge feature detectors in V1 1953, Stephen Kuffler Center-surround feature detectors in retina

6 Some standard models of neural encoding in retina/v1: From Fig. 1 of Carandini et al 2005 Explaining: Retinal receptive fields V1 properties: Orientation tuning, Color tuning Motion direction tuning Spatial frequency tuning Ocular dominance (eye-of-origin tuning) Simple/Complex cells Explaining 80% of the response variance Explaining only 35% of the response variance!! Plus: models of gain control, response normalization by population activities

7 Do we know early vision? What is V1 doing? Do we really know what the early visual system does? Carandini, Demb, Mante, Tolhurst, Dan, Olshausen, Gallant, Rust, 2005 Standard models of neurons in the retina account for 80% of the response variance. However, standard models of V1 capture only 35% of the V1 variances. How close are we to understand V1? Olshausen and Field Standard models combining linear filtering, rectification and squaring, and response Is V1 normalization doing sparse still miss coding 86% of the (more variances efficient in V1 responses. encoding than in retina)? Olshausen and Field 1996: Emergence of simple-cell receptive field properties by learning a sparse coding for natural images Bell and Sejnowski 1997: The independent components of natural scenes are edge filters. Simoncelli and Olshausen 2001: Natural image statistics and neural representation.

8 Do we know early vision? What is V1 doing? Do we really know what the early visual system does? Carandini, Demb, Mante, Tolhurst, Dan, Olshausen, Gallant, Rust, 2005 Standard models of neurons in the retina account for 80% of the response variance. However, standard models of V1 capture only 35% of the V1 variances. How close are we to understand V1? Olshausen and Field Standard models combining linear filtering, rectification and squaring, and response Is V1 normalization doing sparse still miss coding 86% of the (more variances efficient in V1 responses. encoding than in retina)? Other than stereo vision, why not make encoding more efficient in retina already? V1 has ~100 as many cells as retina (Barlow 1981), quite inefficient/redundant! What about contextual influences to neural responses??!!!!

9 The primary visual cortex (V1) 1953, Stephen Kuffler, retina, Hubel and Wiesel, V1 Then Experimentally: V1 and beyond Theoretical/modelling, Reichardt, Marr, etc. 2005: How close are we to understand V1? Olshausen and Field 2005 Do we really know what the early visual system does? Carandini, Demb, Mante, Tolhurst, Dan, Olshausen, Gallant, Rust, 2005 Standard models of V1 neural receptive field combining linear filtering, rectification and squaring, and response normalization captures only 15-35% of the variances in V1 responses. 2012, David Hubel, in answer to What Do You Feel Are the Next Big Questions in the Field? We have some idea for the retina, the lateral geniculate body, and the primary visual cortex, but that s about it. (Hubel & Wiesel 2012, Neuron)

10 The primary visual cortex (V1) 1953, Stephen Kuffler, retina, Hubel and Wiesel, V1 Then Experimentally: V1 and beyond Theoretical/modelling, Reichardt, Marr, etc. 2005: How close are we to understand V1? Olshausen and Field 2005 Do we really know what the early visual system does? Carandini, Demb, Mante, Tolhurst, Dan, Olshausen, Gallant, Rust, 2005 Standard models of V1 neural receptive field combining linear filtering, rectification and squaring, and response normalization captures only 15-35% of the variances in V1 responses. 2012, David Hubel, in answer to What Do You Feel Are the Next Big Questions in the Field? We have some idea for the retina, the lateral geniculate body, and the primary visual cortex, but that s about it. (Hubel & Wiesel 2012, Neuron) Questions: Is a lack of understanding of V1 hindering our progress beyond V1? Physiologically Functionally (in behaviour)?

11 Information bottlenecks in the visual pathway: 10 7 bits/second ~ 10 6 neurons, ~10 spikes/neuron ~1 bit/spike To be or not to be, This is the question.. 40 bits/second (Sziklai, 1956) 10 9 bits/second (Kelly 1962) ~ 25 frames/second, 2000x2000 pixels, 1 byte/pixel

12 Information bottlenecks in the visual pathway: We are nearly blind! Vision ~ Looking (selecting ) + Seeing top-down vs. bottom-up selection 10 7 bits/second ~ 10 6 neurons, ~10 spikes/neuron ~1 bit/spike To be or not to be, This is the question.. 40 bits/second (Sziklai, 1956) Task: find a uniquely oriented bar 10 9 bits/second (Kelly 1962) ~ 25 frames/second, 2000x2000 pixels, 1 byte/pixel

13 Information bottlenecks in the visual pathway: Questions: which brain areas are doing the selection? Frontal? Parietal? top-down vs. bottom-up selection Task: find a uniquely oriented bar 10 9 bits/second (Kelly 1962) ~ 25 frames/second, 2000x2000 pixels, 1 byte/pixel 10 7 bits/second ~ 10 6 neurons, ~10 spikes/neuron ~1 bit/spike To be or not to be, This is the question.. 40 bits/second (Sziklai, 1956)

14 Information bottlenecks in the visual pathway: Questions: which brain areas are doing the selection? Saliency regardless of visual features Frontal? Parietal? Koch & Ullman 1985, Itti & Koch 2001, etc top-down vs. bottom-up selection 10 7 bits/second ~ 10 6 neurons, ~10 spikes/neuron ~1 bit/spike To be or not to be, This is the question.. 40 bits/second (Sziklai, 1956) Task: find a uniquely oriented bar 10 9 bits/second (Kelly 1962) ~ 25 frames/second, 2000x2000 pixels, 1 byte/pixel

15 Eye Extrastriate cortices V 1 LGN Parietal cortex Thalamus Superior colliculus Brain stem, etc to control eye muscle Frontal eye field In lower mammals, a smaller percentage of the total cortex is for prefrontal areas (Fuster 2002) Non-mammals do not have a neocortex Forebrain of fish is more for olfaction than vision (Butler & Hodos 2005) 11.5% 17% 3.5 % 12.5% 29%

16 The V1 Saliency Hypothesis: A bottom-up saliency map in the primary visual cortex (Li 1999, 2002) Retina inputs Saliency map V1 firing rates (highest at each location) = Winner-take-all Superior colliculus

17 The V1 Saliency Hypothesis: A bottom-up saliency map in the primary visual cortex (Li 1999, 2002) Retina inputs V1 firing rates (highest at each location) Winner-take-all Superior colliculus

18 The V1 Saliency Hypothesis: Neural activities as A bottom-up saliency map in the primary visual cortex (Li universal 1999, 2002) currency to bid for visual selection. Retina inputs V1 firing rates (highest at each location) Winner-take-all Superior colliculus

19 The V1 Saliency Hypothesis: A bottom-up saliency map in the primary visual cortex (Li 1999, 2002) Retina inputs V1 V1 firing rates (highest at each location) Winner-take-all Superior colliculus Bosking et al 1997 iso-feature suppression (Blakemore & Tobin 1972, Gilbert & Wiesel 1983, Rockland & Lund 1983, Allman et al 1985, Hirsch & Gilbert 1991, Li & Li 1994, etc)

20 The V1 Saliency Hypothesis: A bottom-up saliency map in the primary visual cortex (Li 1999, 2002) Retina inputs V1 V1 firing rates (highest at each location) Winner-take-all Iso-orientation suppression Iso-color suppression Bosking et al 1997 Bosking et al 1997 iso-feature suppression (Blakemore & Tobin 1972, Gilbert & Wiesel 1983, Rockland & Lund 1983, Allman et al 1985, Hirsch & Gilbert 1991, Li & Li 1994, etc Wachtler et al 2003 Jones et al 2001 ) Superior colliculus Iso-motion (direction) suppression

21 The V1 Saliency Hypothesis: A bottom-up saliency map in the primary visual cortex (Li 1999, 2002) Retina inputs V1 V1 firing rates (highest at each location) Winner-take-all Superior colliculus Bosking et al 1997 Left eye image Right eye image A surprising prediction: an invisible feature attract attention! (Zhaoping 2008, 2012) Escapes iso-orientation suppression Fused perception Escapes iso-eye-of-origin suppression Looking without seeing! Replicated by multiple research groups since! A fingerprint of V1

22 Look for saliency signals in the brain without the awareness/top-down confound Collaboration with Fang Fang (Beijing University), and his student Xilin Zhang (Neuron, 2012) Salient orientation pop-out Quickly masked to prevent awareness (verified) Vernier task probe, may be at cued location or the other hemi-field

23 Look for saliency signals in the brain without the awareness/top-down confound Collaboration with Fang Fang (Beijing University), and his student Xilin Zhang (VSS 2011) Cueing effect Now measure which brain areas are activated by this invisible cue? By fmri, ERP V1, but not IPS, FEF!! i.e.,saliency effect without awareness Orientation contrast at cue

24 Testing the V1 theory on behaving monkeys --- Yan, Zhaoping, & Li, Submitted. Saccade to an uniquely oriented bar ASAP V1 neural responses to input stimulus (spikes/sec) for trials with faster saccades Receptive field for trials with slower or failed saccades Time since visual input onset (second) Quantitative, zero-parameter, predictions from theory Imply that higher visual areas are not involved Solid curve ---Planck s law Squares: --- data points Koene & Zhaoping 2007 Zhaoping & Zhe 2015,

25 Talk outline (1) The functional role of the primary visual cortex (V1) Attentional selection (2) In light of V1 s role a new plan to understanding vision (3) A first example study in this new plan

26 Information bottlenecks in the visual pathway: 10 7 bits/second ~ 10 6 neurons, ~10 spikes/neuron ~1 bit/spike Bottom-up selection To be or not to be, This is the question.. 40 bits/second (Sziklai, 1956) 10 9 bits/second (Kelly 1962) ~ 25 frames/second, 2000x2000 pixels, 1 byte/pixel

27 A new path to understanding vision 20 frames, 20 megabytes/second 40 bits/second Traditional paths to understanding vision (1) Low level vision, mid-level vision, high-level vision To be or not to be, This is the question.. 40 bits/second (Sziklai, 1956) (2) David Marr: primal sketch, 2.5 d sketch, 3-d model bits/second (Kelly 1962) ~ 25 frames/second, 2000x2000 pixels, 1 byte/pixel Bottom-up selection

28 Talk outline (1) The functional role of the primary visual cortex (V1) Attentional selection (2) In light of V1 s role a new plan to understanding vision (3) A first example study in this new plan

29 A new path to understanding vision 20 frames, 20 megabytes/second Looking periphery 40 bits/second Seeing center

30 A new path to understanding vision 20 frames, 20 megabytes/second Looking periphery 40 bits/second Seeing center

31 Looking and seeing --- peripheral and central Two separate processes Demo: (Zhaoping & Guyader 2007) V1: tuned to primitive bars IT & higher areas: X shape recognition, rotationally invariant Add a horizontal bar or a vertical bar to each oblique bar

32 Looking and seeing --- peripheral and central Two separate processes Demo: (Zhaoping & Guyader 2007) Add a horizontal bar or a vertical bar to each oblique bar Add a horizontal bar or a vertical bar to each oblique bar

33 Display span 46x32 degrees in visual angle --- condition A

34 Gaze arrives at target after a few saccades Looking --- mainly by the bottom up saliency of the unique orientation. Then...

35 Gaze dawdled around the target, then abandoned and returned. Seeing

36 A new path to understanding vision 20 frames, 20 megabytes/second Looking 40 bits/second Seeing Peripheral vision Must qualitatively differ in Seeing Central vision Computational algorithms? Focus on feedforwardfeedback / V1 A demo of crowding in the periphery

37 Summation percept dominant Perception =? In V1, signals are efficiently encoded by these two de-correlated channels (Li & Atick 1994) Ocular sum Ocular diff Subject task: report the perceived tilt. C = C Left eye Right eye Zhaoping 2017

38 Why does perception prefer ocular summation? (Zhaoping 2017) Higher areas for analysis-by-synthesis (c.f. predictive coding) Feedforward, feedback, verify, and re-weight (FFVW) If I perceive it, it is likely (prior) shown to both my left and my right eyes, so it should resemble the input in the sum channel!!! The Bayesian(?) monster V1 Feedback to verify Ocular sum Ocular diff Retina Left eye Right eye

39 Why does perception prefer ocular summation? (Zhaoping 2017) Higher areas for analysis-by-synthesis (c.f. predictive coding) Feedforward, feedback, verify, and re-weight (FFVW) If I perceive it, it is likely (prior) shown to both my left and my right eyes, so it should resemble the input in the sum channel!!! The Bayesian(?) monster V1 Feedback to verify Ocular sum Ocular diff Not because periphery cannot see tilt! Retina Left eye Right eye Proposal: Top-down feedback to V1 is weaker or absent in peripheral vision for analysis-by-synthesis!

40 Why does perception prefer ocular summation? (Zhaoping 2017) Higher areas for analysis-by-synthesis (c.f. predictive coding) Feedforward, feedback, verify, and re-weight (FFVW) If I perceive it, it is likely (prior) shown to both my left and my right eyes, so it should resemble the input in the sum channel!!! The Bayesian(?) monster V1 Feedback to verify Ocular sum Ocular diff Retina Left eye Right eye Proposal: Top-down feedback to V1 is weaker or absent in peripheral vision for analysis-by-synthesis!

41 A central disk (non-zero disparity) and a surrounding ring (zero disparity) Testing it in depth perception Zhaoping & Ackermann, 2018 Dots for the central disk correlated A V1 neuron s response to disparity in random dot stereograms Anticorrelated Correlated Dots for the central disk Anti-correlated Disparity (degrees) (Cumming and Parker 1997 ) Proposal: Top-down feedback to V1 is weaker or absent in peripheral vision for analysis-by-synthesis!

42 Proposal: Top-down feedback to V1 is weaker or absent in peripheral vision for analysis-by-synthesis! Dots for the central disk correlated A central disk (non-zero disparity) and a surrounding ring (zero disparity) Testing it in depth perception Zhaoping & Ackermann, 2018 No reversed depth percept V1 s report vetoed prior expects correlated inputs between the eyes Dots for the central disk Anti-correlated Top-down feedback to verify V1 s report V1 feeds forward reverse depth to higher brain areas!

43 Proposal: Top-down feedback to V1 is weaker or absent in peripheral vision for analysis-by-synthesis! Dots for the central disk correlated A central disk (non-zero disparity) and a surrounding ring (zero disparity) Testing it in depth perception Zhaoping & Ackermann, 2018 No reversed depth percept If peripheral vision has no feedback Dots for the central disk Anti-correlated V1 feeds forward reverse depth to higher brain areas!

44 Proposal: Top-down feedback to V1 is weaker or absent in peripheral vision for analysis-by-synthesis! A central disk (non-zero disparity) and a surrounding ring (zero disparity) Testing it in depth perception Zhaoping & Ackermann, 2018 Dots for the central disk correlated Accuracy reporting disparity-defined depth Dots for the central disk Anti-correlated

45 Thanks to Funding source: Gatsby Foundation, EPSRC, BBSRC. Collaborators: Wu Li, Yin Yan, Ansgar Koene, Li Zhe, Joealle Ackermann, etc A new path to understanding vision Falsifiable 20 frames, 20 megabytes/second 40 bits/second Opening the window to our brain Theory motivate inspire Experiments