A computational Approach to Behavior

Transcription

1 A computational Approach to Behavior Pietro Perona California Institute of Technology 26 March 2017 AAAI Symposium on ``Computational Principles of Natural and Artificial Intelligence Stanford University

2 California gull vs. alkali fly (Larus californicus vs Ephydra hians) Mono Lake, CA N, '33.4"W

3 brain+ body development, regulation genes perception learning decision, control behavior natl. selection gene expression environment

4 construction HW+ SW design perception learning decision, control behavior analysis of performance environment

5 What is `behavior?

6 6

7 Zidane in a dish

8 Courtship in Drosophila Marla B. Sokolowski Nature Reviews Genetics 2, (2001)

9

10

11 Phenomenology: many scales 1 day 1 hour 1 minute 1 second 1 frame hike walk court dine flirt open door have drink write `A step reach glance 10 pix 100 pix 1000 pix movemes actions activities

12 Parametrization

13 [Eyjolfsdottir, 2014]

14

15

16 Ethograms [Dankert et al., Nature Methods, April 2009] by hand it would have taken ( )*1/3 * 3 * 7 = 400h

17 Levels of enlightenment Description Phenomena Mechanisms

18 FINDING AGGRESSION NEURONS IN THE FLY X X=ion channel gene LIBRARIES OF Drosophila STRAINS WITH DIFFERENT GENETICALLY LABELED POPULATIONS OF NEURONS warmth-sensitive ion channel 27 o C hyperaggressive?

19 Analysis by synthesis [Braitenberg, 1984]

20 MOTOR PLANNING Action Perception group-level goals and plans SOCIAL NETWORK individual goals and plans PREFRONTAL CORTEX INDIVIDUAL interaction, cooperation, competition SOCIAL INTERACTION plans, goals, behavior, relationships... COGNITION REPRESENTATIONAL motor programs MOTOR CORTEX, BASAL GANGLIA etc. sensor-based control SPINAL CORD pose, movemes, actions, activities, objects, scenes RECOGNITION images, trajectories IMAGING,TRACKING SENSORY World

21 Levels of analysis Description Phenomena Grammar Purpose Mechanisms Performance Ontogeny Phylogeny

22 From phenomena to mechanisms: a simple example

23 [Branson et al. Nature Methods, Jun. 09]

24

25 [Branson et al. Nature Methods, Jun. 09]

26 T-Stops

27 X-Stops

28 An engineers solution Wang et al

29 Try implementing it with 10 5 neurons

30

31

32 Regressive motion

33

34 Experiment [Zabala 12]

35 Model [Chalupka 15]

36

37 Regressive Looming

38 Levels of enlightenment Description - fly trajectories Phenomena - stops, chases, (actions) Grammar - T-stops,X-stops, (interactions) Purpose - Avoid predation and collisions Mechanisms - regressive motion(implem.?) Performance - Mobility,safety Ontogeny - Learning?? Phylogeny -??

39 Mechanism inference [Eyjolfsdottir 2016,2017]

40 Fly-centric features flies walls eat v fight reproduce fwd h x yaw side escape len w-len w-ang v: sensory input x: motor control

41 Model architecture yi ^ h 2 i-1 ^ h 2 i ^ h 2 i+1 generative h 2 i-1 ^ yi h 2 i h 2 i+1 discriminative ^ h 1 i-1 ^ h 1 i ^ h 1 i+1 h 1 i-1 h 1 i h 1 i+1 ^ xi-1 ^ xi ^ xi+1 vi-1 xi-1 vi xi vi+1 xi+1 Simulation: - prediction as input - compute vi on the fly [Eyjolfsdottir 2016,2017]

42 Sanity check: learning on a synthetic fly

43 Can the model learn generative control laws? SynthFly Simulation Control laws: 1) walk forward with small randomness 2) rotate when close to boundary (away from higher visibility) 3) walk towards obstacle, when visible 4) extend wing when half way towards it 5) rotate when close to obstacle (alternate between left and right) 6) repeat from 1) obstacle visible

44 Interesting units hidden unit 91 hidden unit 175 left/right wing angle distance to object turn synthetic fly r l r l r l r l r l r l r l r l r time simulation r l r l l l l l r l time r l r l l l l l r l r l r l l l l l r l r l r l l l l l r l

45 Classification

46

47 Discovery

48 Discovery: FlyBowl tsne dimensionality reduction female male right wing extension left wing extension h l i: hidden states of unsupervised model

49 Discovery: IAM-OnDB stroke length writer identity

50 Prediction

51

52

53 Simulation

54 fly vision chamber vision Simulation: FlyBowl

55 Simulation: FlyBowl simulated real

56 Simulation - handwriting generated activate text character injections e s r

57 Summary Brains, behavior and intelligence Description Phenomena Mechanisms Relationship with AI?

58 Collaborators Heiko Dankert Kristin Branson Xavi Burgos Piotr Dollar Eyrún Eyjólfsdóttir Cristina Segalin Grant van Horn David Anderson Michael Dickinson + Carlos Gonzalez Shay Ohayon Roian Egnor Michael Maire John Bender Tim Lebestky Alice Robie Erik Hoopfer