Visual Reasoning in Human Brain Reasoning strategy Performance

Visual Reasoning in Human Brain Reasoning strategy Performance improvement of 3 D analysis 1

Outline • Visual reasoning and response • Some examples • Principle and implementation • Cognitive science • Engineering approximation • Summary and conclusion * Framework of PM • Keyword: vision, content(cause) vs. context, architecture, Bayesian inference, predictive coding, free-energy, generative model, neuronal dynamic 2

Sophisticated Processing behind the Human Vision Hollow Face Illusion *Ref: https: //www. youtube. com/watch? v=Baofyu. CXZ_0, https: //www. youtube. com/watch? v=1 ass. N_WJCSE 3

Key Components: Content, Context & Likelihood *The hollow face illusion *Ref: Reflections on agranular architecture: predictive coding in the motor cortex 4

3 Stages in the Active Inference Flow *Ref: Reflections on agranular architecture: predictive coding in the motor cortex 5

Neurobiological Implementation Context Content *Ref: Cerebral hierarchies: predictive processing, precision and the pulvinar. 7

How to Make an Active Inference – an Toy Example Generative Model *Ref: https: //nl. wikipedia. org/wiki/Teleo bjectief, https: //medium. com/@solopchuk/intu itions-on-predictive-coding-and-thefree-energy-principle-3 fc 5 bcedc 754 8

Free-energy Principle • The brain avoids surprise by having a good model of the environment. • Free-energy function F = Ω(surprise(sensory input)) Remain alive maintain homeostasis avoid surprising states avoid surprising observations by minimizing approximation to surprise (free energy) *: www. thespruceeats. com *Ref: https: //en. wikipedia. org/wiki/Free_energy_principle 9

Predictive Coding • Based on generative model and free-energy principle. Prediction Hidden states of the world Prediction of the state Differential matrix operator A mixture of prediction errors Free energy Sensory input 10

Formulations of Generative Model Sensory input Level index Hidden state’ state and cause Gaussian random noise. log precisions 11

Make and Improve the Prediction • Combine generative model and predictive coding 12

Overall Flow of the Reasoning 13

Some Results • Signal estimation. 14

Some Results • Figure–ground segregation. 15

Conclusion and To-do List • Reasoning model in human brain • From cognitive science and engineering views. • Brain is a inference machine [Helmholtz, 1860 s] • Active reasoning rather than passive fileting, e. g. , using hypothesis and estimating iteratively (hierarchically) • Predictive coding framework. • TBD • Details for implementation (F, f, g) • Find instances and NN experiments • Performance comparison of NN blocks * Digital detox 16

A new human f. MRI study shows how early visual cortex makes sense of complex visual scenes by segregating foreground and background, and by highlighting outlier objects. The findings are consistent with two attractive theories: biased competition and predictive coding. Bartels, Current Biology, Vol 24, No 13, Cell. Press 17

NN of Predictive Coding *Ref: A Hierarchical Predictive Coding Model of Object Recognition in Natural Images 18

Q&A 19