Principles Underlying the Construction of Brain-Based Devices Jeff Krichmar The Neurosciences Institute San Diego, California, USA test behavior in real world compare with empirical data develop theory create simulation
Construction of an Intelligent Machine Following the Brain Based Model Design should be constrained by these principles: Active sensing and autonomous movement in the environment. Organize the signals from the environment into categories without a priori knowledge or instruction. Incorporate a simulated brain with detailed neural dynamics and neuroanatomy. Engage in a behavioral task and adaptation of behavior when an important environmental event occurs. Allow comparisons with experimental data acquired from animal systems.
Active Sensing and Autonomous Movement in the Environment Darwin IV-VI 1992 - 1998 Darwin VII-VIII 1999 - 2002 Darwin IX-X 2003 - present BrainWorks 2004 - present
Organize the Signals from the Environment into Categories without A Priori Knowledge or Instruction Seth et al, Cerebral Cortex, November 2004, V 14 N 11 Fabre-Thorpe, Phil. Trans. R. Soc. Lond. B (2003) 358, 1215–1223
Incorporate A Simulated Brain With Detailed Neural Dynamics And Neuroanatomy The anatomy of the hippocampus is unique to the nervous system and may play a role in its function, which appears to be important for the acquisition and recall of episodic memories. The picture on the left shows that the hippocampus receives input from many cortical different regions of the brain and that this information is looped through the hippocampus at many levels before returning to the cortex. The picture on the right shows the micro anatomy within the hippocampus itself. There are subregions with specific connectivity such as the dentate gyrus (DG), CA3 and CA1 that receive input from the entorhinal cortex (EC).
Incorporate A Simulated Brain With Detailed Neural Dynamics And Neuroanatomy V1 Color Camera Width V2/4 ODOMETRY HD IT Pr ATN MHDG HIPPOCAMPUS S R+ IR Platform Wall R- MOTOR BF ECinFB DGFB CA3FB CA1FB ECin DG CA3 CA1 Cortex CA3FF CA1FF ECout The hippcampus model for the Darwin X brain-based device was based on the anatomy shown in the previous slide. The diagram to the left shows the high level anatomy of the hippocampus model. The model receives visual input from cortical areas that correspond to object recognition (IT), and spatial position of visual features (Pr). The model also receives input from odometry readings that correspond to Darwin X’s heading (ATN). The model receives positive value if IR sensors detect a hidden platform on the floor of Darwin X’s enclosure and negative value if IR sensors detect a wall. The diagram to the right shows the micro anatomy of the hippocampus model. The model contains approximately 90,000 neuronal units and 1.4 million synaptic connections between those units. ECoutFB S MHDG voltage independent inhibitory plastic value dependent voltage dependent
Engage in a Behavioral Task And Adapt Behavior When An Important Environmental Event Occurs This is our variation of the Morris water maze, without the water, used to test Darwin X’s spatial and episodic memory. Darwin X cannot see the platform because it is the same color as the flooring in the room. However, the platform is reflective and is detected by Darwin X’s IR sensor. Landmarks of colored paper are placed on the wall. Darwin X is started from one of four locations and explores the room until it finds the platform. After it finds the platform, a new trial is started.
Engage in a Behavioral Task And Adapt Behavior When An Important Environmental Event Occurs This movie clip shows Darwin X early on during training in the task where it is clearly moving about the room randomly. By the eighth trial Darwin X, is moving purposefully toward the hidden platform from any starting location in the room. This movie is sped up eight times the real speed.
Allow Comparisons with Experimental Data Acquired from Animal Systems
Allow Comparisons with Experimental Data Acquired from Animal Systems CA1 ECout CA3 DG ECin
Incorporate A Simulated Brain With Detailed Neural Dynamics And Neuroanatomy predictive input reflex response = error signal reflex “Preflex”
Pre-Cerebellar Nuclei Incorporate A Simulated Brain With Detailed Neural Dynamics And Neuroanatomy Camera excitatory Motion Area (MT) inhibitory climbing fiber error signal Pre-Cerebellar Nuclei LTD LTD Purkinje Cells Turn Purkinje Cells Velocity Inferior Olive Turn Inferior Olive Velocity Deep Cerebellar Nuclei Turn Deep Cerebellar Nuclei Velocity Error signal Error signal LTP “Preflex” “Preflex” Reflex Reflex IR Turn Motor Turn Motor Velocity IR Velocity
Engage in a Behavioral Task and Adapt Behavior when an Important Environmental Event Occurs Un-Trained Trained
Allow Comparisons with Experimental Data Acquired from Animal Systems LTD Pre-Cerebellar Nuclei LTD Purkinje Cells Velocity Weight Matrices (initially, all weights were equal) Pre-Cerebellar-NucleiPurkinje Cells for velocity White = maximum Black = minimum More widespread LTD for sharper courses results in lower velocity
Development of Intelligent Machines that follow Neurobiological and Cognitive Principles in their Construction
Build A Brain Team Jason Fleischer Jeff McKinstry Don Hutson Botond Szatmary Anil Seth Jim Snook Krichmar Brian Cox Alisha Lawson Thomas Allen Donatello Darwin V Darwin X BrainWorks Segway B
Construction of an Intelligent Machine Following the Brain Based Model Design should be constrained by these principles: Active sensing and autonomous movement in the environment. Organizing the signals from the environment into categories without a priori knowledge or instruction. Incorporating a simulated brain with detailed neural dynamics and neuroanatomy. Engaging in a behavioral task and adaptation of behavior when an important environmental event occurs. Allowing comparisons with experimental data acquired from animal systems.