Cortex learning models (Weber at al. 2006, Weber and Triesch, 2006, Weber and Wermter 2006/7)

 Download zip file 
Help downloading and running models
Accession:83528
A simulator and the configuration files for three publications are provided. First, "A hybrid generative and predictive model of the motor cortex" (Weber at al. 2006) which uses reinforcement learning to set up a toy action scheme, then uses unsupervised learning to "copy" the learnt action, and an attractor network to predict the hidden code of the unsupervised network. Second, "A Self-Organizing Map of Sigma-Pi Units" (Weber and Wermter 2006/7) learns frame of reference transformations on population codes in an unsupervised manner. Third, "A possible representation of reward in the learning of saccades" (Weber and Triesch, 2006) implements saccade learning with two possible learning schemes for horizontal and vertical saccades, respectively.
References:
1 . Weber C, Wermter S, Elshaw M (2006) A hybrid generative and predictive model of the motor cortex. Neural Netw 19:339-53 [PubMed]
2 . Weber C, Triesch J (2006) A possible representation of reward in the learning of saccades Proc. of the Sixth International Workshop on Epigenetic Robots :153-60
3 . Weber C, Wermter S (2006) A self-organizing map of sigma-pi units Neurocomputing 70(13-15):2552-2560
Model Information (Click on a link to find other models with that property)
Model Type: Connectionist Network;
Brain Region(s)/Organism:
Cell Type(s):
Channel(s):
Gap Junctions:
Receptor(s):
Gene(s):
Transmitter(s):
Simulation Environment: C or C++ program;
Model Concept(s): Rate-coding model neurons; Reinforcement Learning; Unsupervised Learning; Attractor Neural Network; Winner-take-all; Hebbian plasticity; Olfaction;
Implementer(s): Weber, Cornelius [cweber at fias.uni-frankfurt.de]; Elshaw, Mark [mark.elshaw at sunderland.ac.uk];
/**
This contains some additional code to analyze the network. Produces Fig.8 of
"A hybrid generative and predictive model of the motor cortex"
 Cornelius Weber, Stefan Wermter and Mark Elshaw
 Neural Networks, 19 (4), 339-353 (2006)
Learning and weight-export is out-commented here.
Weights have to be trained earlier by the file CortexDocking(.enr).
To display the flow field of the reinforcement-trained network:
- set iterWreinforce   1000000       (or any large number)
- After running cococo with this file, then run the flowfield.tcl file:
  wish -f flowfield.tcl
  "wish" is the widget shell, the TCL/TK interpreter.
  This produces also the eps file  flow.eps  for visualization.
To display the flowfield of the Helmholtz machine network (the bu and td weights):
- set iterWreinforce   0             (so not to go in there)
- set iterWbu          200000        (or any large number)
- Then run the flowfield.tcl file:
  wish -f flowfield.tcl
To display the flowfield of the predictive model (lateral weights V):
- set iterWreinforce   0             (so not to go in there)
- set iterWbu          0             (so not to go in there)
- set iterV            100000        (or any large number)
- Then run the flowfield.tcl file:
  wish -f flowfield.tcl
To display the flowfields with a different robot angle,
change the last argument of the function total_scan_coord (which is here set to 0.0).
**/
global {
 iter         0
 areas        10
 mult         1
 d           "/tmp/coco"
}
all {
 d_a          0
 d_b          0
}
/**area 4: place cells**/
area 4 {
 d_a          112        /**16 * num_angle**/
 d_b          24
}
/**area 5: critic -- holds value function**/
area 5 {
 d_a          1
 d_b          1
}
/**area 7: real world coordinates -- x,y,phi**/
area 7 {
 d_a          3
 d_b          1
}
/**area 6: actor -- steers the agent into N E S W**/
area 6 {
 d_a          1
 d_b          4
}
/**area 3: the perceived Gaussian in NON-blown-up coordinates**/
area 3 {
 d_a          16
 d_b          24
}
/**area 8: phi as Gaussian -- HD cells -- angle input for the cortex**/
area 8 {
 d_a          7
 d_b          1
}
/**area 9: motor cortex -- its input is (i) the perceived location from area 3, (ii) the angle from area 9 and (iii) the motor output from area 6**/
area 9 {
 d_a          16
 d_b          16
}
/**Actor-Critic**/
/**Helmholtz Machine**/
/**Lateral Predictive Weights**/
series (1) {
 if (1 = 0)
 sw (0; 1; alltime) {
  5(n) {N,w; weight_list_alloc_full; 4    ;   ; -0.01+0.01} /**critic weights**/
  6(n) {N,w; weight_list_alloc_full; 4    ;   ; -0.01+0.01} /**actor weights**/
 }
 if (1 = 1)
 sw (0; 1; alltime) {
  5(n) {N,w+d; weight_list_alloc_import; 4;   ;    }
  6(n) {N,w+d; weight_list_alloc_import; 4;   ;    }
 }
 sw (0; 1; order) {
  5(o){R; local_const       ; ,   ; ,   ; 1.0 } /**so that total_init_coord chooses a new position**/
  9(o){O; local_const       ; ,   ; ,   ; 0.0 }              /**    init   **/
  3(o){O; local_const       ; ,   ; ,   ; 0.0 }              /**    the    **/
  6(o){O; local_const       ; ,   ; ,   ; 0.0 }              /**    old    **/
  8(o){O; local_const       ; ,   ; ,   ; 0.0 }              /**activations**/
 }
}
/**training of the reinforcement module to perform the docking**/
series (1000000) {
 sw (0; 1; order) {
  7(o){P  ; local_copy      ; ,   ; Q,  ;      } /****/
  7(t){P  ; total_init_coord; 5   ; R   ;16+24, 50, 0.0} /**init random new (x,y,phi), if R != 0, else leave as is <- make sure orange visible; set phi = 0 for simplicity!**/
  7(t){P  ; total_scan_coord;     ;     ; 1, 16+24, 16+24, 0.0}       /**angles tested: 0.0, 0.4363323, -0.4363323**/
  4(t){P  ; total_angl_coord; 7   ;     ;1.5+0.5,7+0.78539816} /**compute the perceived orange**/
  5(o){K,w; weight_list_feed; 4   ; P   ;      } /**Eq.2:  critic activation at time t**/
        
  6(o){N,w; weight_list_feed; 4   ; P   ;      } /**Eq. before 9: actor inner activation**/
  6(t){M  ; total_rand_winner;    ; N   ;      } /**Eq.9:  choose action stochastically**/
  6(t){M  ; total_winner    ;     ; N   ; 1, 1.0}
  6(t){M  ; total_scan_coord;     ;     ; 2    }
  6(t){M  ; total_winner    ;     ; N   ; 1, 1.0}
  7(t){Q  ; total_move_coord; 6+7 ; M+P ; 0.9+0.1} /**old position P and action M (q_00~velocity,q_01~angularvelocity) -> new position Q**/
  4(t){Q  ; total_angl_coord; 7   ;     ;1.5+0.5,7+0.78539816} /**compute the perceived orange**/
  5(o){R  ; feed_dock_reward; 4,7 ; Q,Q ;15+11,7+16+24,2+0,-0.3,8.0} /**set reward R=1 at target pos q_00+q_01, and R = q_30 at boundary; 0.0 for nicer presentation**/
  5(o){L,w; weight_list_feed; 4   ; Q   ;      } /**Eq.2:  critic activation at time t+1             **/
  5(o){L  ; local_mult_const; ,   ; ,   ;0.9} /**Eq.7:  discount factor gamma                     \    compute **/
  5(o){T  ; local_sub       ; ,   ; L, K;      } /**Eq.7:  difference critic act at t+1 minus at t    |   delta   **/
  5(o){T  ; local_sum       ; ,   ; R,  ;      } /**Eq.7:  add reward                                /            **/
  6(t){T  ; total_spread_xy ; 5   ;     ; 2    } /**                      delta on all action units  **/
  6(o){O  ; local_mult      ; ,   ; T, M;      } /**Eq.10: delta * g_j    delta on active action unit**/
/*5(o){N,w; weight_list_hebb; 5, 4; T, P; 0.1  }*/ /**Eq.8:  delta * f_i(p_t)                          **/
/*6(o){N,w; weight_list_hebb; 6, 4; O, P; 0.1  }*/ /**Eq.10: (delta*g_j) * f_i(p_t)                    **/
  4(t){R; total_pause     ;     ;     ;  0  } /**in ms**/
 }
 if (iter % 100)
 sw (0; 1; alltime) {
  4(o){P,d; observe_act     ;     ;     ;      }
  4(o){Q,d; observe_act     ;     ;     ;      }
  5(o){R,d; observe_act     ;     ;     ;      }
  5(o){L,d; observe_act     ;     ;     ;      }
  6(o){N,d; observe_act     ;     ;     ;      }
  6(o){M,d; observe_act     ;     ;     ;      }
  7(o){P,d; observe_act     ;     ;     ;      }
  7(o){Q,d; observe_act     ;     ;     ;      }
/*6(n){N,w+d; weight_list_export; 4;    ;      }*/
/*5(n){N,w+d; weight_list_export; 4;    ;      }*/
 }
}
series (1) {
 if (1 = 0)
 sw (0; 1; alltime) {
  3(n) {N,w; weight_list_alloc_full; 9    ;   ; -0.01+0.01}   /****/
  6(n) {N,w; weight_list_alloc_full; 9    ;   ; -0.01+0.01}   /****/
  8(n) {N,w; weight_list_alloc_full; 9    ;   ; -0.01+0.01}   /****/
  9(n) {N,w; weight_list_alloc_full; 3    ;   ; -0.01+0.01}   /****/
  9(n) {N,w; weight_list_alloc_full; 6    ;   ; -0.01+0.01}   /****/
  9(n) {N,w; weight_list_alloc_full; 8    ;   ; -0.01+0.01}   /****/
 }
 if (1 = 1)
 sw (0; 1; alltime) {
  3(n) {N,w+d; weight_list_alloc_import; 9;   ;    }
  6(n) {N,w+d; weight_list_alloc_import; 9;   ;    }
  8(n) {N,w+d; weight_list_alloc_import; 9;   ;    }
  9(n) {N,w+d; weight_list_alloc_import; 3;   ;    }
  9(n) {N,w+d; weight_list_alloc_import; 6;   ;    }
  9(n) {N,w+d; weight_list_alloc_import; 8;   ;    }
 }
}
/**training of the motor cortex (only HM, i.e. BU/TD, not lateral yet)**/
series (200000) {
 sw (0; 1; order) {
  3(o){N,w; weight_list_rectify;  9  ;      ; 0.0, 1}  /** all weights W **/
  6(o){N,w; weight_list_rectify;  9  ;      ; 0.0, 1}  /**     are       **/
  8(o){N,w; weight_list_rectify;  9  ;      ; 0.0, 1}  /**   positive    **/
  9(o){N,w; weight_list_rectify;3+6+8; ++   ; 0.0, 1}  /**     only      **/
 }
 sw (0; 1; order) {
  7(o){P  ; local_copy      ; ,   ; Q,  ;      } /****/
  7(t){P  ; total_init_coord; 5   ; R   ;16+24, 50, 0.0} /**init random new (x,y,phi), if R != 0, else leave as is <- make sure orange visible; set phi = 0 for simplicity!**/
  /*7(t){P  ; total_scan_coord;     ;     ; 1, 16+24, 16+24, 0.0}*/       /**angles tested: 0.0, 0.4363323, -0.4363323**/
  4(t){P  ; total_angl_coord; 7   ;     ;1.5+0.5,7+0.78539816} /**compute the perceived orange**/
  5(o){K,w; weight_list_feed; 4   ; P   ;      } /**Eq.2:  critic activation at time t**/
        
  6(o){N,w; weight_list_feed; 4   ; P   ;      } /**Eq. before 9: actor inner activation**/
  6(t){M  ; total_rand_winner;    ; N   ;      } /**Eq.9:  choose action stochastically**/
  /*6(t){M  ; total_winner    ;     ; N   ; 1, 1.0}*/
  /*6(t){M  ; total_scan_coord;     ;     ; 2    } */
  6(t){M  ; total_winner    ;     ; N   ; 1, 1.0}
  7(t){Q  ; total_move_coord; 6+7 ; M+P ; 0.9+0.1} /**old position P and action M (q_00~velocity,q_01~angularvelocity) -> new position Q**/
  4(t){Q  ; total_angl_coord; 7   ;     ;1.5+0.5,7+0.78539816} /**compute the perceived orange**/
  5(o){R  ; feed_dock_reward; 4,7 ; Q,Q ;15+11,7+16+24,2+0,-0.3,8.0} /**set reward R=1 at target pos q_00+q_01, and R = q_30 at boundary; 0.0 for nicer presentation**/
 }
 /**WAKE: TD weights W_(368)9 learn**/
 sw (0; 1; order) {
  3(t){T; total_imag_coord  ; 7   ; P   ;1, 1.5}                     /**compute the perceived orange (on the original "where" area, NOT on the angle-blown-up input)**/
  3(o){T; local_mult_const  ; ,   ; ,   ; 1.0 }                     /**!!!**/
  6(o){T; local_copy        ; ,   ; M,  ;         }                     /**the action**/
  6(o){T; local_mult_const  ; ,   ; ,   ; 1.0 }                     /**!!!**/
  8(t){T; total_imag_coord  ; 7   ; P   ;2, 0.5, 7+0.78539816}   /**the heading direction cells**/
  8(o){T; local_mult_const  ; ,   ; ,   ; 2.0 }                     /**!!!**/
 }
 sw (0; 1; order) {
  9(o){T,w; weight_list_feed;3+6+8;T+T+T;     }
  9(o){Q,w; weight_list_feed; 3   ; T   ;     }                         /** only  **/
  9(o){R,w; weight_list_feed; 6   ; T   ;     }                         /**  for  **/
  9(o){S,w; weight_list_feed; 8   ; T   ;     }                         /**observe**/
  9(o){T; local_mean_01     ; ,   ; ,   ; 64+2.0+0+0+1}
  3(o){S,w; weight_list_feed; 9   ; T   ;     }
  3(o){R; local_sub         ; ,   ; T, S;     }
/*3(o){N,w; weight_list_hebb; 3, 9; R, T; 0.01}*/
  6(o){S,w; weight_list_feed; 9   ; T   ;     }
  6(o){R; local_sub         ; ,   ; T, S;     }      /**to plot the TD errors R (continuous) with init on area 6**/
/*6(o){N,w; weight_list_hebb; 6, 9; R, T; 0.01}*/
  8(o){S,w; weight_list_feed; 9   ; T   ;     }
  8(o){R; local_sub         ; ,   ; T, S;     }
/*8(o){N,w; weight_list_hebb; 8, 9; R, T; 0.01}*/
/*3(o){N,w; weight_list_decay ; 9   ;     ; 0.00015 }*/
/*6(o){N,w; weight_list_decay ; 9   ;     ; 0.00015 }*/
/*8(o){N,w; weight_list_decay ; 9   ;     ; 0.00015 }*/
 }
 /**to plot the TD errors S (continuous) if NO init on area 6, but during testing -- comment this out to plot the training errors, i.e. with correct area 6 init**/
 sw (0; 1; order) {
  9(o){L,w; weight_list_feed    ;3+8;T+T;     }
  9(o){L; local_mean_01     ; ,   ; ,   ; 64+2.0+0+0+1}
  6(o){S,w; weight_list_feed    ; 9   ; L   ;     }
  6(o){S; local_sub         ; ,   ; T, S;     }
 }
/**SLEEP: BU weights W_9(368) learn**/
 sw (0; 1; order) {
  9(t){N; total_topo_gibbs_01;,   ; ,   ; 10+0.75, 1.0+1.0, 1}    /**init sleep phantasy**/
  3(o){N,w; weight_list_feed; 9   ;     ;     }
  6(o){N,w; weight_list_feed; 9   ;     ;     }
  8(o){N,w; weight_list_feed; 9   ;     ;     }
  9(o){M,w; weight_list_feed;3+6+8;N+N+N;     }
  9(o){M; local_mean_01     ; ,   ; ,   ; 64+2.0+0+0+1}
  9(o){N; local_sub         ; ,   ; N, M;     }
/*9(o){N,w; weight_list_hebb;9,3+6+8;N,N+N+N; 0.01}*/
/*9(o){N,w; weight_list_decay ;3+6+8; ,   ; 0.00015}*/
 }
 /**test whether the feed-back weights W_69 can already reconstruct the optimal motor output**/
 sw (0; 1; order) {
  7(t){P; total_scan_coord  ;     ;     ; 1, 16+24, 16+24, 0.43633231}/**/ /**init x,y on grid, given phi; write to file**/ /**comment this line OUT to obtain the errors!**/
  3(t){P; total_imag_coord  ; 7   ;     ; 1, 1.5}                    /**compute the perceived orange**/
  3(o){P; local_mult_const  ; ,   ; ,   ; 1.0  }                     /**!!!**/
  8(t){P; total_imag_coord  ; 7   ; P   ; 2, 0.5, 7+0.78539816}  /**the heading direction cells**/
  8(o){P; local_mult_const  ; ,   ; ,   ; 2.0  }                     /**!!!**/
  9(o){Q,w; weight_list_feed    ; 3   ; P   ;     }                         /** only  **/
  9(o){S,w; weight_list_feed    ; 8   ; P   ;     }                         /**observe**/
 }
 sw (0; 1; order) {
  9(o){Q; local_mult_const  ; ,   ; ,   ; 1.0}
  9(o){O; local_sum         ; ,   ; Q, S;     }
  9(o){O; local_mean_01     ; ,   ; ,   ; 64+2.0+0+0+1}
  6(o){O,w; weight_list_feed    ; 9   ;     ;     }
  6(t){O; total_winner      ;     ;     ; 1, 1.0}                       /**only theoretically ... (for display reasons ...)**/
  6(t){O; total_scan_coord  ;     ;     ; 2   }                         /**write the 4 motor values to same file as above**/
/*6(o){S,w; weight_list_feed    ; 9   ; O   ;     }*/                                                                                /**to obtain the errors**/
 }
 if (iter % 100)
 sw (0; 1; alltime) {
  3(o){T,d; observe_act     ;     ;     ;      }
  6(o){T,d; observe_act     ;     ;     ;      }
  8(o){T,d; observe_act     ;     ;     ;      }
  9(o){T,d; observe_act     ;     ;     ;      }
  3(o){N,d; observe_act     ;     ;     ;      }
  6(o){N,d; observe_act     ;     ;     ;      }
  8(o){N,d; observe_act     ;     ;     ;      }
  9(o){N,d; observe_act     ;     ;     ;      }
  9(o){M,d; observe_act     ;     ;     ;      }
  9(o){Q,d; observe_act     ;     ;     ;      }
  9(o){R,d; observe_act     ;     ;     ;      }
  9(o){S,d; observe_act     ;     ;     ;      }
 }
/*
 if (iter % 500)
 sw (0; 1; alltime) {
  9(n){N,w+d; weight_list_export; 3;    ;      }
  9(n){N,w+d; weight_list_export; 6;    ;      }
  9(n){N,w+d; weight_list_export; 8;    ;      }
  3(n){N,w+d; weight_list_export; 9;    ;      }
  6(n){N,w+d; weight_list_export; 9;    ;      }
  8(n){N,w+d; weight_list_export; 9;    ;      }
 }
*/
}
 /**lateral attractor weights V (separate relaxation with WAKE data)**/
 /**the relaxation dynamics might solve the problem of conflicting BU inputs from motor, angle and location ... !**/
 /**might relax together with motor output area 6, because the classification is relevant w.r.t. the motor output anyway ... !**/
series (1) {
 if (1 = 0)
 sw (0; 1; alltime) {
  3(n) {N,v; weight_list_alloc_full; 9    ;   ; -0.01+0.01}   /****/
  6(n) {N,v; weight_list_alloc_full; 9    ;   ; -0.01+0.01}   /****/
  8(n) {N,v; weight_list_alloc_full; 9    ;   ; -0.01+0.01}   /****/
  9(n) {N,v; weight_list_alloc_full; 3    ;   ; -0.01+0.01}   /****/
  9(n) {N,v; weight_list_alloc_full; 6    ;   ; -0.01+0.01}   /****/
  9(n) {N,v; weight_list_alloc_full; 8    ;   ; -0.01+0.01}   /****/
  3(n) {N,v; weight_list_alloc_full; 3    ;   ; -0.01+0.01}   /****/
  6(n) {N,v; weight_list_alloc_full; 6    ;   ; -0.01+0.01}   /****/
  8(n) {N,v; weight_list_alloc_full; 8    ;   ; -0.01+0.01}   /****/
  9(n) {N,v; weight_list_alloc_full; 9    ;   ; -0.01+0.01}   /****/
 }
 if (1 = 0)
 sw (0; 1; order) {
  3(o){N,v; weight_list_cutself;     ;      ;     }
  6(o){N,v; weight_list_cutself;     ;      ;     }
  8(o){N,v; weight_list_cutself;     ;      ;     }
  9(o){N,v; weight_list_cutself;     ;      ;     }
 }
 if (1 = 1)
 sw (0; 1; alltime) {
  3(n) {N,v+d; weight_list_alloc_import; 9;   ;    }
  6(n) {N,v+d; weight_list_alloc_import; 9;   ;    }
  8(n) {N,v+d; weight_list_alloc_import; 9;   ;    }
  9(n) {N,v+d; weight_list_alloc_import; 3;   ;    }
  9(n) {N,v+d; weight_list_alloc_import; 6;   ;    }
  9(n) {N,v+d; weight_list_alloc_import; 8;   ;    }
  3(n) {N,v+d; weight_list_alloc_import; 3;   ;    }
  6(n) {N,v+d; weight_list_alloc_import; 6;   ;    }
  8(n) {N,v+d; weight_list_alloc_import; 8;   ;    }
  9(n) {N,v+d; weight_list_alloc_import; 9;   ;    }
 }
}
series (100000) {
 /**begin of copy&paste from above**/
 sw (0; 1; order) {
  7(o){P  ; local_copy      ; ,   ; Q,  ;      } /****/
  7(t){P  ; total_init_coord; 5   ; R   ;16+24, 50, 0.0} /**init random new (x,y,phi), if R != 0, else leave as is <- make sure orange visible; set phi = 0 for simplicity!**/
  /*7(t){P  ; total_scan_coord;     ;     ; 1, 16+24, 16+24, 0.0}*/       /**angles tested: 0.0, 0.4363323, -0.4363323**/
  4(t){P  ; total_angl_coord; 7   ;     ;1.5+0.5,7+0.78539816} /**compute the perceived orange**/
  5(o){K,w; weight_list_feed; 4   ; P   ;      } /**Eq.2:  critic activation at time t**/
        
  6(o){N,w; weight_list_feed; 4   ; P   ;      } /**Eq. before 9: actor inner activation**/
  6(t){M  ; total_rand_winner;    ; N   ;      } /**Eq.9:  choose action stochastically**/
  /*6(t){M  ; total_winner    ;     ; N   ; 1, 1.0}*/
  /*6(t){M  ; total_scan_coord;     ;     ; 2    } */
  6(t){M  ; total_winner    ;     ; N   ; 1, 1.0}
  7(t){Q  ; total_move_coord; 6+7 ; M+P ; 0.9+0.1} /**old position P and action M (q_00~velocity,q_01~angularvelocity) -> new position Q**/
  4(t){Q  ; total_angl_coord; 7   ;     ;1.5+0.5,7+0.78539816} /**compute the perceived orange**/
  5(o){R  ; feed_dock_reward; 4,7 ; Q,Q ;15+11,7+16+24,2+0,-0.3,8.0} /**set reward R=1 at target pos q_00+q_01, and R = q_30 at boundary; 0.0 for nicer presentation**/
 }
 /**WAKE: TD weights W_(368)9 learn**/
 sw (0; 1; order) {
  3(t){T; total_imag_coord  ; 7   ; P   ;1, 1.5}                     /**compute the perceived orange (on the original "where" area, NOT on the angle-blown-up input)**/
  3(o){T; local_mult_const  ; ,   ; ,   ; 1.0 }                     /**!!!**/
  6(o){T; local_copy        ; ,   ; M,  ;         }                     /**the action**/
  6(o){T; local_mult_const  ; ,   ; ,   ; 1.0 }                     /**!!!**/
  8(t){T; total_imag_coord  ; 7   ; P   ;2, 0.5, 7+0.78539816}   /**the heading direction cells**/
  8(o){T; local_mult_const  ; ,   ; ,   ; 2.0 }                     /**!!!**/
 }
 sw (0; 1; order) {
  9(o){T,w; weight_list_feed;3+6+8;T+T+T;     }
  9(o){T; local_mean_01     ; ,   ; ,   ; 64+2.0+0+0+1}
 }
 /**end of copy&paste from above**/
 sw (0; 1; order) {
  3(o){T; local_rectify     ; ,   ; ,   ; 0.8+-1}   /**just for consistency with the 2 lines below**/
  6(o){T; local_rectify     ; ,   ; ,   ; 0.8+-1}   /**because value was previously multiplied by scale6>1**/
  8(o){T; local_rectify     ; ,   ; ,   ; 0.8+-1}   /**because value was previously multiplied by scale8>1**/
 }
 sw (0; 1; order) {
  9(o){P,v; weight_list_feed;3+6+8+9;O+O+O+O; }    /**see 06.gaussequence.enr for average between BU & lateral input -- related to prediction !**/
      {P; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}    /**works more robustly with kurt=8 instead of 64 as for BU cells -- 0.7 trick to keep act's high!**/
  3(o){P,v; weight_list_feed; 9+3 ; O+O ;     }    /**see 06.gaussequence.enr for average between BU & lateral input -- related to prediction !**/
      {P; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}    /**works more robustly with kurt=8 instead of 64 as for BU cells -- 0.7 trick to keep act's high!**/
  6(o){P,v; weight_list_feed; 9+6 ; O+O ;     }    /**see 06.gaussequence.enr for average between BU & lateral input -- related to prediction !**/
      {P; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}    /**works more robustly with kurt=8 instead of 64 as for BU cells -- 0.7 trick to keep act's high!**/
  8(o){P,v; weight_list_feed; 9+8 ; O+O ;     }    /**see 06.gaussequence.enr for average between BU & lateral input -- related to prediction !**/
      {P; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}    /**works more robustly with kurt=8 instead of 64 as for BU cells -- 0.7 trick to keep act's high!**/
 }
 sw (0; 1; order) { /**attention: it may not work to learn only at the last step, because the pre-synaptic value might be too small!!!**/
/*9(o){N,v; weight_list_decay ; 3+6+8+9;     ; 0.00001 }*/
/*3(o){N,v; weight_list_decay ; 3+9    ;     ; 0.00001 }*/
/*6(o){N,v; weight_list_decay ; 6+9    ;     ; 0.00001 }*/
/*8(o){N,v; weight_list_decay ; 8+9    ;     ; 0.00001 }*/
  9(o){K; local_sub         ; ,   ; T, P;     }
  3(o){K; local_sub         ; ,   ; T, P;     }
  6(o){K; local_sub         ; ,   ; T, P;     }      /**to plot the lat errors K (continuous) with init on area 6**/
  8(o){K; local_sub         ; ,   ; T, P;     }
/*9(o){N,v; weight_list_hebb;,3+6+8+9;K,O+O+O+O; 0.01}*/
/*3(o){N,v; weight_list_hebb; ,9+3;K,O+O; 0.01}*/
/*6(o){N,v; weight_list_hebb; ,9+6;K,O+O; 0.01}*/
/*8(o){N,v; weight_list_hebb; ,9+8;K,O+O; 0.01}*/
  9(o){Q,v; weight_list_feed ; 3   ; P   ;     }                         /** only  **/    /**here  at**/
  9(o){R,v; weight_list_feed ; 6   ; P   ;     }                         /**  for  **/    /**separate time**/
  9(o){S,v; weight_list_feed ; 8   ; P   ;     }                         /**observe**/    /**from above**/
 }
 sw (0; 1; order) {
  9(o){O; local_copy        ; ,   ; T,  ;     }                         /**      the       **/
  3(o){O; local_copy        ; ,   ; T,  ;     }                         /**      old       **/
  6(o){O; local_copy        ; ,   ; T,  ;     }                         /**  activations   **/
  8(o){O; local_copy        ; ,   ; T,  ;     }                         /**for earlier init**/
 }
 /**only for observe**/
 sw (1; 10; order) {
  9(o){P,v; weight_list_feed;3+6+8+9;+++; -1 }
      {P; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}
  3(o){P,v; weight_list_feed; 9+3 ; +   ; -1 }
      {P; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}
  6(o){P,v; weight_list_feed; 9+6 ; +   ; -1 }
      {P; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}
  8(o){P,v; weight_list_feed; 9+8 ; +   ; -1 }
      {P; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}
 }
[
 /**to plot the lat errors L (continuous) if NO init on area 6, but during training -- comment this out to plot the training errors, i.e. with correct area 6 init**/
 sw (0; 1; order) {
  9(o){P  ; local_const       ; ,   ; ,   ; 0.0 }
  3(o){P  ; local_copy        ; ,   ; T,  ;     }
  6(o){P  ; local_const       ; ,   ; ,   ; 0.0 }
  8(o){P  ; local_copy        ; ,   ; T,  ;     }
 }
 sw (1; 10; order) {
  3(o){P  ; single_copy       ;     ;     ; -1  }                         /**clamping the other inputs!**/
  8(o){P  ; single_copy       ;     ;     ; -1  }                         /**(better performance than unclamped)**/
 }
 sw (1; 10; propto) {
  9(o){P,v; weight_list_feed  ;3+6+8+9;+++; -1  }    /**see 06.gaussequence.enr for average between BU & lateral input -- related to prediction !**/
      {P  ; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}    /**works more robustly with kurt=8 instead of 64 as for BU cells -- 0.7 trick to keep act's high!**/
  6(o){P,v; weight_list_feed  ; 9+6 ; +   ; -1  }    /**see 06.gaussequence.enr for average between BU & lateral input -- related to prediction !**/
      {P  ; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}    /**works more robustly with kurt=8 instead of 64 as for BU cells -- 0.7 trick to keep act's high!**/
 }
 sw (0; 10; order) {
  6(t){L  ; total_copy_back   ;     ; P   ;     }
 }
 sw (0; 1; order) {
  6(o){L  ; local_sub         ; ,   ; T, L;     }
 }
]
 /**test whether the "lateral" weights V can reconstruct the optimal motor output**/
 /**areas 3 and 8 clamped <-- better than totally free running relaxation!**/
 sw (0; 1; order) {
  7(t){T  ; total_scan_coord  ;     ;     ; 1, 16+24, 16+24, 0.43633231}/**/ /**init x,y on grid given phi; write to file**/ /**comment this line OUT to obtain the errors!**/
/*7(o){T  ; local_copy        ; ,   ; P,  ;          }*/                      /**comment this line IN to obtain the errors!**/
  3(t){T  ; total_imag_coord  ; 7   ;     ; 1, 1.5}                    /**compute the perceived orange**/
  3(o){T  ; local_mult_const  ; ,   ; ,   ; 1.0  }                    /**!!!**/
  8(t){T  ; total_imag_coord  ; 7   ;     ; 2, 0.5, 7+0.78539816}  /**the heading direction cells**/
  8(o){T  ; local_mult_const  ; ,   ; ,   ; 2.0  }                    /**!!!**/
  9(o){T,w; weight_list_feed  ;3+8  ;T+T  ;     }
  9(o){T  ; local_mean_01     ; ,   ; ,   ; 64+2.0+0+0+1}
 }
 sw (1; 10; order) {
  9(o){T  ; single_copy       ;     ; T   ; -1  }    /**target values for whole sequence -- here just for display**/
  3(o){T  ; single_copy       ;     ; T   ; -1  }    /**target values for whole sequence**/
  8(o){T  ; single_copy       ;     ; T   ; -1  }    /**target values for whole sequence**/
 }
 sw (0; 10; order) {
  3(o){T  ; local_rectify     ; ,   ; ,   ; 0.8+-1}  /**just for consistency with the 2 lines below**/
  8(o){T  ; local_rectify     ; ,   ; ,   ; 0.8+-1}  /**because value was previously multiplied by scale8>1**/
 }
 sw (0; 1; order) {
  9(o){P  ; local_copy        ; ,   ; T,  ;     }
  3(o){P  ; local_copy        ; ,   ; T,  ;     }
  6(o){P  ; local_const       ; ,   ; ,   ; 0.0 }
  8(o){P  ; local_copy        ; ,   ; T,  ;     }
 }
 sw (1; 10; order) {
  3(o){P  ; local_copy        ; ,   ; T,  ;     }                         /**clamping the other inputs!**/
  8(o){P  ; local_copy        ; ,   ; T,  ;     }                         /**(better performance than unclamped)**/
 }
 sw (1; 10; propto) {
  9(o){P,v; weight_list_feed;3+6+8+9;+++; -1  }    /**see 06.gaussequence.enr for average between BU & lateral input -- related to prediction !**/
      {P; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}    /**works more robustly with kurt=8 instead of 64 as for BU cells -- 0.7 trick to keep act's high!**/
  6(o){P,v; weight_list_feed; 9+6 ; +   ; -1  }    /**see 06.gaussequence.enr for average between BU & lateral input -- related to prediction !**/
      {P; local_mean_01     ; ,   ; ,   ; 8+2.0+0+0+0.7}    /**works more robustly with kurt=8 instead of 64 as for BU cells -- 0.7 trick to keep act's high!**/
 }
 sw (9; 10; order) {
  6(t){O; total_winner      ;     ; P   ; 1 ,1.0}          /**only theoretically ... (for display reasons ...)**/
  6(t){O; total_scan_coord  ;     ;     ; 2   }                         /**write the 4 motor values to same file as above**/
 }
 if (iter % 100)
 sw (0; 10; alltime) {
  9(o){O,d; observe_act     ;     ;     ;      }
  6(o){O,d; observe_act     ;     ;     ;      }
  3(o){P,d; observe_act     ;     ;     ;      }
  6(o){P,d; observe_act     ;     ;     ;      }
  8(o){P,d; observe_act     ;     ;     ;      }
  9(o){P,d; observe_act     ;     ;     ;      }
 }
 if (iter % 100)
 sw (0; 1; alltime) {
  3(o){T,d; observe_act     ;     ;     ;      }
  6(o){T,d; observe_act     ;     ;     ;      }
  8(o){T,d; observe_act     ;     ;     ;      }
  9(o){T,d; observe_act     ;     ;     ;      }
 }
/*
 if (iter % 500)
 sw (0; 1; alltime) {
  9(n){N,v+d; weight_list_export; 3;    ;      }
  9(n){N,v+d; weight_list_export; 6;    ;      }
  9(n){N,v+d; weight_list_export; 8;    ;      }
  3(n){N,v+d; weight_list_export; 9;    ;      }
  6(n){N,v+d; weight_list_export; 9;    ;      }
  8(n){N,v+d; weight_list_export; 9;    ;      }
  3(n){N,v+d; weight_list_export; 3;    ;      }
  6(n){N,v+d; weight_list_export; 6;    ;      }
  8(n){N,v+d; weight_list_export; 8;    ;      }
  9(n){N,v+d; weight_list_export; 9;    ;      }
 }
*/
}

Loading data, please wait...