Created with ❤️ by Machine Learning & Simulation.
💡 rightclick and open in a new tab to get the flow charts in full size
Primal  Reverse1  Reverse2  Reverse3  Reverse4 

sup1nonetrueprimal onestep supervised training

sup1nonetruefull_gradient full gradient


sup2nonetrueprimal twostep supervised training with aggregation over time

sup2nonetruefull_gradient full gradient

sup2nonetrueno_net_bptt no backpropagation through time over the network


sup2nonefalseprimal twostep supervised training with loss only at final state

sup2nonefalsefull_gradient full gradient

sup2nonefalseno_net_bptt no backpropagation through time over the network


sup3nonetrueprimal threestep supervised training with aggregation over time

sup3nonetruefull_gradient full gradient

sup3nonetrueno_net_bptt no backpropagation through time over the network


sup3nonefalseprimal threestep supervised training with loss only at final state

sup3nonefalsefull_gradient full gradient

sup3nonefalseno_net_bptt no backpropagation through time over the network


sup4nonetrueprimal fourstep supervised training with aggregation over time

sup4nonetruefull_gradient full gradient

sup4nonetrueno_net_bptt no backpropagation through time over the network

sup4nonetruecut_every_2_net_bptt stop backpropagation through time over the network every two network calls


mix11trueprimal mixedchain with one network followed by one physics step with aggregation over time

mix11truefull_gradient full gradient


mix11falseprimal mixedchain with one network followed by one physics step with loss only at final state

mix11falsefull_gradient full gradient


mix21trueprimal mixedchain with two network followed by one physics step with aggregation over time

mix21truefull_gradient full gradient

mix21trueno_net_bptt no backpropagation through time over the network


mix21falseprimal mixedchain with two network followed by one physics step with loss only at final state

mix21falsefull_gradient full gradient

mix21falseno_net_bptt no backpropagation through time over the network


div21trueprimal divertedchain with two steps in main chain and one branch length

div21truefull_gradient full gradient

div21trueno_dp no differentiable physics

div21trueno_net_bptt no backpropagation through time over the network

div21trueno_dpno_net_bptt no differentiable physics and no backpropagation through time over the network

div31trueprimal diverted chain with three steps in main chain and one branch step length

div31truefull_gradient full gradient

todo  todo  todo 
div41trueprimal diverted chain with four steps in main chain and one branch step length

div41truefull_gradient full gradient

todo  todo  todo 
div32trueprimal diverted chain with three steps in main chain and branch chain of length two

div32truefull_gradient full gradient

todo  todo  todo 
div32falseprimal diverted chain with three steps in main chain and branch chain of length two with loss only at final state of the branch

div32falsefull_gradient full gradient

todo  todo  todo 
div42trueprimal diverted chain with four steps in main chain and branch chain of length two

div42truefull_gradient full gradient

todo  todo  todo 
div42falseprimal diverted chain with four steps in main chain and branch chain of length two with loss only at final state of the branch

div42falsefull_gradient full gradient

todo  todo  todo 
tf31trueprimal threestep supervised rollout with state reset after each prediction

tf31truefull_gradient full gradient


tf42trueprimal fourstep supervised rollout with state reset after every second prediction

tf42truefull_gradient full gradient

tf42trueno_net_bptt no backpropagation through time over the network


tf42falseprimal fourstep supervised rollout with state reset after every second prediction; loss only at end of each forcing period

tf42falsefull_gradient full gradient

tf42falseno_net_bptt no backpropagation through time over the network


res1nonefalseprimal onestep residuum training

res1nonefalsefull_gradient full gradient


res2nonetrueprimal twostep residuum training with aggregation over time

res2nonetruefull_gradient full gradient

res2nonetrueno_net_bptt no backpropagation through time over the network

Todo  Todo 
res3nonetrueprimal threestep residuum training with aggregation over time

res3nonetruefull_gradient full gradient

res3nonetrueno_net_bptt no backpropagation through time over the network

Todo  Todo 