The U-Net CNN with differentiable plasticity learning method implementation
The learner architecture
The main learner part is based on U-Net architecture which demonstrate good prediction power in medical images segmentation tasks. Taking into account that both tasks - medical and seismical images segmentation - are very close we decided to use this network architecture for this task. But, taking into account, that for seismical images we need to identify specific parts of the images corresponding to the salt accumulations, which rather is not clear segmentation task, we decided to augment the architecture with synaptic plasticity rules. The introduction of plasticity rules allows to implement long-term memory into our architecture allowing it to maximize influence of previously seen training samples on the current optimization step.
The synaptic plasticity is a major mechanism for continuous learning during life-time implemented in the human brain, which makes it so efficient in assimilation of novel data based on previous experience. The plasticity of the synaptic weights implemented by adjust of weights during inference depending on training signals received from the environment.
In the ANN the plasticity rules can be implemented in different ways. In this work we will consider plastic rules implemented separately from inference part of the ANN architecture. We started with simple Hebbian rule, which stores plastic coefficients in Hebbian trace during lifetime and applies learned plastic part at the final stage of inference routine, i.e. before final layer of our network architecture. Then we continued with more advanced Oja rule which provides workaround over Hebbian tendency to decay plastic coefficients to zero during life-time. The Oja rule can maintain stable weight values in plastic part indefinitely in the absence of stimulation, thus allowing stable long-term memories, while still preventing runaway divergences.
The OJA rule gives considerable predictive performance boost over HEBB rule due to it's ability to maintain learned weights indefinitely, thus allowing stable long-memories. Due to this effect knowledge inferred from previously seen training samples greatly determines the update in the current optimization step.