To implement the confusion scheme on this dataset a few modification has been
made to the original method, first as the data is high dimensional I've used a
deeper network with convolutional layers as opposed to the shallow fully
connected network suggested by the paper.
Secondly the Haldane model in the phase φ=-π/2 has two critical
frequencies, theoretically the Haldane model is topologically trivial in ω
<6.1 and ω>
6.8 and topological, C=-1, in between.
So we have a double phase transition on our hand and the universal
This is a try to apply the DCN method to detect topological phase transition of the Haldane model from snapshots of the ultra-cold atoms experiment. The dataset is the same as the one used in the paper Unsupervised machine learning of topological phase transitions from experimental data and you can access it from here .
I suspect the problem is either
This is a try to generate new samples from this dataset by adding two more question neurons to the original model that has been used in order to remove the micromotions phase. The idea is that if one can change the micromotion of sampled images with experiment why can't you just generate new images in the desired parameters (with specific frequencies and phases in our case). The problem is that I couldn't find a clear way to evaluate the generated images, and it raises the question wether the generated snapshots preserve meaningful physical properties. If you have any suggestion regarding this I would be happy to discuss this.
I defined the environment as a set of nodes living in the parameter space, presenting the phase diagram of the many-body physical system under study. The agent can take action to change the coordinates of these nodes to obtain the physical system's phase diagram. The rewarding system for this environment is defined as the performance of a shallow neural network trained with the suggested phase diagram by the agent as an indicator of how well it is doing (confusion scheme). Here I used the QLearning method which requires a discrete action space and by using SAC we can use a continous action space.
Coming soon...