• Implement adverserial Lavg -> Lmax
• Robust optimization experiment
• Active learning experiment
• ICML Feb 24!!!!
• Figure out the theory part of domain adaptation stuff
• Domain adaptation as a discriminator
• Domain adaptation experiment
• Active domain adaptation experiment
• ICCV March 17
• RL experiment with robots
• Counter-Factual learning
• NIPS May 19

• Cifar10 training and test (full data/nothing new using tf_base)
• Set the seed training_set and save it num_images = 5000
• Random active learning test 0.1/0.2/0.3/0.4/0.5/0.6/0.7/0.8/0.9/1.0
• Plot the accuracy vs training data (baseline 1) see acc_vs_size.png

• L_avg -> L_max experiment with the same setup
• No random selection, just a sampling w/ replacement
• Refactor the code
• Re-run the baseline 1
• Run the baseline 2
• Plot the accuracy vs training data (baseline 2)
• Consider unbiasing the gradients samples are pretty uniform
• Consider dropping one fc layer the reason is about distribution difference

• Consider a few tricks
  • Re-initialize everything effective but not much
  • Keep a validation and learn adverserial only on the validation (no contamination since actual network never sees it) effective but not much
• Sample new data with the learned model
• May be a diversity trick? (still a valid thing) does not seem necessary since tSNE is pretty diverse
  • Diversity is a submodular function if defined as sum of total probability covered around each ball
  • Theory suggests a covering ball so let's use that
• Combinatorial Algorithm: start with greedy 2-OPT solution, then refine it using integer programming and binary search if feasible. this is pretty feasible actually somehow Gurobi is more efficient than greedy one to improve the solution
• To match theory and practice, put feature learning in both players
• Include gradient reversal layer
  • Seems like best option for now
  • Step 1: vanilla reversal Note: ADAM is using second derivative which is crap in adverserial setting so use momentum
  • Step 1.5: Implement reversal with single output (so it can learn data distribution)
  • Step 2: vanilla(so) reversal+loss_rescale
  • Step 3: Reversal(so) domain estimate + sampling
  • Step 4: Reversal(so and not/so) + combinatorial sampling (this is desired simply because theory)
• Try with oracle loss still worse than random may be it is bringing sort of a bias
• Look at the tSNE plot and see is it because of diversity it is pretty diverse
• Exploration works so test different degree of exploration 0.2 seems like a nice one, may be 0.25
• Consider normalizing stuff since they become crazy (may be remove batch norm)
• Use BiGAN or ALI as semi-supervised algorithm

• k-k^\prime
• maximum uncertanity
• uncertanity based sampling
• oracle uncertanity based sampling

• Modify the loss/adversery to be applicable to domain adaptation

• Combine everything

• 109 0/5k/10k/15k
• 110 20k/25k/30k/35k
• 106 40k/45k

50,55,62,68,70,73,76,78,79,81

• get the top 5000 expected loss make it 1 rest 0
• combine with gamma = 0.01 or 0.02

• Descrive the active learning with pool problem, state it is a weakly supervised problem and need to be treated like one
• Discuss p(x) p_n(x) p_\hat{n}(x) and give the basic idea behind loss re-scale and discuss how it can be discussed through alternating minimization (Adverserial Weak Supervision)
• Discuss the theoretical aspect of active learning
  • Review robustness and generalization
  • Lemma (VGG is robust)
  • Theorem: Any robust algorithm is robust with less samples if ()
• Discuss the empirical setup and explain the two concepts (fixed budget, single step)
  • Representations should be as close as possible so it is easier to cover the same space with less points
    • Gradient reversal layer
  • The bound depends solely on (\gamma) hence solve combinatorial optimization to have minimum ball
    • Binary search over a submodular problem
• Experiments
  • MNIST
  • Cifar 10 / Cifar 100 on VGG

• Get the features and look at the tSNE
• Sample far points and try this
• Implement the combinatorial algorithm for N-D, try with 2D t-SNE points
• Experiment the active learning