- 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 = 5000Random active learning test 0.1/0.2/0.3/0.4/0.5/0.6/0.7/0.8/0.9/1.0Plot the accuracy vs training data (baseline 1)see acc_vs_size.png
L_avg -> L_max experiment with the same setupNo random selection, just a sampling w/ replacementRefactor the codeRe-run the baseline 1Run the baseline 2Plot the accuracy vs training data (baseline 2)Consider unbiasing the gradientssamples are pretty uniformConsider dropping one fc layerthe reason is about distribution difference
- Consider a few tricks
Re-initialize everythingeffective 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 diverseDiversity is a submodular function if defined as sum of total probability covered around each ballTheory 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 solutionTo match theory and practice, put feature learning in both players- Include gradient reversal layer
- Seems like best option for now
Step 1: vanilla reversalNote: ADAM is using second derivative which is crap in adverserial setting so use momentumStep 1.5: Implement reversal with single output (so it can learn data distribution)Step 2: vanilla(so) reversal+loss_rescaleStep 3: Reversal(so) domain estimate + samplingStep 4: Reversal(so and not/so) + combinatorial sampling (this is desired simply because theory)
Try with oracle lossstill worse than random may be it is bringing sort of a biasLook at the tSNE plot and see is it because of diversityit is pretty diverseExploration works so test different degree of exploration0.2 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 0combine 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
- Representations should be as close as possible so it is easier to cover the same space with less points
- 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