Adversarial Domain Adaptation (ADA)

Multi-class classification of images coming from multiple sources (domains).

• Objective

• DataSet

• Methodology

• Results

• Conclusion

• The primary objective of this project is to implement the algorithm of domain-adaptation, which is based on adversarial training (through negative-gradient) from the paper 1 and is in the main branch.

• The models are trained on 11-classes and from three (3) sources and tested on the fourth source.

• To perform a preliminary analysis on the performance of models:

• Trained with sources-class confounding alternatively (similar to GAN) to learn features independent of sources of origin of data. The code is in branch ada_confound.

• A plain naive-training i.e. what if we consider the all source data together and simply train it as a plain multi-class classification problem. The code is in branch plain.

• Data is obtained from 2019 Visual Domain Adaptation Challenge. For simplicity, we only pick 11 classes, some of which are similar and only four sources.

• Class Labels:

  • [0: shorts, 1: snake, 2: table, 3: tiger, 4: toothbrush, 5: tree, 6: truck, 7: van, 8: wine_glass, 9: wristwatch, 10: yoga]

• Source Labels:

  • Training: [0: quickdraw, 1: real, 2: sketch]
  • Testing: [3: infograph]

• Image Exmaples:

Real
Sketch
Quickdraw
Infographs
Src/Cls Labels	Shorts	Snake	Table	Tiger	Toothbrush	Tree	Truck	Van	Wine Glass	Wrist Watch	Yoga

Figure 1. Sample Data (source)

• Download the dataset used in this project from here.

• To download the entire dataset, go to the official page to the VisDA challenge.

Adversarial domain adaptation by Ajakan1 et. al 1 proposes a novel way to learn a representation of data that tries to be independent of the source (domain/origin) of data. To achieve this gradient from source-class loss is reversed to update the parameters of feature-learner. A visual representation is given below:

Figure 2. Model Architecture (source)

• The network is divided into three parts:

  • Feature Representation (FR): the deep architecture in green.
  • Label Predictor (LP): in blue.
  • Domain Predictor (DP): in pink.

• To update the parameters of FR, the network computes the forward pass for all the parts but the total loss is LP loss minus DP loss. Thus gradient being reversed from DP. Please note that only the parameters of FR are updated here.

• Next, is to update the parameters of the part of the network that learns Labels (cls-lbl) and Domains (src-lbl) discriminator respectively.

• Forward pass for the LP and DP network and respectively update the parameters.

• We observe that training is adversarial since updating FR params gradient is reversed, while for Label and Domain parts gradient flows as usual.

• The following experts from the paper summarize the computation well:

Figure 3. Loss Functions (source): Equations 13, 14, & 15 repectively gives the update rule for the params of the three-part network.

• Models are trained for all three different approaches. On average, 400 images are used per-class per-domain, which makes a total of 13K.

• The results presented here in Table 1 are from a preliminary analysis to check which of the three approaches works best. Is the paper (1) claims are on the right track or does plain-classification network also work equally well?

• Results from models saved at different checkpoint and top-3 are presented in the table as follows:

Table 1: Top-3 checkpoints results

• Although, for 11 classes, the result for even the best approach (ADA Paper) is just better than random guessing (100/11 = 9.1%) but provides the empirical-evidence to explore the adversarial-domain adaptation approach much further in case of multi-source data.

• Moreover, the models are trained only on a very small dataset i.e. 400 images per class and source and despite that, naive training clearly shows the difference.

A lot of experimentation is needed to see the method's efficacy and also compare with different adaptation techniques.

[1]. Ganin, Yaroslav, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. "Domain-adversarial training of neural networks." The Journal of Machine Learning Research 17, no. 1 (2016): 2096-2030.

[2]. Ajakan, Hana, Pascal Germain, Hugo Larochelle, François Laviolette, and Mario Marchand. "Domain-adversarial neural networks." arXiv preprint arXiv:1412.4446 (2014).

[3]. Visual Domain Adaptation Challenge (VisDA-2019), "Multi-Source Domain Adaptation and Semi-Supervised Domain Adaptation" tasks, 2019-2020