It presents an informal discussion about do’s and dont’s of research — more specifically experimental study of algorithms. The author presents 10 important principles backed by examples from his long experience. I am listing points from the reading.

#### 1. Solve a problem worth solving.
Take some time to find a good problem. Decide what questions you want to answer before you actually start finding an answer. Finding a new solution would take up time and resources. So think before you compute. Do exploratory research and experimentation. But do not get into the endless loop of exploration.

#### 2. Tie your paper to the literature.
Know the prior art for your problem. Find what are the existing solutions. Think if another solution is actually needed and if yes, what needs to be improved. Tell how your work relates to existing literature. This will help you and your readers. It will make your experiments newsworthy.

#### 3. Use instance testbeds that can support general conclusions.
When doing research, especially experimental research, you are highly likely to use data — for making some hypothesis, for validating results and so on. More general your data set is, more general or applicable your results would be. Applying conclusions drawn from a specific data set on a general problem would be disastrous.

#### 4. Document everything.
Document your code. Document your data sets. Document your experimental setup, your results, your mistakes, your conclusions. Document everything that some other person would need to replicate your experiments. This will help your future-self save some time and he would be grateful to you. Read #6. When using randomly generated data sets, use the same data set for benchmarking all approaches. It would reduce time and variance. Use sampling and bootstrapping.

#### 5. Use reasonably efficient implementations.
Seems obvious. But efficiency requires effort and we may be tempted especially when execution time in not a metric of interest. More efficient code means you can perform more experiments and on larger data sets. But do not over-optimise. Remember — Premature optimization is the root of all evil.

#### 6. Ensure Reproducibility.
Provide information on how to reproduce your experiments. Read #4. See gitxiv. Share your code and data set with others. If the data set is generated using some algorithm, share its details. Use reproducible standards of comparison. Do not say “results after running the algorithm for 1 hour”. Machines are evolving every day and hence doing more operations than before for the same amount of time.

#### 7. Ensure Compatibility.
Comparability goes beyond reproducibility and documentation. Benchmark the machine speed and the other experimental setup that you use so that other researchers can accurately compare their result with yours. Take a full backup of your code and data.

#### 8. Report the full story.
Report the complete story, including the parts you do not like. Do not hide anomalous results. Instead, try to explain them. It could lead to more insights to the problem and to the solution. Report the full running time, even if time is not a metric of interest from the problem’s point of view. Many times readers want to see how your solution performs on the scale of time.

#### 9. Draw well-justified conclusions and look for explanations.
Summarize trends, infer implications and provide conclusions. If there are no conclusions and inferences, then you fail the newsworthiness test. Probably you picked the wrong question or your solution is incomplete. Justify your conclusions with data. If needed, use techniques like profiling.

#### 10. Present Your Data in Informative Ways.
Your data representation should highlight what you want the reader to infer from the data. This can include things like appropriate ordering of columns in a table or choice of variables to be shown along the axis. If you want the reader to compare percentage rise in time, provide a column for that. Do not make the reader do the arithmetic. Take care of not overwhelming your readers with data alone. Put additional data in appendix.

In the area of explaining model predictions over images, there are two main strains of technique: methods that look for pixels that have the highest gradient effect on the output class, and assign those as the “reason” for the class, and approaches that ask which pixel regions are most responsible for a given classification, in the sense that the classification would change the most if they were substituted with some uninformative reference value. 

The tricky thing about the second class of methods is that you need to decide what to use as your uninformative fill-in value. It’s easy enough to conceptually pose the problem of “what would our model predict if it couldn’t see this region of pixels,” but as a practical matter, these models take in full images, and you have to put *something* to give the classifier in a region, if you’re testing what the score would be if you removed the information contained in the pixels in that region. What should you fill in instead? The simplest answers are things like “zeros”, or “a constant value” or “white noise”. But all of these are very off-distribution for the model; it wouldn’t have typically seen images that resemble white noise, or all zeros, or all a single value. So if you measure the change in your model score from an off-distribution baseline to your existing pixels, you may not be getting the marginal value of the pixels, so much as the marginal disutility of having something so different from what the model has previously seen. There are other, somewhat more sensible approaches, like blurring out the areas around the pixel region of interest, but these experience less intense forms of the same issue. 

This paper proposes instead, using generative models to fill in the regions conditioned on the surrounding pixels, and use that as a reference. The notion here is that a conditioned generative model, like a GAN or VAE, can take into account the surrounding pixels, and “imagine” a fill-in that flows smoothly from the surrounding pixels, and looks generally like an image, but which doesn’t contain the information from the pixels in the region being tested, since it wasn’t conditioned on that. 

https://i.imgur.com/2fKnY0M.png

Using this approach, the authors run two types of test: one where they optimize to find the smallest region they can remove from the image, and have it switch class (Smallest Deletion Region, or SDR), and also the smallest informative region that can be added to an otherwise uninformative image, and have the model predict the class connected to that region. They find that regions calculated using their generative model fill in, and specifically with GANs, find a smaller and more compact pixel region as their explanation for the prediction, which is consistent with both human intuitions and also with a higher qualitative sensibleness of the explanations found. 

Stanford’s paper on Global Vectors for Word Representation proposes one of few extremely popular word embedding methods in NLP. GloVe takes advantage of both global corpus statistics and local context window methods by constructing the word-context co-occurrence matrix and reducing its dimensionality by preserving as much of the variance as possible. It builds a feature space with additive compositionality while preserving statistically meaningful word occurrence information extracted from the data corpus.

Authors start with building the counts matrix X, where X_ij is the number of times word w_j appears in the context w_i. The corresponding probabilities can be calculated as P_ij = X_ij/X_i. The GloVe model tries to fit a function F(w_i, w_j, w.hat_k) = P_ik/Pjk, which represents ratio of probabilities of the word w.hat_k appearing in the context of words w_i and w_j respectively. The model is expected to produce large output for w.hat_k relevant to w_i and not w_j and vice versa. When w.hat_k is equivalently relevant or irrelevant to w_i and w_j, the output is close to unity. For practical reasons, the authors simplified the model with a linear equation w_i.T*w.hat_k + b_i + b_j = log(1+X_ik), where the biases b_i, b_j represent log(1+X_i) and log(1+X_j) respectively. As an optimization objective, GloVe uses weighted squared error J = Sum_ij(f(X_ij)*(w_i.T*w.hat_k + b_i + b_j - log(1+X_ik))^2), where the weighting function mitigates the effect of noisy rare word occurrences (X_ij ~ 0). The choice of the weighting function f(x) is somewhat arbitrary and driven by empirical observations. The best performing f(x) = min{1, (x/x_max)^0.75}, with x_max = 100. Authors refer to 0.75 power of the skip-gram model, which appears to be used in negative sampling distribution in skip-gram model, which also functions as a weighting coefficient. The resulting log-bilinear regression model is then optimized with AdaGrad and the combination of the resulting matrices W + W.hat is used as the word embeddings.

Authors evaluate GloVe performance on a variety of NLP tasks and compare against Skip-Gram, CBOW, SVD and HPCA. Although GloVe demonstrated advantageous metrics in both performance and training time (co-occurrence matrix is built in parallel), the model is not very different from Word2Vec. Another questionable argument of the GloVe authors is that the log squared error function is superior to the cross entropy loss in Skip-Gram and ivLBL models, because of higher robustness in the long tails distributions. However, it is not apparently correct, as Skip-Gram uses stochastic optimization, which inherently mitigates the long tails vulnerability. In fact, GloVe shows similar performance to Word2Vec in numerous NLP problems, yet the later historically gained a larger popularity.

This paper proposed a way to do classification using primitive concepts such as color, shape, texture, etc.

The framework is simple, they have two sub-models:

(1) the first one is a trained VL model such as CLIP,  ViLT, and ALBEF. The input of this step is the primitive concepts or let's say, attribute concepts and an image, then the output will be the scores for each concept.

(2) the second one is a linear model that uses the concepts and their scores to do classification. This model is trained in a supervised manner.

https://i.imgur.com/7WMmGyv.png

Kanbak et al. propose ManiFool, a method to determine a network’s invariance to transformations by iteratively finding adversarial transformations. In particular, given a class of transformations to consider, ManiFool iteratively alternates two steps. First, a gradient step is taken in order to move into an adversarial direction; then, the obtained perturbation/direction is projected back to the space of allowed transformations. While the details are slightly more involved, I found that this approach is similar to the general projected gradient ascent approach to finding adversarial examples. By finding worst-case transformations for a set of test samples, Kanbak et al. Are able to quantify the invariance of a network against specific transformations. Furthermore, they show that adversarial fine-tuning using the found adversarial transformations allows to boost invariance, while only incurring a small loss in general accuracy. Examples of the found adversarial transformations are shown in Figure 1.

https://i.imgur.com/h83RdE8.png
Figure 1: The proposed attack method allows to consider different classes of transformations as shown in these examples.

Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/).