arXiv/viXra - Introduction

Learning the difference with machines.

Which is which?

Here are two research paper titles:

An Efficient Lattice Algorithm for the Libor Market Model

and

Atomic Entanglement vs Photonic Visibility for Quantum Criticality of Hybrid System

One example comes from arXiv and the other from viXra. For the unfamiliar:

• arXiv hosts nearly two million papers and is the standard, global repository for fields such as Computer Science, Statistics, Quantitative Finance, Mathematics, and, in particular, Physics, the discipline for which it was originally established.
• viXra hosts nearly forty thousand papers and was created as an alternative to arXiv for "scientists who find they are unable to submit their articles to arXiv.org because of Cornell University's policy of endorsements and moderation designed to filter out e-prints that they consider inappropriate."

Question: Can you tell which title above came from an arXiv paper and which is from viXra?I guessed incorrectly for both of these titles, which can be found here and here.

Humans vs. Machines

Humans

The general task of determining the source of a given paper from, say, its title or abstract alone is non-trivial, particularly for those with little exposure to technical articles.

At garrettgoon.com/arxiv-vixra-quiz I have made a small quiz where you can test your own abilities. As of November 5, 2021 there have been over 25,000 total guesses with the following results: Participants are requested to self-identify as experts if they regularly read technical research articles. These experts guessed correctly on \approx 78\% of the title questions and \approx 71\% of the abstract questions. Non-experts guessed correctly on \approx 71\% and \approx 66\% of these same tasks. Approximately three-fifths of all guesses came from experts.

• Having seen the title alone, participants guess the correct source \approx 69\% of the time.
• Having seen the abstract alone, participants guess the correct source \approx 75\% of the time.

A summarizing graphic can also be found above.

Machines

Though humans perform significantly better than random, there is room for improvement and surely one can teach a computer to outperform the human baseline arXiv itself uses Machine Learning models to assess its submissions. Published results include studies of plagiarism detection and the correlation between submission time and citation accumulation. Paul Ginsparg (arXiv founder) gives fascinating talks on the subject. . This particular task of text classification is one of the classic problems of Machine Learning (ML), falling under the subcategory of Natural Language Processing (NLP).

In the below series of posts I describe in detail the process of building increasingly sophisticated models for this classification task In order to limit the scope of the project, most of the models focus on the (harder) problem of classifying papers based on title alone. . Along the way, I also attempt to explain various ML and Data Science concepts in my own words with the hope that others may find the explanations informative. These often take the form of (long) footnotes Here's an example: many common computations in ML are paralellized using broadcasting and vectorized code. These are some of the most important practices to develop for efficiency. Coming from physics, the tensor operations one performs in these calculations are also initially somewhat bizarre looking (at least for me) and take some getting used to. Below I explain broadcasting in the way I would have found helpful when initially learning about them.

Broadcasting involves operations between tensors of different sizes, where a tensor is an array-like object T_{i_0 i_2\ldots i_N} with indices i_k\in\{0, \ldots, d_{k}-1\} spanning different ranges, in general. In pytorch, such a tensor is represented as an instance of the torch.Tensor class, call it t, which is said to be (d_0, d_1, ...., d_N)-shaped. In addition to the usual matrix-multiplications and dot-products, in ML one often wants to do something like the following: take (r, c)-shaped matrix and rescale all of the rows by the values contained in an (c, )-shaped vector. This is an uncommon operation in physics, where all tensor manipulations tend to be invariant or co-variant under rotations (or other group transformations) of the constituent tensors. An explicit, simple example with a matrix m and vector v: m = \begin{pmatrix} 2 & 1 \\ 0 & 2 \\ 3 & 3 \end{pmatrix}\ , \quad v = \begin{pmatrix} -1 \\ 0 \\ 4 \end{pmatrix}\quad \longrightarrow \quad m * v \equiv \begin{pmatrix} -2 & -1 \\ 0 & 0 \\ 12 & 12 \end{pmatrix}\ , or in component form: m_{ij}, v_{k} \longrightarrow \left(m*v\right)_{ij} \equiv m_{ij}v_{j}. Because the entries of the m*v tensor can all be computed independently from each other, the calculation is amenable to parallelization and corresponding vectorized pytorch code for the above is simply the concise syntax below (using random entries instead of the preceding integers): m = torch.randn(2, 3) # (2, 3)-shaped. v = torch.randn(3) # (3, )-shaped. mv = m * v # (2, 3)-shaped.
This type of broadcasting works with all of the usual infix operations (*, /, +, -) and is possible for tensors T, U, (t, u in code below) which satisfy the following criteria. Writing out the indices, we take T_{i_M\ldots i_{0}} to be (M+1)-dimensional and U_{j_N\ldots j_{0}} to be (N+1)-dimensional with M \le N (without loss of generality), where the ordering of the index subscripts is reversed from usual conventions for reasons that will soon become apparent. Writing out the sizes of these dimensions in code notation as (t_M, ..., t_1, t_0) and (u_N,..., u_1, u_0) for T and U, respectively, T and U can be broadcast together if the sizes of each of the final M + 1 dimensions of both tensors (all of the ones they have in common) are equal to one other or at least one of the dimensions is a trivial, singleton dimension of size one. That is for each k in (M, ..., 1, 0), at least one of the following must be true:

A) t_k == u_k, so that the indices i_k and j_k span the same range of values.
B) t_k == 1 or u_k == 1, so that at least one of i_k or j_k only takes on the value 0.

When the above holds, T and U can be broadcast into a new tensor V which has N+1 indices V_{j_N\ldots j_{0}}, like the larger tensor U, and is of shape (u_N, ..., max(t_M, u_M), ..., max(t_0, u_0)). For instance: t = torch.randn(5, 1, 3) u = torch.randn(2, 1) v = t + u # (5, 2, 3)-shaped The components of V come from performing the infix operation between T and U using the natural components of these latter tensors. That is, when T, say, has an index which ranges over the same values as the corresponding index in V , then the usual, non-trivial components of T are used. Otherwise, if the relevant index in T is a trivial singleton index or is completely missing (which is equivalent to inserting a dummy singleton index at the relevant location), the single value of the tenor available at this index is used in the computation for all values of the index in V. When a single value is used for multiple components of V in this way, one says that it has been broadcast across a larger dimension. Analogous statements hold for U. In the code example above, this works out as follows: V_{ijk} is a three-dimensional, (5, 2, 3)-shaped tensor whose components are V_{ijk} = T_{i0k} + U_{j0}\ . There is not really any other reasonable way that we could have combined T and U into a V-shaped object. In general, one can think about broadcasting as first taking T and U, inserting dummy singleton indexes into the smaller tensor until the two are of the same dimension (T_{i_M\ldots i_{0}}\longrightarrow T_{i_{N}\ldots i_{M+1}i_M\ldots i_{0}} in the present case, i_k\in \{0\} for k > M), broadcasting all singleton dimensions in both tensors as needed, and then finally performing the relevant infix operation between the now-properly-sized tensors.

A final example of when this is useful and how to translate index notation to vectorized code. Imagine that T_{ik} is an (n_t, d)-shaped tensor corresponding to n_t different vectors (T's first index) which are each d-dimensional (T's second index) and U_{jk} is a similar (n_u, d)-shaped tensor. If we wanted to get the squared-distances between all such vectors, a (n_t, n_u)-shaped tensor V_{ij}, we could write this in index form as V_{ij} = \sum_{k}\left(T_{ik} - U_{jk}\right)^2 \ . In the equivalent vectorized code computation, we first need to re-shape both tensors so that the length-n_t component of t and length-n_u component of u lie along independent dimensions, allowing them to be broadcast together into the desired output shape. The requisite singleton dimensions can be inserted in pytorch as in t.unsqueeze(dim) to perform the insertion at dimension dim or, alternatively, by slicing the Tensor object with None inserted in the appropriate location. Both methods are shown below. After this re-shaping, it is a simple matter of taking differences, squaring, and summing appropriately. Very explicitly: n_t, n_u, d = torch.randint(1, 100, (3,)) # Random dimensions. t = torch.randn(n_t, d) # (n_t, d)shaped. u = torch.randn(n_u, d) # (n_u, d)shaped. t = t.unsqueeze(dim=1) # (n_t, 1, d)-shaped. u = u[None, ...] # (1, n_u, d)-shaped. t_u_diff = t - u # (n_t, n_u, d)-shaped. t_u_diff_squared = t_u_diff ** 2 # (n_t, n_u, d)-shaped. v = t_u_diff_squared.sum(dim=-1) # (n_t, n_u)-shaped.
to which more technical points are also relegated.

Project Posts

Links to all posts in this series. Note: all code for this project can be found on my GitHub page.

Acknowledgments

Thank you to the Distill team for making their article template publicly available. I also gratefully acknowledge useful conversations with Matt Gormley, Matt Malloy, Thomas Schaaf, and Rami Vanguri in the course of this project.