5 Must-Read Data Science Papers (and How to Use Them)


By Nicole Janeway Bills, Data Scientist at Atlas Research.

Photo by Rabie Madaci on Unsplash.

Data science is perhaps a younger area, however that doesn’t imply you received’t face expectations about having an consciousness of sure subjects. This article covers a number of of an important latest developments and influential thought items.

Topics lined in these papers vary from the orchestration of the DS workflow to breakthroughs in quicker neural networks to a rethinking of our elementary method to downside fixing with statistics. For every paper, I provide concepts for how one can apply these concepts to your individual work


#1 — Hidden Technical Debt in Machine Learning Systems


The workforce at Google Research supplies clear directions on antipatterns to keep away from when organising your data science workflow. This paper borrows the metaphor of technical debt from software program engineering and applies it to data science.

by way of DataBricks.

As the subsequent paper explores in better element, constructing a machine learning product is a extremely specialised subset of software program engineering, so it is smart that many classes drawn from this self-discipline will apply to data science as properly.

How to use: observe the specialists’ sensible suggestions to streamline growth and manufacturing.


#2 — Software 2.0


This traditional submit from Andrej Karpathy articulated the paradigm that machine learning fashions are software program functions with code based mostly on information.

If data science is software program, what precisely are we constructing in the direction of? Ben Bengafort explored this query in an influential weblog submit known as “The Age of the Data Product.”

The information product represents the operationalization part of an ML undertaking. Photo by Noémi Macavei-Katócz on Unsplash.

How to use: learn extra about how the info product suits into the mannequin choice course of.


#3 — BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding


In this paper, the workforce at Google Research put ahead the pure language processing (NLP) mannequin that represented a step-function enhance in our capabilities in for textual content evaluation.

Though there’s some controversy over precisely why BERT works so properly, this can be a nice reminder that the machine learning area might have uncovered profitable approaches with out absolutely understanding how they work. As with nature, synthetic neural networks are steeped in thriller.

In this pleasant clip, the Director of Data Science at Nordstrom explains how synthetic neural nets draw inspiration from nature.

How to use:

  • The BERT paper is imminently readable and comprises some urged default hyperparameter settings as a useful place to begin (see Appendix A.3).
  • Whether or not you’re new to NLP, try Jay Alammar’s “A Visual Guide to Using BERT for the First Time” for an enthralling illustration of BERT’s capabilities.
  • Also, try ktrain, a bundle that sits atop Keras (which in flip sits atop TensorFlow) that permits you to effortlessly implement BERT in your work. Arun Maiya developed this highly effective library to allow velocity to perception for NLP, picture recognition, and graph-based approaches.


#4 — The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks


While NLP fashions are getting bigger (see GPT-3’s 175 billion parameters), there’s been an orthogonal effort to discover smaller, quicker, extra environment friendly neural networks. These networks promise faster runtimes, decrease coaching prices, and fewer demand for compute assets.

In this groundbreaking paper, machine learning wiz children Jonathan Frankle and Michael Carbin define a pruning method to uncover sparse sub-networks that may attain comparable efficiency to the unique, considerably bigger neural community.

by way of Nolan Day’s “Breaking down the Lottery Ticket Hypothesis.”

The Lottery Ticket refers to the connections with preliminary weights that make them significantly efficient. The discovering presents many benefits in storage, runtime, and computational efficiency – and received a greatest paper award at ICLR 2019. Further analysis has constructed on this method, proving its applicability and making use of it to an initially sparse community.

How to use:

  • Consider pruning your neural nets earlier than placing them into manufacturing. Pruning community weights can cut back the variety of parameters by 90%+ whereas nonetheless reaching the identical stage of efficiency as the unique community.
  • Also, try this episode of the Data Exchange podcast the place Ben Lorica talks to Neural Magic, a startup that’s wanting to capitalize on strategies resembling pruning and quantization with a slick UI that makes reaching sparsity simpler.

Read extra:


#5 — Releasing the death-grip of null speculation statistical testing (p < .05)


Classical speculation testing leads to over-certainty and produces the false concept that causes have been recognized by way of statistical strategies. (Read extra)

Hypothesis testing predates the usage of computer systems. Given the challenges related to this method (resembling the truth that even statisticians discover it almost unimaginable to clarify p-value), it might be time to take into account options resembling considerably exact end result testing (SPOT).

“Significant” by way of xkcd.

How to use:

Sign up to get notified when “Resources to Supercharge your Data Science in the Last Months of 2020” comes out

Original. Reposted with permission.


Bio: Nicole Janeway Bills is a machine learning engineer with expertise in business and federal consulting. Proficient in Python, SQL, and Tableau, Nicole has enterprise expertise in pure language processing (NLP), cloud computing, statistical testing, pricing evaluation, and ETL processes, and goals to use this background to join information with enterprise outcomes and proceed to develop technical skillsets.



Source hyperlink

Write a comment