Written by Younes Bensouda Mourri & Adam Hodges

We live in a time of rapidly accelerating advancements in technology. This can be seen in the field of machine translation which, over the course of a few years, has gone from rule-based translation techniques to neural network techniques to the most recent approach known as the Transformer—a novel neural network model developed by Ashish Vaswani and colleagues.

As fields like machine translation rapidly evolve, online courses must teach content that keeps up with the latest advances. How can educators in technology fields incorporate the latest research into their teaching and make complex ideas accessible to learners?

In this article, we illustrate how to apply Coursera’s pedagogy best practices to an approach adopted by the deeplearning.ai team for crafting lessons based on the latest research.

Assessing the Current Literature  

The first step is to assess the research literature to determine what advances are worth teaching. As you read research in your field, ask yourself these three questions.

1. Does the research stand out?

Take, for example, Vasawani and colleagues’ paper, “Attention is All You Need,” in which they introduce the Transformer, a novel network architecture “based solely on attention mechanisms.”

This paper made a substantial impact on machine translation techniques by allowing for faster computation and better results. Previous models relied on a general-purpose encoder-decoder framework. In that previous framework, translating an English sentence (e.g., I am happy because I am learning) into its Arabic equivalent (أنا سعيد لأنني أتعلم) would first require processing “I am happy because I am” before translating the word “learning.” This sequential computation takes time.

The Transformer model introduced a way to use parallel computing to achieve this task much more quickly, representing an important breakthrough in machine translation techniques.

2. Are the ideas useful in industry?

Answering this question allows us to understand the incentives for teaching a new approach.

Prior to 2014, Google used statistical machine translation for their Google translate system and then moved to neural machine translation. But today, their engineers have adopted the Transformer framework. This will have profound impact on how machine translation is used industry-wide as demand continues to grow for effective machine translation tools.

According to Grand View Research, the machine translation market is expected to reach $983.3 million by 2022. Learners entering this market obviously need to be well-versed in the latest machine translation techniques. Since the Transformer model was developed by engineers working at Google who have started to implement it in their own translation systems, the ideas are clearly relevant to industry applications.

3. Will the ideas stand the test of time?

No one wants to spend time learning something that will become obsolete in the next few months. So before deciding whether to teach the latest research, it’s important to consider the value of doing so for learners entering a fast-evolving field.

Vasawani and colleagues’ paper is one of the first research papers to not only introduce parallel computing to machine translation tasks but to also show how to reduce sequential computation in those tasks. It marks a transition into a new era of machine translation.

The foundational nature of the paper means others will continue to build upon its ideas. Teaching the ideas will allow learners to start with a solid foundation in the current state of the art even as the field continues to rapidly evolve.

Developing Assignments and Lectures

Once you’ve identified current research advances you want to incorporate into your course, the next step is to use Coursera’s “backwards design” approach to create a lesson that aligns learning objectives with appropriate assignments and instructional materials.

1. Define the Learning Objectives

Learning objectives should use specific, measurable action verbs to clearly define the desired outcome for learners (e.g., By the end of this lesson, you will be able to apply the Transformer model to a machine translation problem). Once the end goal is defined, then work backwards to develop an assessment that will allow learners to demonstrate that outcome.

2. Create the assessment

Based on the learning objective above, we might create an assessment where learners need to write code for our previously mentioned translation problem—translating the English sentence (I am happy because I am learning) into Arabic (أنا سعيد لأنني أتعلم).

In this task, we want the learners to focus on writing code that brings the most value to the new Transformer model—in this case, we want them to write code that allows for parallel computing in this machine translation problem. To scaffold this task, we might provide them with some starter code. In the instructions, we would direct them to add the code for parallel computing according to the techniques of the Transformer model.

3. Create the lectures

The final step in the “backwards design” process is to devise the instructional materials that will teach the concepts and skills learners need to know to be successful in the assessment.

In our case, we want learners to code a particular part of the Transformer model. So, we need to develop lectures that explain the “self-attention” mechanism and demonstrate how it is used for parallel computing.

The key concepts and skills learners need to know will be presented in videos of 4-7 minutes in length. Each short video lecture will cover a particular idea. When stacked together, they will provide the instruction the learners need to be successful in the assessment and to fulfill the intended learning outcome.

Machine translation is certainly not the only field that is rapidly evolving. All technology educators are faced with the challenge of keeping course content fresh and relevant. Adapt the ideas found here to your own domain of expertise to keep your courses aligned with the latest advances in your field.

Younes Bensouda Mourri teaches courses in Artificial Intelligence, Machine Learning, and AI pedagogy at Stanford University where he is a graduate student. He developed content for Coursera’s most popular Machine Learning and Deep Learning courses.

Adam Hodges, Ph.D., is a Sr. Teaching & Learning Specialist at Coursera. He works with university and industry partners to help them produce effective courses on Coursera.

The post How to Design Lessons Based on Current Research appeared first on Coursera Blog.