How Deep Learning Techniques are Transforming Ed Tech
Eric Cosyn, Senior Director of Applied Research
The recent breakthroughs in generative AI and large language models created concerns in the education community for their potential misuses on matters such as truthfulness, verifiability, and student’s probity. If caution around these technologies is warranted, it is also doubtless that in time they will make their way into the ed tech products deployed in the classroom and positively impact them. They are the latest advances from the deep learning renaissance initiated in the early ’10s, a period of time where we experienced a series of research breakthroughs in pattern recognition that enable us to leverage artificial neural networks (or, deep learning) in transformative ways.
It’s worthwhile to have a look at how deep learning is already changing ed tech. An adaptive learning system such as ALEKS offers us a good example.
ALEKS uses a cognitive model known as Knowledge Space Theory that captures how students’ knowledge of an academic field is structured. ALEKS relies on this structure to run adaptive assessment engines that draw powerful inferences to uncover a student’s knowledge and to prompt the student with new topics that they are most ready to learn. ALEKS is an empirically grounded model and data shape its knowledge structures and the inferences made by its assessment engines. Deep learning models, on the other hand, are mechanisms that automatically build programs that capture highly non-linear relationships between the data they were exposed to. As such, it was only natural to seek their use to make ALEKS more efficient. Over the last few years, deep learning models made a critical impact in three different areas of ALEKS. Let’s have a look at them in turn.
Finding Subtle Patterns in Assessment Data
The ALEKS adaptive assessment engine updates, after each question, the probability estimates that the student knows any of the other topics in the ALEKS course product. The probabilistic inferences that the engine draws are informed by sophisticated statistical analyses of millions of student assessments. Deep learning neural networks now allow ALEKS to leverage its vast data store even more effectively by finding subtle patterns in the assessment data beyond the reach of traditional methods. The challenge here was to develop a deep neural network architecture that outputs probability estimates consistent with the ALEKS knowledge structure, ensuring that the student’s dynamic learning path is pedagogically sound at all points. The use of probability estimates from the deep learning model accelerates the determination of the student’s knowledge as the assessment proceeds, resulting in an assessment that is both more accurate and shorter by 20%.
Consolidating of Student’s Learning
Consolidation of student’s learning is the second area where deep learning models are making an impact. In ALEKS, students progress through a cycle of multiple learning sessions followed by a progress assessment. The role of the progress assessment is to enforce spacing, retrieval practice, and interleaved practice to consolidate long-term mastery of the learned material. Following the progress assessment, ALEKS may require additional practice for some of the topics recently learned. Recent research demonstrated that, given the highly structured approach to learning in ALEKS — students can only learn those topics for which they have the required background knowledge — , there is sometimes a greater benefit in letting the students learn new material built on top of the topics they just learned than in testing them on these topics. Here again, ad hoc deep learning neural networks trained on hundreds of millions of learning datapoints capture subtle patterns between how a topic was learned together with other relevant features and its long-term retention to determine whether it should be tested or not by the progress assessment. The cumulative effect of this improvement is not marginal: over the length of the course, students cover 9% more material without negative impact on long-term retention.
Gathering Clear, Actionable Data for Instructors
Our last example concerns the information that the learning system reports to the instructor. By design, ALEKS tracks the minute interactions of a student with its adaptive assessment and learning engines. That wealth of information is synthesized in a number of reports that, altogether, delineate a precise picture of the student. These reports allow an experienced instructor to catch changes in the student’s performance that point to possible issues that require intervention. An example would be a slowdown in learning beyond what is expected from tackling more advanced material as the student progresses. But monitoring a large class may quickly become a daunting task for instructors notoriously short of time. Here again we take advantage of the ability of deep learning neural networks to categorize students’ learning patterns. The noteworthy cases are then automatically surfaced in a clear, understandable way for instructors, helping them identify which student needs helps for what. This tool, ALEKS Insights, allows the instructor to spend less time shuffling through data and more time where their action is most critical and valuable, which is on one-on-one interaction with students.
The three examples above illustrate how deep learning techniques are transforming ed tech now. We can confidently say that their impact will only grow. The key takeaway is that these techniques are tools and not substitutes for the model underlying an ed tech product. Their value is only extent to the pedagogical value of the model that makes use of them. The ALEKS learning system is in a privilege situation in that regard. Based on strong foundations and meant to be driven, measured, and validated by data, ALEKS is best positioned to take advantage of the “deep learning” revolution to successfully improve student outcomes. For instance, the emergence of highly functional large language models has the potential to enable the delivery of instructional material in a more conversational mode, where both student’s queries and the system answers are flowing in natural language. Such an application is one of the next directions in the development of more effective instructional systems.
Eric Cosyn co-founded ALEKS Corporation while completing his PhD degree in Mathematical Behavioral Sciences at the University of California, Irvine. Through his long career in Ed Tech development, he advocated for and implemented solutions grounded in strong learning science principles. As Senior Director of Applied Research, he leads the team responsible for developing the mathematical and cognitive models driving knowledge assessment, acquisition, and retention in all ALEKS course products.
For more on AI in K-12 education, see:
References
J. Matayoshi, E. Cosyn, & H. Uzun (2022). Does Practice Make Perfect? Analyzing the Relationship Between Higher Mastery and Forgetting in an Adaptive Learning System. Proceedings of the 15th International Conference on Educational Data Mining.
J. Matayoshi, H. Uzun, & E. Cosyn (2022). Using a Randomized Experiment to Compare the Performance of Two Adaptive Assessment Engines. Proceedings of the 15th International Conference on Educational Data Mining.
J. Matayoshi, E. Cosyn, & H. Uzun (2021). Evaluating the Impact of Research-Based Updates to an Adaptive Learning System. Proceedings of the 22nd International Conference on Artificial Intelligence in Education.
J. Matayoshi, H. Uzun, & E. Cosyn (2020). Studying Retrieval Practice in an Intelligent Tutoring System. Proceedings of the 7th ACM Conference on Learning at Scale.
