Technical Summary: Teaching Machine Learning to Software Engineers

Problem Statement

While Machine Learning (ML) and Artificial Intelligence (AI) components are increasingly embedded in software products, undergraduate Software Engineering (SE) curricula lack systematic preparation for building, testing, deploying, and maintaining AI/ML-based systems. Although universities have introduced AI/ML minors, these often treat the discipline as separate from core SE education rather than integrating necessary competencies into the foundational curriculum. Consequently, graduates are frequently ill-equipped to handle the unique lifecycle challenges of ML systems, such as data dependencies, model versioning, monitoring, and the integration of probabilistic behaviors into deterministic software architectures.

Methodology

The authors employed a two-step, evidence-based approach to identify gaps and prioritize integration strategies:

1. Topic Framework Construction (RQ1):

   • Sourcing: A comprehensive inventory of ML-related topics was compiled from two sources: (S1) peer-reviewed literature on SE for ML (2020–2025) and (S2) practitioner descriptions of production ML pipelines.
   • Categorization: Through an iterative normalization process, 36 distinct topics were organized into five categories:
     1. Mathematical Foundations (MF): Concepts for reasoning about ML behavior (e.g., probability, linear algebra, optimization).
     2. Data Structures and Algorithms (DSA): Algorithmic concepts specific to ML computation (e.g., tensors, computational graphs, specific search structures).
     3. Platform and Computing (PC): Infrastructure for scalable ML workloads (e.g., storage, HPC, networking).
     4. Model Development, Maintenance, and Operations (MDMO): Lifecycle management including testing, deployment, monitoring, and versioning.
     5. Applications of ML in Software Engineering (AML): Using ML to support SE tasks (e.g., automated debugging, code quality analysis).
   • Curriculum Mapping: The framework was mapped against required core courses from five representative universities (U1–U5) to identify consistently covered topics versus gaps.

2. Instructor Survey (RQ2 & RQ3):

   • Participants: 36 valid responses from SE instructors across 36 universities in North America, Asia, Europe, and Australia.
   • Instrument: A survey asking instructors to classify each topic as: (a) covered to expected depth, (b) not covered but should be, (c) not covered and not relevant, or (d) unsure.
   • Analysis: Responses were aggregated to determine coverage rates and support levels (prioritization) for each topic. Qualitative data from open-ended questions were analyzed to identify barriers to integration.

Key Results

1. Coverage Gaps (RQ2)

• Mathematical Foundations (MF): Basic topics (probability rules, gradient descent, evaluation metrics) show high coverage. However, advanced topics such as Maximum-Likelihood Estimation (MLE), matrix decompositions, and advanced optimization methods remain underrepresented.
• Data Structures and Algorithms (DSA): Core topics like tensors and computational graphs have moderate-to-high coverage. Specific structures like KD-trees and Autoencoders show low coverage despite being relevant.
• Platform and Computing (PC): Core infrastructure (databases, data pipelines, containers/VMs) is well-covered. Distributed storage, high-performance computing (HPC), and load balancing show moderate coverage with significant variation.
• MDMO and AML: These categories exhibit the most significant gaps. Topics regarding model lifecycle management (versioning, monitoring), data-centric requirements, prompt engineering, and automated debugging have low reported coverage across curricula.

2. Instructor Priorities (RQ3)

• High Support / Low Coverage: There is a strong consensus among instructors that MDMO and AML topics are critical but currently missing. Specifically, model monitoring, debugging, ethical compliance, and using ML for code quality/security are viewed as high-priority inclusions.
• Low Support / Low Coverage: Topics such as data lakes, ML-specific compilers, and networking for ML received low support for inclusion, likely due to prerequisite complexity or perceived lower relevance for general SE roles.
• Barriers to Integration: Instructors cited three primary challenges:
  1. Expertise Gaps: A lack of faculty with dual expertise in SE and ML.
  2. Overloaded Curricula: Difficulty adding new content without displacing core SE fundamentals.
  3. Process and Resources: Bureaucratic hurdles in curriculum revision and the need for computational resources (e.g., GPUs) for hands-on learning.

Key Contributions

• Structured Topic Inventory: The paper provides a defined, operationalized inventory of 36 ML-related topics relevant to SE, categorized into five distinct domains.
• Empirical Gap Analysis: It offers the first cross-curricular analysis mapping these topics against existing SE requirements, identifying specific areas where education lags behind industry practice.
• Prioritization Framework: By combining coverage data with instructor sentiment, the authors derive a prioritization matrix that distinguishes between "foundational" (high coverage/high support) and "critical gaps" (low coverage/high support) topics.
• Integration Guidelines: The study proposes that high-priority topics (specifically MDMO and AML) should be integrated incrementally into existing SE courses rather than requiring entirely new standalone courses, addressing the constraint of overloaded curricula.

Significance and Claims

The paper claims to provide evidence-based guidance for integrating AI/ML content into core SE education. It argues that while universities have begun offering AI minors, the critical need is to embed ML competencies directly into the standard SE curriculum to prepare engineers for the reality of maintaining AI-driven systems.

The authors position their work as a bridge between theoretical engineering challenges (identified in prior literature) and practical curriculum design. By identifying that Model Development, Maintenance, and Operations (MLOps) and Applications of ML in SE are the most significant gaps with the highest instructor support, the paper advocates for a shift in focus: moving beyond just teaching ML algorithms to teaching the engineering practices required to deploy and maintain them. The study concludes that addressing these gaps through modular integration is the most feasible path forward, given current institutional constraints.