Machine Learning: The High Interest Credit Card of Technical Debt

By W. B. Roberts, David Sculley, Gary Holt, D. Golovin, Eugene Davydov, Todd Phillips, D. Ebner, Vinay Chaudhary, Michael Young. 2014

328 citations recorded by the source corpus.

Subfields: Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Safety, Robustness & Interpretability

Citations in the corpus

PromptChainer: Chaining Large Language Model Prompts through Visual Programming (2022, background): “3 The likelihood of cascading errors: The black-box nature of LLMs means that sub-optimal or even unsafe output in a single step could potentially lead to cascading errors across an LLM chain (see [14] for a similar observation in traditional machine learning pipelines).”
Large scale distributed neural network training through online distillation (2018, background): “To reproduce a given final model, one needs the entire history of teacher models and everything required to reproduce them which can easily result in what Sculley et al. (2014) refers to as “pipeline jungle” and unnecessary, undesirable data dependencies.”
BoostClean: Automated Error Detection and Repair for Machine Learning (2017, background): “Recent papers and surveys of analysts suggest that such problems are pervasive [27,31,47].”
Failing Loudly: An Empirical Study of Methods for Detecting Dataset Shift (2018, background): “Unfortunately, software systems based on machine learning are notoriously hard to test and maintain [42].”
Deep Learning: A Critical Appraisal (2018, background): “As a team of authors at Google put it in 2014, in the title of an important, and as yet unanswered essay (Sculley, Phillips, Ebner, Chaudhary, & Young, 2014), machine learning is “the high-interest credit card of technical debt”, meaning that is comparatively easy to make systems that work in some…”
Closing the AI accountability gap: defining an end-to-end framework for internal algorithmic auditing (2020, background): “rtificial intelligence is a tool that mixes data sources together. As Scully et al. point out, artificial intelligence models create entanglementandmaketheisolationofimprovementseffectivelyimpossible [67], which they call Change Anything Change Everything. We suggest that by having explicit document”
DeepTest: Automated Testing of Deep-Neural-Network-Driven Autonomous Cars (2017, methodology): “it a bus while shifting lane "The car assumed that the bus would yield when it attempted to merge back into traffic" [12] systemsincludingself-drivingcar,speechrecognition,imagesearch, etc. [22, 73]. Our experience with traditional software has shown that it is hard to build robust safety-”
Data Management Challenges in Production Machine Learning (2017, background): “As stated in [16], it is crucial to clearly identify explicit and implicit data dependencies (and accordingly avoid needlessly introducing new dependencies) in order to develop a maintainable machine learning pipeline.”
NAIS: Neural Attentive Item Similarity Model for Recommendation (2018, methodology): “For the embedding size k, we test the values of [8, 16, 32, 64], and set the attention factor a same as the embedding size in each setting.”
Clinical artificial intelligence quality improvement: towards continual monitoring and updating of AI algorithms in healthcare (2022, background): “It can be dif ﬁ cult to anticipate how these models will behave in real-world settings over time, as their complexity can obscure potential failure modes 8 .”
A Survey of Modern Questions and Challenges in Feature Extraction (2015, methodology): “…projection, clustering, nearest neighbor, dimensionality reduction, manifold learning, distance metric learning, multiple kernel learning, representation learning, embedding, PCA, LDA, RCA, kernel PCA, Isomap, LLE, Laplacian eigenmap, neural networks, autoencoders, restricted Boltzmann machines”
The ML test score: A rubric for ML production readiness and technical debt reduction (2017, background): “Data 3: No feature’s cost is too much: It is not only a waste of computing resources, but also an ongoing maintenance burden to include -features that add only minimal predictive benefit [1].”
Defining Explainable AI for Requirements Analysis (2018, neutral): “Indeed, while good software engineering practice has resulted in software becoming more modular, reliable, testable, predictable and accountable, machine learned systems are becoming more tightly coupled and their interactions increasingly obfuscated and opaque [25].”
Fast Matrix Factorization for Online Recommendation with Implicit Feedback (2016, background): “More importantly, the low eﬃciency makes it even more dif-ﬁcult to deploy implicit MF method online [29].”
Making Contextual Decisions with Low Technical Debt (2016, methodology): “L G ] 9 M a y 2017 collected to train the system [12, 43].”
A systematic comparison of feature space effects on disease classifier performance for phenotype identification of five diseases (2015, background): “Not all steps in the pipeline are equally easy to implement or maintain in a production environment (Sculley et al., 2014).”