Note: This is the first of several lectures from the series on AI in supply chain management I prepared for a course available on MIT Learn called Supply Chain Technology and Systems (SC4x). The visual companion (slides) is available here. This post is a companion written form of the lecture. The lecture videos were designed for the course. If you are interested in having me present this material to your organization, let me know.
New here? This post is part of a MIT AI Lecture series in the newsletter. I also write about AI safety in robotics and wearables, AI ethics, endurance running, and other topics.
A Lecture, Preserved in Writing
It can be awkward to give an introductory lecture on AI in 2026. The field is moving faster than the curriculum can keep up with, and everyone in the room has strong feelings about what “AI” is and what impact it is having on our world. Supply chain management (SCM) had a similar — though a little less charged — moment during the Covid-19 pandemic. Since then, it has become a common part of our vernacular. The inflection point for AI was the introduction of ChatGPT in 2022. Before then, SCM had been quietly running on algorithms for nearly eighty years. Most of those algorithms were not “AI” by the current understanding of the term, but they used many of the methods that now form the basis for modern AI.
This introductory lecture sits at the intersection of SCM and AI. My goal is to provide a map of where the field came from, what its competing intellectual traditions actually are, and why the systems on the current frontier — the ones being built in MIT labs and multi-billion-dollar “startups” — look the way they do. It is, in other words, a vocabulary and a history.
This post is that lecture, preserved in writing. The visual companion (slides) is available online here. I delivered a version of this lecture in SC4x. The reason I’m reproducing the lecture here as a written piece, and not as a video embed, is partly personal: I am a writer and reader by nature. I prefer to read books and articles — on physical paper if available — and perform better on the page than on the stage. It is also for access: A lecture that exists as a written essay alongside its slides can be re-read, quoted, cited, and updated for a decade. A lecture that exists only as a video on a courseware platform cannot.
The central story of the post is this: AI in supply chain management is a story of two intellectual traditions converging. One is mathematical optimization, born in the U.S. Air Force in 1947 and refined for seven decades under the discipline of operations research (OR). The other is machine learning, born in psychology and statistics in the 1950s and refined — through two long winters and a recent massive acceleration — into the family of systems we now call AI. Until about ten years ago, those two traditions evolved on largely separate timelines. They are now converging, and the most consequential systems in SCM over the next decade will be hybrid descendants of both.
The human dimension to this story doesn’t often receive much space in introductory lectures, but it should. The capabilities don’t deploy themselves. Whether the hybrid systems of the next decade actually land safely in a warehouse, a procurement function, or a port depends on a body of work in human-centered AI that I’ll argue, by the end of this post, is not a side topic. It is central to the deployment of these systems.
The lecture is organized into seven parts. This post follows the same arc in descriptive form.
The Problem That Analytics Has Always Been Trying to Solve
Before we dive into the specifics of AI, it’s important that we have a shared understanding of what we mean by “analytics” and how different analytical approaches serve different purposes. This foundation will help us appreciate why certain AI techniques are well-suited to specific supply chain challenges.
One of the most widely used frameworks (DDPP) for categorizing analytics is a four-level progression from descriptive to prescriptive analytics1. It is focused on what kind of question you are asking of the data or model in the problem‑solving process. Descriptive analytics tells you what happened, diagnostic analytics tells you why, predictive analytics tells you what is likely to happen next, and prescriptive analytics tells you what to do about it. The progression in the DDPP framework represents an increasing level of sophistication and value, but also increasing complexity. Each tier requires more from the system underneath it, and each pulls the human further out of the loop.
Another framework for understanding analytical problem-solving is focused on what kind of model you are using2. At its core, analytical problem-solving involves identifying and modeling relationships between outcome variables (Y) and various factors (X). This relationship can be written as Y = f(X).
Mast et al.3 described three types of models. Causal models identify the variables that change Y when you intervene, and they are the only models that support honest “what-if” reasoning. Correlational models are powerful for prediction from observational data but cannot tell you what will happen if you change something. Deductive models derive Y from universal laws (physics, mathematics, probability theory) and are the foundation of optimization. SCM needs all three. Forecasting is correlational. Capacity planning is deductive. Pricing and policy are causal. Conflating them is one way AI projects in supply chains can go wrong.
The third framework I want to mention is the distinction between deterministic and probabilistic models. The Economic Order Quantity formula is deterministic. It assumes a fixed demand rate, lead times, and costs, and produces a clean answer. The world is not deterministic, though. Lead times range from two to sixteen days. Demand follows Poisson or Normal distributions or, more often, distributions we cannot identify. Probabilistic models embrace that uncertainty at the cost of computational complexity and interpretive difficulty. We will encounter this tension across the rest of this post: between provable rigor and adaptive flexibility, between mathematical optimization and machine learning. Every algorithm we will look at sits somewhere on this grid, and the frontier we are heading toward is largely about systems that can move across the grid as conditions change.
Origins in Operations Research: A Mathematician at the Pentagon in 1947
We start this journey by tracing a mathematical approach to the question “what should we do?” in a large network almost eighty years ago. In the summer of 1947, a young mathematician, George Dantzig, was asked by the United States Air Force to mechanize the planning process.
Military logistics are incredibly complex. Imagine trying to optimize resource allocation across a military operation involving thousands of variables and constraints, all calculated by hand. The need for a systematic computational approach was clear. Dantzig was assigned to find a general-purpose method and produced the Simplex Method4.
The idea was revolutionary. Linear programming allows us to optimize an objective function — such as minimizing cost or maximizing throughput — subject to linear constraints. The Simplex Method provides an efficient way to find the optimal solution among many possible solutions. Its practical applications in supply chain management are profound. Transportation planning, production scheduling, inventory optimization, and network design all still rely on the foundations that Dantzig established.
The geometry of the algorithm is easier to draw than describe, but here’s an attempt in three dimensions. Imagine wearing magnetic boots and walking along the outside of a strange metal building — the kind of angular structure you might see on MIT’s campus. The walls are flat, so they meet at sharp corners. You start at one corner, and your goal is to reach the highest point on the building. At each step, you move along an edge to a neighboring corner that takes you higher. Eventually, no adjacent corner is higher than where you stand. This is the intuition behind the Simplex Method applied to a maximization problem.
In more technical terms, a linear programming problem with n variables and m constraints defines a convex polytope in n-dimensional space. The optimal solution sits at one of the polytope’s vertices. The Simplex Method walks from vertex to adjacent vertex along the edges, improving the objective at each step, until it reaches the best one.
Dantzig initially feared the approach would not work in higher dimensions — our magnetic boots intuition only works in three dimensions — because there can be exponentially many vertices in a polytope, and a naive walk could be intractable. In the summer of 1947, the economist Leonid Hurwicz visited the Pentagon and discussed the problem with Dantzig; out of that conversation came an alternative geometric framing that worked by columns rather than rows. They called this “climbing the bean pole”.5
The first real test of the method came in 1948 with Stigler’s minimum-cost diet problem. The challenge was to find the cheapest combination of foods that would meet all nutritional requirements. The Simplex Method solved it effectively, demonstrating its practical power.
As Allyn Jackson6 illustrates in a 75-year retrospective on the Simplex Method, it was the preferred method for linear programming for nearly four decades. In the 1980s, it was challenged by Karmarkar’s (1984) interior-point methods, which traversed the interior of the feasible region rather than its vertices and had better worst-case guarantees. In 2001, Spielman & Teng’s smoothed analysis formally explained why the simplex method works so well in practice on real-world inputs but not on worst-case instances. Today, simplex still ships inside Excel Solver and all common optimization engines.
The Simplex Method and linear programming are part of the broader field of OR, with many other applications in supply chain management. Network optimization handles last-mile routing and global flow problems. Integer programming handles warehouse siting, vehicle dispatch, and production scheduling, where decisions must be whole numbers. Simulation modeling lets us test policies against a virtual distribution center without disrupting the real one. Queueing theory tells us how long the line will be at the inbound dock and what to do about it.
A defining property of these methods is provable rigor. The algorithms used in OR come with mathematical guarantees: an optimal solution when the assumptions hold, and a bounded one when they don’t. This rigor also comes with tradeoffs. Classical OR assumes the world is more deterministic than it is. The problems OR struggles with (uncertainty, noise, novelty) were picked up by a different tradition evolving in mostly different buildings and on mostly different funding lines, over almost exactly the same seventy years.
The Parallel Track: Seventy Years of AI and ML
While operations research focused on mathematical optimization, a parallel evolution was occurring in the fields of symbolic reasoning and machine learning. The history of AI is often characterized by cycles of enthusiasm and disappointment (”AI winters”).
One starting point is Frank Rosenblatt’s perceptron from 1957. This is a precursor to the powerful neural networks used by modern AI systems. Rosenblatt’s perceptron was a learning rule that adjusted weights to classify patterns, and it was limited to linearly separable problems. The limitation, and Minsky and Papert’s 1969 critique of the perceptron, resulted in the first AI winter, with little funding and interest in neural networks7. It took several decades before the field took another significant step forward with John Hopfield’s work in 1982, which introduced one of the earliest recurrent neural network architectures. The 80s also saw Geoffrey Hinton and Terry Sejnowski’s Boltzmann Machines in 1985 and the re-popularization of the backpropagation algorithm by Rumelhart, Hinton, and Williams in 1986. Taken together, these developments enabled efficient training of multilayer networks, but interest stalled due to the challenge of scaling these methods to real-world problems (the second AI winter).
While interest in neural networks was cyclical, an entire generation of other algorithms quietly carried the field forward, including k-Nearest Neighbors (Fix & Hodges, 1950s), Naive Bayes (1950s), Support Vector Machines (Vapnik & Chervonenkis, 1960s), and Decision Trees (Quinlan’s ID3 in the 1980s). Most do not match the power of modern deep learning, but all of them are still in active production use; SVMs in particular remain a default for structured-data classification.
The 1980s and 1990s also saw a significant detour. While statistical learning was struggling through its winter, symbolic AI became the dominant area of research and the first commercially successful AI. These expert systems, or Good Old-Fashioned AI (GOFAI), were a form of rule-based intelligence that encoded human expertise as IF-THEN logic through knowledge engineering8.
IBM’s Logistics Management System is an example of an expert system used for SCM. It was deployed at the company’s Burlington, Vermont, semiconductor facility in the 1980s. Sullivan and Fordyce9 describe the LMS as evolving through three stages: passive decision support (queries, reports, what-if analysis), proactive decision support (automated monitoring and recommendations), and finally, automated decision-making (heuristic-driven execution on a limited set of processes where the rules were understood well enough to support it). LMS won the Franz Edelman Award and is the conceptual ancestor of many SCM systems used today.
But many expert systems died in enterprise deployment. Gill10 argues that early expert systems face two significant challenges. The first challenge was technical and economic. Knowledge acquisition is expensive, rules are hard to maintain, and adapting to evolving domains requires constant re-engineering. The second challenge was human and organizational: misalignment with user needs, conflicts with IT strategy, insufficient long-term ownership, and a lack of mechanisms to transfer knowledge across people and time. We see many of the same failure modes in enterprise AI deployments today. This is particularly true in physical automation, where novel edge cases break rule-based systems exactly as Gill described in 1995. The Friday robotics thread of this newsletter will return to this topic in detail.
The dominance of expert systems was short-lived. The next inflection point came in 2012 with the ImageNet competition and AlexNet. The latter was a deep convolutional neural network developed by Krizhevsky, Sutskever, and Hinton. It cut the error rate on the 1.2-million-image classification task by more than ten percentage points11. The ImageNet training dataset, built by Fei-Fei Li and colleagues12, had grown from 3.2 million images in 2009 to more than 14 million by 2011. Within three years, deep networks had matched and then exceeded the human baseline on the challenge13. Before AlexNet, SVMs and boosting methods dominated machine learning. After AlexNet, deep learning did.
The conceptual shift behind ImageNet is feature learning. Traditional machine learning with algorithms like SVM requires humans to hand-design the features the model uses. This is a slow, domain-specific, expert-driven process. Deep neural networks, in contrast, learn features automatically, transforming each layer’s outputs into representations that better reduce prediction error at the next layer14. This is arguably one of the most important changes with modern AI and led to the development of LLMs five years later.
The transformer architecture, introduced by Vaswani and colleagues at Google in 2017, did for natural language what AlexNet had done for vision15. Its key insight is an attention mechanism that lets a model weigh the relevance of every token in a sequence against every other token. This solved the contextual-understanding and long-range-dependency problems that limited earlier language models. It was also parallelizable in a way that recurrent architectures were not, which brings significant computational efficiencies.
The LLM family tree branched quickly with encoder-only and encoder-decoder models like T5 and UL2 for classification and translation tasks, and decoder-only models like GPT and Claude for autoregressive generation16. For SCM, these models bring a variety of applications, including generative summarization of supplier communications, code generation for internal tooling, document classification for procurement, and natural-language interfaces over enterprise data.
AI in physical automation has not enjoyed the same near-vertical acceleration curve as have applications in vision and language. Autonomous mobile robots can navigate warehouses using SLAM and computer vision and AI has led to advances in pick-and-place tasks, but physical automation faces constraints that have slowed progress compared to digital native AI applications, including capital expenditure (retrofitting warehouses for ASRS or robot fleets is enormous, with ROI horizons measured in years), limited flexibility (robots designed for specific tasks fail at novel edge cases like spilled liquids or crushed boxes), and energy density (humanoid and general-purpose robots remain bottlenecked by battery technology). Physical AI is real, but it has not yet seen its ImageNet inflection point.
What “AI” Means Now
By 2026, “AI” in supply chain management means more things than any single audience reliably agrees on. The rest of this post depends on a shared vocabulary, starting with the term itself, which is hard to pin down. One definition that synthesizes prior work is from Rai et al.17:
The ability of a machine to perform cognitive functions that we associate with human minds, such as perceiving, reasoning, learning, interacting with the environment, problem solving, decision-making, and even demonstrating creativity.
This definition does not resolve the philosophical debate about intelligence, but it is useful as a point of comparison with human cognitive functions.
One distinction often used in the field of AI is narrow versus general AI. This distinction is somewhat arbitrary because there’s no clear understanding of where the distinction can be defined, but it’s something we often encounter, so it’s worth dissecting. Narrow AI is the type of AI that we find in SCM and all other applications. It comprises specialized systems that perform specific tasks and are often better than humans at those tasks. The limitation of narrow AI is that models need to be retrained or use different algorithms to handle different tasks. General AI, or artificial general intelligence (AGI), in contrast, is the theoretical concept of a system that works across domains like humans, and presumably with superhuman performance. Foundation models are steps in this direction, but AGI is contested enough that it is best considered a research question rather than a deployment category.18
Within the broad category of narrow AI, there is a taxonomy that begins with the distinction between machine learning (ML) and symbolic AI, as we saw in the evolutionary trajectory. ML systems can be further categorized by learning type, including supervised, unsupervised, and reinforcement learning. Within each of those, there are dozens of algorithms. The visual companion to this post (here) illustrates this with diagrams that are worth looking at to ground the vocabulary visually. This taxonomy also blurs with modern AI systems like OpenAI’s ChatGPT, which often utilize all three types of learning and have rule-based capabilities and constraints.
Two additional terms that are important, especially with the rise of LLMs, are discriminative and generative.19 Discriminative systems predict a class or value given inputs. These are typical machine learning systems that perform tasks such as identifying fraudulent transactions, estimating the probability that a shipment will be late, and predicting next quarter’s demand. Generative systems, in contrast, learn the underlying distribution of the data and produce new instances. These are the LLM chatbots that have become commonplace. Generative systems can also be used to generate synthetic supply-chain data for stress testing, scenario generation for resilience planning, and augmented training data where the real data is sparse.
We encountered many of these terms across the historical evolution of OR and AI, and the landscape will continue to shift as AI systems progress. This taxonomy serves as a conceptual map of the AI systems we’ll encounter in SCM, whether for evaluating vendor claims, understanding research, or making informed decisions about AI adoption.
The Human Dimension
As algorithmic complexity and capabilities increase, the intersection of AI and human agency becomes a focal point of risk and ethics. For me, the human dimension is the most important topic in AI. Technology alone doesn’t create value. It emerges from the interaction between technology and people. Getting this interaction right is essential for successful AI implementation.
One framework that I find useful is human-centered AI. For most of the history of automation, designers treated the relationship between human control and machine automation as a one-dimensional trade-off: more automation meant less human control, and vice versa. The classic scale of automation, from manual control to full automation, lives on that one-dimensional spectrum. Beer et al.20 apply the scale to autonomy in robotics, where a robot can sense, plan, and act toward a goal without external control. The conceptual shift toward human-centered AI introduces a two-dimensional grid, with human control on one axis and automation on the other. According to Shneiderman21, the quadrant for “reliable, safe, and trustworthy” systems is the upper right — a high level of human control and automation. This means we should be building systems that are more capable machines and that simultaneously empower humans.
The distinction between human-in-the-loop and human-on-the-loop is one way to describe human control in the context of system design22. That design choice, however, depends on whether the human can understand what the machine is doing. This is the explainability problem (XAI). The importance of explanations has roots in human psychology23. A fundamental challenge with many AI systems is that they are “black boxes,” and XAI seeks to use interpretability and explainability to predict similar events in the future, diagnose, assess blame or guilt, justify or rationalize an action, and for aesthetic pleasure24. Human-Centered Explainable AI (HCXAI) extends this further, arguing that algorithmic transparency is not enough. With HCXAI, an explanation must be designed for the human in front of it, in their sociotechnical context, with attention to the cognitive biases it might exploit and the over-trust it might induce. Bad explanations give users false confidence and shift accountability away from the system.25
Around all of this sits a layer of governance. Shneiderman’s “Bridging the Gap Between Ethics and Practice” paper26 offers a useful nested framework centered on reliable systems at the engineering layer, then moving outward to a safety culture at the organizational layer, and trustworthy certification at the industry layer. Each layer has a specific job. Engineering ensures the system does what it says. Organizational culture ensures that the people building it can raise concerns without being punished. Certification ensures that third parties (auditors, regulators, customers) can verify the first two. The Volkswagen emissions scandal — which I write about here — shows what happens when the two outer layers fail.
The technical frontier in the next section only matters if it is deployable, and that depends on autonomy levels, explainability, and governance just as much as on benchmark scores.
The Convergence
The historical arc points toward a frontier that is reshaping SCM right now. Some of these technologies are already being deployed, while others remain on the research bench. Either way, both areas are worth tracking now.
Two important shifts define this frontier. The first is in the optimization layer, where machine learning is being grafted onto operations research. The second is in the perception-and-action layer, where vision, language, and physical control are being unified inside single models. I will introduce both of these areas briefly here. They are both the subject of upcoming posts in this series as well.
One of the most exciting developments in supply chain analytics is the convergence of classical optimization with modern machine learning. The Simplex Method provides mathematical rigor. It’s deterministic and explainable. But it assumes known parameters and struggles with uncertainty. Deep learning, in contrast, is adaptable. It handles uncertainty and learns from data. It’s probabilistic and flexible. But it does not provide optimality guarantees and can be difficult to interpret — the “black box”. The emerging paradigm combines both into what we might call “augmented intelligence”. These are systems that are both adaptive and optimal. Machine learning handles uncertainty and learns patterns, while mathematical optimization finds the best solution given those patterns. This convergence represents a maturation of the field toward pragmatic integration of complementary strengths.27
Four threads are particularly active in the SCM literature right now. Natural language to optimization model (NL2OPT) systems translate a planner’s prose description of a problem into a formal mathematical model that a solver can work with. This reduces the distance between articulating a problem and solving it. Neural combinatorial optimization (NCO) applies learned policies to hard OR problems like vehicle routing and scheduling; the variant with reinforcement learning, NCO-RL, has produced credible end-to-end policies that outperform hand-tuned heuristics on several problem families.28 Learning-augmented stochastic systems are bringing modern ML to queueing theory, inventory management, and Markov decision processes. Generative simulation is making it cheaper and faster to build the simulation models that supply chains have always relied on.29
The second area where AI is reshaping SCM is in physical automation. Computer vision, in both its traditional and modern forms, enables a machine to label an image. SLAM (simultaneous localization and mapping) enables a robot to build a geometric model of its environment and localize itself within it. Both have been around for years. What is new is the integration of language, vision, and motor control inside a single learned system. Vision-Language-Action (VLA) models — such as Google’s RT-2, Physical Intelligence’s π0 — take in pixels and natural-language instructions and output motor commands, learning manipulation from demonstrations rather than from hand-coded rules.30 They generalize across objects and situations in ways that the previous generation of warehouse robots cannot. The frontier beyond VLAs is spatial intelligence, a term Fei-Fei Li uses for systems that understand 3D geometry, physics, and the affordances of objects in space, not just their pixels.31 A spatially intelligent robot does not navigate to the coordinates (X, Y); it understands “pick up the fragile red box on the top shelf” and infers the motor controls to execute the request without crushing the box. World Labs’ Marble and Google DeepMind’s Genie 3 are the early public examples of multimodal world models trained to represent how objects exist, interact, and evolve in three dimensions.32
Most of these systems are still research demos rather than production systems, but the trajectory is clear, and they will likely have a huge impact on the next generation of SCM systems — the ones that will route freight amid disruption, run warehouses that handle edge cases, and build digital twins with generative priors. These systems will combine the rigor of the OR tradition with the adaptability of ML, and they will succeed or fail in deployment based on the human-dimension work in the previous section.
What to Take Away, and What Comes Next
Three things to take away from this post.
First, bridging rigor and adaptability is a defining conceptual shift with AI in SCM. Mathematical optimization and machine learning are no longer separate disciplines that happen to share a building. The next generation of systems will combine both, and the practitioners who can speak both languages will be the ones who build and interact with them successfully.
Second, the path to physical AI depends on spatial intelligence. Computer vision is well-established. VLAs and world models are the frontier. The warehouses, ports, and last-mile networks that absorb that generation of robots will look very different from what they do today.
Third, human-centered AI should not be optional. Shneiderman’s two-dimensional reframing provides a starting point for understanding how these frontier systems can be deployed responsibly. Levels of autonomy, human-in-the-loop design, explainability, and governance are essential parts of the story.
This is the first in a series of posts on MIT AI Lectures. Lecture 2: AI-Enabled Systems in SCM covers digital twins, robotics, large language models, and governance frameworks. Lecture 3: Applied AI Methods covers retrieval-augmented generation, prompt engineering, optimization applications, and more. Lecture 4: AI Safety covers interpretability and white-box analysis, alignment, cybersecurity, and more. To receive future releases directly in your inbox, subscribe to the newsletter. It’s free. If you find it valuable, consider passing it along to someone else who might.
Lepenioti, K., Bousdekis, A., Apostolou, D., & Mentzas, G. (2020). Prescriptive analytics: Literature review and research challenges. International Journal of Information Management, 50, 57–70. https://doi.org/10.1016/j.ijinfomgt.2019.04.003
Mast, J. de, Steiner, S. H., Nuijten, W. P. M., & Kapitan, D. (2023). Analytical Problem Solving Based on Causal, Correlational and Deductive Models. The American Statistician, 77(1), 51–61. https://doi.org/10.1080/00031305.2021.2023633
Mast, J. de, Steiner, S. H., Nuijten, W. P. M., & Kapitan, D. (2023). Analytical Problem Solving Based on Causal, Correlational and Deductive Models. The American Statistician, 77(1), 51–61. https://doi.org/10.1080/00031305.2021.2023633
Dantzig, G. (1987). Origins of the Simplex Method. Stanford University, Technical ReportSOL 87-5. https://apps.dtic.mil/sti/tr/pdf/ADA182708.pdf
Dantzig, G. (1987). Origins of the Simplex Method. Stanford University, Technical ReportSOL 87-5. https://apps.dtic.mil/sti/tr/pdf/ADA182708.pdf
Jackson, A. (2023). Why the Simplex Method, at Age 75, is Still the Go-To Algorithm. Communications of the ACM. https://cacm.acm.org/news/why-the-simplex-method-at-age-75-is-still-the-go-to-algorithm/
Fradkov, A. L. (2020). Early History of Machine Learning. IFAC-PapersOnLine, 53(2), 1385–1390. https://doi.org/10.1016/j.ifacol.2020.12.1888
Oravec, J. A. (2014). Expert Systems and Knowledge-Based Engineering (1984-1991): Implications for Instructional Systems Research. International Journal of Designs for Learning, 5(2). https://doi.org/10.14434/ijdl.v5i2.12891
Sullivan, G., & Fordyce, K. (1990). IBM Burlington’s Logistics Management System. Interfaces, 20(1), 43–64. https://doi.org/10.1287/inte.20.1.43
Gill, T. G. (1995). Early Expert Systems: Where Are They Now? MIS Quarterly, 19(1), 51–81. https://doi.org/10.2307/249711
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems 25 (NIPS 2012). https://papers.nips.cc/paper_files/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009). ImageNet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–255. https://doi.org/10.1109/cvpr.2009.5206848
He, K., Zhang, X., Ren, S., & Sun, J. (2015). Deep Residual Learning for Image Recognition. arXiv. https://doi.org/10.48550/arxiv.1512.03385
Bengio, Y., Courville, A., & Vincent, P. (2012). Representation Learning: A Review and New Perspectives. arXiv. https://doi.org/10.48550/arxiv.1206.5538
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention Is All You Need. arXiv. https://doi.org/10.48550/arxiv.1706.03762
Yang, J., Jin, H., Tang, R., Han, X., Feng, Q., Jiang, H., Yin, B., & Hu, X. (2023). Harnessing the Power of LLMs in Practice: A Survey on ChatGPT and Beyond. arXiv. https://doi.org/10.48550/arxiv.2304.13712
Rai, A., Constantinides, P., & Sarker, S. (2019). Editor’s Comments: Next-Generation Digital Platforms: Toward Human–AI Hybrids. MIS Quarterly, 43. https://dl.acm.org/doi/10.5555/3370135.3370136
Russell, S., & Norvig, P. (2021). Artificial Intelligence: A Modern Approach (4th ed.). Pearson. https://www.pearson.com/en-us/subject-catalog/p/artificial-intelligence-a-modern-approach/P200000003500/9780137505135
Adapted from Dr. Ian McCulloh at Johns Hopkins University.
Beer, J. M., Fisk, A. D., & Rogers, W. A. (2014). Toward a framework for levels of robot autonomy in human-robot interaction. Journal of Human-Robot Interaction, 3(2), 74–99. https://doi.org/10.5898/jhri.3.2.beer
Shneiderman, B. (2020b). Human-Centered Artificial Intelligence: Reliable, Safe & Trustworthy. International Journal of Human–Computer Interaction, 36(6), 495–504. https://doi.org/10.1080/10447318.2020.1741118
Cheruvu, R. (2025). Handbook of Human-Centered Artificial Intelligence. 1–47. https://doi.org/10.1007/978-981-97-8440-0_75-1
Keil, F. C. (2006). Explanation and Understanding. Annual Review of Psychology, 57(1), 227–254. https://doi.org/10.1146/annurev.psych.57.102904.190100
Ridley, M. (2025). Human‐centered explainable artificial intelligence: An Annual Review of Information Science and Technology (ARIST) paper. Journal of the Association for Information Science and Technology, 76(1), 98–120. https://doi.org/10.1002/asi.24889
Ridley, M. (2025). Human‐centered explainable artificial intelligence: An Annual Review of Information Science and Technology (ARIST) paper. Journal of the Association for Information Science and Technology, 76(1), 98–120. https://doi.org/10.1002/asi.24889
Shneiderman, B. (2020a). Bridging the Gap Between Ethics and Practice. ACM Transactions on Interactive Intelligent Systems (TiiS), 10(4), 1–31. https://doi.org/10.1145/3419764
Bengio, Y., Lodi, A., & Prouvost, A. (2021). Machine learning for combinatorial optimization: A methodological tour d’horizon. European Journal of Operational Research, 290(2), 405–421. https://doi.org/10.1016/j.ejor.2020.07.063
Chung, K. T., Lee, C. K. M., & Tsang, Y. P. (2025). Neural combinatorial optimization with reinforcement learning in industrial engineering: a survey. Artificial Intelligence Review, 58(5), 130. https://doi.org/10.1007/s10462-024-11045-1
Boone, T., Fahimnia, B., Ganeshan, R., Herold, D. M., & Sanders, N. R. (2025). Generative AI: Opportunities, challenges, and research directions for supply chain resilience. Transportation Research Part E: Logistics and Transportation Review, 199, 104135. https://doi.org/10.1016/j.tre.2025.104135
Zhou, Q., & Sheu, J.-B. (2025). The use of Generative Artificial Intelligence (GenAI) in operations research: review and future research agenda. Journal of the Operational Research Society, ahead-of-print(ahead-of-print), 1–21. https://doi.org/10.1080/01605682.2025.2561762
Sapkota, R., Cao, Y., Roumeliotis, K. I., & Karkee, M. (2025). Vision-Language-Action Models: Concepts, Progress, Applications and Challenges. arXiv. https://doi.org/10.48550/arxiv.2505.04769
