AAAI-24 Spring Symposium on Human-Like Learning
- Important Event Information
- Symposium Schedule
- Motivation and Background
- Themes of the Symposium
- Symposium Organizers
- References
Important Event Information
| Key Dates All deadlines are 11:59pm UTC-12:00 (anywhere on Earth) |
|
|---|---|
| Abstracts Due | January 10, 2024 |
| AAAI Symposium Registration Opens | January 10th, 2024 |
| Selected Presenters Notified | January 15, 2024 |
| Finalized Abstracts Due from Authors | January, 26, 2024 |
| AAAI Proceedings Camera Ready Copy Deadline | Febuary 15th, 2024 |
| Deadline for Symposium Registration | Febuary 29th, 2024 |
| Finalized Schedule Released | March 2nd, 2024 |
| Symposium Takes Place at Stanford University | March 25-27, 2024 |
Event Location
- All Symposia Rooms will be in the Lane History Corner/Bldg. 200 – located at 450 Jane Stanford Way Bldg 200, Stanford, CA 94305.
- Our symposium will be located in Room 105.
- The poster session will take place just outside of the building.
- Reception – Tresidder Oak Lounge, second floor of the Tresidder Student Union Building
- Plenary Session – Bishop Auditorium
Speaker and Poster Presenter Details
- Each speaker will each be alloted 30 minutes, with 20 minutes for their presentation and 10 minutes for questions.
- For the poster presenters, AAAI will provide foam poster boards and easels. Poster boards are 30”x40” and can be either portrait or landscape orientation.
Parking Information
- General information: https://transportation.stanford.edu/how-purchase-visitor-parking
- “How to” guide for visitor zone parking : https://transportation.stanford.edu/how-guide-visitor-zone-parking
- Link to the map with parking zone numbers. Lots 7282 and 7213 are the closest for Symposia sessions. https://drive.google.com/file/d/1VCyTJOvNkDE43PpLI-cGtR30hylolWuo/view
Registration Details
You can register at the AAAI 2024 Spring Symposium page.
Submission Details
- We invite submissions of abstracts. We have no hard length requirements, but recommend a short (minimum one paragraph) or extended format (up to 2 pages).
- Abstracts can be brief, but should provide sufficient detail for the organizing committee to determine the focus and substance of your proposed talk for planning purposes. Also, we plan to link to the titles/abstracts in the schedule and symposia proceedings.
- Abstracts should be formatted using the AAAI Author Kit (https://aaai.org/authorkit24-2/).
- Submit abstracts via EasyChair to the “Symposium on Human-Like Learning”.
- Participants will have the opportunity to finalize their abstract (by January 26th, see table above) before it appears on the website or in the symposia proceedings.
- Submissions not selected for talks will be invited to participate in a poster session.
- If there is interest at the symposia, we will explore a possible journal special issue or edited volume based on submissions.
- SSS-24 accepted authors will have the opportunity to publish their work with AAAI Press in the Proceedings of the AAAI Symposium Series. The AAAI CRC will be collected via a separate platform. Authors should NOT upload the AAAI CRC to Easy Chair. Authors with accepted papers should expect to receive an email invitation to submit their files to the AAAI CRC site from proceedings-questions@aaai.org. If you do not receive AAAI CRC submission instructions within two weeks of the CRC deadline, please check your spam inbox and contact proceedings-questions@aaai.org. If authors are interested in having work published by AAAI Press, note that the proceedings will NOT be published before the program. Note: Participating in the Proceedings of the AAAI Symposium Series publication is optional. If you do not want your paper published, please ignore e-mails from proceedings-questions@aaai.org. If you do not submit anything by the AAAI CRC deadline, then we will understand that you are opting out of the proceedings.
Symposium Schedule
The event will occur at the AAAI Spring Symposium at Stanford from March 25th to March 27th. The schedule below shows the talk titles / activities and the respective authors. Note, all times are in the timezone at the symposium’s location (Stanford, CA / PST).
For talks, each presentation will be allotted 30 minutes, with 20 minutes for the presentation and 10 minutes for discussion. The talks are roughly organized into thematic groups that will be presented together, with the intention of highlighting commonalities across research efforts and paradigms. The first session provides a general framing for human-like learning research and follow talks explore human-like learning as it relates to representation, decision making, analogy, concept learning, language (and LLMs), evaluation, and task learning.
On the first day, there will also be a poster session. For poster presenters, AAAI will provide foam poster boards and easels. Poster boards are 30”x40” and can be either portrait or landscape orientation. The first day will conclude with the AAAI symposium reception.
On the second and third day, we have dedicated time for four group discussion sessions to explore the following questions:
- What are the key characteristics of human learning we should aim to reproduce, and how might different approaches reproduce these characteristics?
- How should we evaluate human-like learning systems?
- What are the potential broader impacts for human-like learning research?
The symposium will conclude with a discussion of the steps we should take next to advance this research program. E.g., we might discuss follow on events or an edited book or journal special topic issue.
Day 1 Schedule (3/25)
Day 2 Schedule (3/26)
| Start | End | Title / Activity | Authors |
|---|---|---|---|
| 9:00 | 9:30 | Analogy as the Swiss Army Knife of Human-like Learning | Kenneth D. Forbus |
| 9:30 | 10:00 | Human-Like Learning of Social Reasoning via Analogy | Irina Rabkina |
| 10:00 | 10:30 | Pushing the Limits of Learning from Limited Data | Maya Malaviya, Ilia Sucholutsky and Thomas L. Griffiths |
| 10:30 | 11:00 | Break | |
| 11:00 | 11:30 | Avoiding Catastrophic Forgetting in Visual Classification using Human-Like Concept Formation | Nicki Barari, Xin Lian and Christopher J. MacLellan |
| 11:30 | 12:00 | Human-like Learning in Temporally Structured Environments | Matt Jones, Tyler Scott and Michael Mozer |
| 12:00 | 12:30 | Discussion 1: What are the key characteristics of human learning we should aim to reproduce, and how might different approaches reproduce these characteristics? | |
| 12:30 | 2:00 | Lunch | |
| 2:00 | 2:30 | Constructing Deep Concepts through Shallow Search | Bonan Zhao, Christopher G. Lucas, and Neil R. Bramley |
| 2:30 | 3:00 | Toward Autonomy: Metacognitive Learning for Enhanced AI Performance | Brendan Conway-Smith and Robert West |
| 3:00 | 3:30 | Do Large Language Models Learn to Human-Like Learn? | Jesse Roberts |
| 3:30 | 4:00 | Break | |
| 4:00 | 4:30 | Turtle-like Geometry Learning: How Humans and Machines Differ in Learning Turtle Geometry | Sina Rismanchian, Shayan Doroudi and Yasaman Razeghi |
| 4:30 | 5:00 | Turing-like Experiment in a Cyber Defense Game | Yinuo Du, Baptiste Prebot and Cleotilde Gonzalez |
| 5:00 | 5:30 | Discussion 2: How should we evaluate human-like learning systems? | |
| 6:00 | 7:00 | AAAI Plenary Session |
Day 3 Schedule (3/27)
| Start | End | Title / Activity | Authors |
|---|---|---|---|
| 9:00 | 9:30 | Decomposed Inductive Procedure Learning: Learning Academic Tasks with Human-Like Data Efficiency | Daniel Weitekamp |
| 9:30 | 10:00 | Modeling Human-Like Acquisition of Language and Concepts | Peter Lindes and Steven Jones |
| 10:00 | 10:30 | Rapid Acquisition of Hierarchical Plan Knowledge | Pat Langley |
| 10:30 | 11:00 | Break | |
| 11:00 | 12:30 | Discussion 3: What are the potential broader impacts of human-like learning research? | |
| 11:00 | 12:30 | Final Discussion: What steps should we take after the symposium to advance this reserach program? |
Motivation and Background
Recent machine-learning research has made incredible progress across a wide range of tasks. While many systems can achieve human-like performance, one area that is currently under explored is how to realize human-like learning capabilities within these systems. For example, machine learning typically employs batch training and requires more data and computation than people to achieve similar capabilities. The resulting models are effective, but difficult to update in the face of new data without costly retraining. In contrast, humans excel at rapidly assimilating new information on the fly from a limited number of examples. More research is needed to investigate human-like capabilities, such as efficient, incremental learning, and to explore the design of artificial systems that can also exhibit them.
Historically, the fields of Artificial Intelligence and Cognitive Science have maintained close theoretical ties. As a result, early machine-learning research placed a strong emphasis on human-like learning. For example, early research with systems like ACT (Anderson, 1982; 1983) and EPAM (Feigenbaum, 1959; Feigenbaum & Simon, 1984) aimed to both improve our understanding of human learning and to reproduce human-like learning capabilities within artificial systems. Other early research with Soar (Laird, Newell, & Rosenbloom, 1987) composed many human-like capabilities to propose a unified account of human cognition (both problem-solving and learning). Research within the Parallel Distributed Processing group (Rumelhart, McClelland, and the PDP Group, 1986) explored how parallel processing and distributed representations—often within neural networks—could account for human cognition and learning. Despite the success of these early efforts (many of which led directly to today’s dominant paradigms), the connection between Artificial Intelligence and Cognitive Science has weakened over the past decades; there has been an increasing emphasis on functional applications and the development of artificial systems that process data in ways and at scales that exceed human processing capacities. A symposium on human-like learning is timely because it can re-strengthen this theoretical connection and enable researchers to revisit this topic in light of recent computational advancements. A renewed focus on human learning should prove valuable, both in enhancing our comprehension of human capabilities and in enabling new advancements in applications, such as more efficient machine-learning systems.
Themes of the Symposium
At a high-level, the symposium aims to encourage discussion on the following themes:
- Identification of key characteristics of human-like learning to target in AI/ML research, along with what makes these characteristics challenging for current approaches;
- Ongoing and proposed research into how to create artificial systems that exhibit human-like learning;
- Approaches for evaluating human-like learning systems; and
- Exploration of the broader context for and impacts of human-like learning systems, such as how they might complement/benefit current machine learning systems and humanity.
Characteristics of Human Learning
There are several characteristics of human learning that could serve as targets for a general research program. Langley (2022) highlights several that could serve as anchors to motivate the symposium discussions:
- Expertise is acquired in a piecemeal manner, with one element being added at a time.
- Learning is an incremental activity that processes one experience at a time.
- Learning is guided by prior experience, which aids in the interpretation of new experiences.
- Expertise is acquired and refined rapidly, from few training cases.
This list is not exhaustive and only serves to provide initial motivation. We will encourage participants to propose, highlight, and discuss features of human learning that are central to their work and could inspire future computational work. A major output of the symposium will be a list of human-like capabilities that could be the focus for a sustained research program. We also hope that by encouraging researchers to emphasize and discuss human-like capabilities, rather than emphasizing their particular methodological approach, we can foster more effective communication between people that come from different disciplines (e.g., Psychology or AI) and methodological perspectives (e.g., Cognitive Architectures and Deep Learning) backgrounds.
Creating Human-Like Learning Systems
Given these characteristics, the major focus of the symposium will be to explore ongoing and proposed research into how to create artificial systems that exhibit human-like learning capabilities. Many different research efforts—spanning multiple methodological paradigms—have emphasized some aspects of human-like learning:
- Cognitive architectures: Researchers in this area have identified and operationalized several human-like learning assumptions, such as the assumption that “learning occurs online and incrementally” (Laird, Lebiere, & Rosenbloom, 2017).
- Interactive task learning: This paradigm emphasizes the incremental, multimodal acquisition of new task knowledge from a limited amount of interactive instruction provided by human teachers (Laird et al., 2017); early examples include systems like AnimNL (Webber, Badler, Di Eugenio, Geib, Levison, & Moore, 1995) and more recent examples include AILEEN (Mohan, Klenk, Shreve, Evans, Ang, & Maxwell, 2020) and ONYX (Ruoff, Myers, & Maedche, 2023).
- Extended/continual learning: A class of approaches that emphasizes learning of different types of knowledge or functions over extended durations (e.g., years); a notable example is NELL (Mitchell et al., 2018), which continually crawls the web and grows and refines its relational ontology based on what it learns.
- Probabilistic programming: This paradigm emphasizes the use of Bayesian priors (over both internal structures and parameters) to guide and constrain learning, which enables the acquisition of new concepts with few examples (Lake, Salakhutdinov, & Tenenbaum, 2015).
- Concept formation: Systems like Cobweb (Fisher & Langley, 1990; Fisher & Yoo, 1993; MacLellan, Matsakis, & Langley, 2022) emphasize the incremental, piecemeal acquisition of concepts and rapid, few-example learning.
- Analogical and case-based learning: The SAGE (McLure, Friedman, & Forbus, 2015) system leverages analogical processing to support efficient, few example learning over structured representations.
- Logic-based learning: Recent work on this topic emphasizes the induction of modular knowledge structures that can be composed at performance and guide subsequent learning (Cropper, Dumančić, & Muggleton, 2020).
- Simulated students: Efforts to simulate human novices has highlighted the importance of prior knowledge and self-explanation in guiding subsequent learning (VanLehn, Jones, & Chi, 1991) and emphasized comparison of human and machine behavior—both correct and incorrect—as a tool to better understand and simulate human learning (MacLellan, Harpstead, Patel, Koedinger, 2016; Weitekamp, Ye, Rachatasumrit, Harpstead, & Koedinger, 2020).
- Human-like neural network approaches: Although most current neural network approaches do not emphasize human-like learning, there are some efforts that do. One recent example investigates how to model humans’ ability to learn increasingly complex skills across a lifespan without major instances of catastrophic interference or forgetting (Lyndgaard, Tidler, Provine, & Varma, 2022).
These are just a few examples of research lines that emphasize human-like learning. The symposium aims to bring together individual researchers working across different paradigms who may feel isolated in their commitment to human-like learning, and to create a community that can foster collaboration and inspire innovation on this topic. Many of these efforts share similar goals, but there has only been limited discussion about common insights for creating human-like learning systems that span across paradigms. The symposium will encourage discussion on this topic by having speakers representing different paradigms present in the same session and facilitating a broader group discussion about cross-paradigm commonalities.
Evaluating Human-Like Learning Systems
Research on this topic will need some way to measure progress. Thus, another major aim of the symposium will be to discuss how to evaluate systems with human-like capabilities. Here are a few potential evaluations participants might discuss:
- assessing systems based on their learning efficiency, rather than final performance;
- measuring fits to human behavioral data, both quantitative and qualitative;
- establishing a set of target capabilities, with qualitative measures of attainment; and
- evaluating human-like learning systems in terms of ease of use and acceptance by human users who are interacting with them.
These are just a selection of several potential evaluation ideas and a key deliverable of the symposium will be to solicit participants ideas about evaluations and to encourage formation of a community consensus around appropriate evaluations.
Broader Impacts of Human-Like Learning Systems
Lastly, we envision that research into human-like learning will have substantial broader impacts across several research communities. For example, it could lead to improvements in the efficiency and sustainability of machine-learning models, and enable them to be more adaptable and flexible. It could also contribute to a deeper understanding of human learning, and lead to the reproduction and modeling of human learning capabilities. Finally, it could lead to more user-friendly AI systems with more relatable, human-like capabilities. A final aim of the symposium will be to reflect on and discuss the broader implications for research into human-like learning.
Symposium Organizers
- Christopher J. MacLellan, Teachable AI Lab, Georgia Institute of Technology, cmaclell@gatech.edu
- Ute Schmid, University of Bamberg, ute.schmid@uni-bamberg.de
- Douglas Fisher, Vanderbilt University, douglas.h.fisher@vanderbilt.edu
- Randolph M. Jones, Soar Technology, LLC, rjones@soartech.com
References
- Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological review, 89(4), 369.
- Anderson, J. R. (1983). A spreading activation theory of memory. Journal of verbal learning and verbal behavior, 22(3), 261-295.
- Cropper, A., Dumančić, S., & Muggleton, S. H. (2020). Turning 30: New ideas in inductive logic programming. arXiv preprint arXiv:2002.11002.
- Feigenbaum, E.A. (1959) An Information Processing Theory of Verbal Learning. RAND Report P-1817. Santa Monica, CA: RAND Corporation.
- Feigenbaum, E. A., & Simon, H. A. (1984). EPAM-like models of recognition and learning. Cognitive Science, 8, 305–336.
- Fisher, D., & Langley, P. (1990). “The Structure and Formation of Natural Categories,” in G. Bower (ed.), The Psychology of Learning and Motivation, 26, San Diego, CA: Academic Press, 241–284.
- Fisher, D., & Yoo, J. (1993). “Problem solving, categorization, and concept learning: A unifying view,” in G. Nakamura, R. Taraban, & D. Medin (eds.), The Psychology of Learning and Motivation, 29, San Diego, CA: Academic Press, 219–255.
- Laird, J. E., Gluck, K., Anderson, J., Forbus, K. D., Jenkins, O. C., Lebiere, C., … & Kirk, J. R. (2017). Interactive task learning. IEEE Intelligent Systems, 32(4), 6-21.
- Laird, J. E., Lebiere, C. & Rosenbloom, P. S. (2017). A Standard Model for the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics, AI Magazine 38(4). https://doi.org/10.1609/aimag.v38i4.2744
- Laird, J. E., Newell, A., & Rosenbloom, P. S. (1987). Soar: An architecture for general intelligence. Artificial intelligence, 33(1), 1-64.
- Lake, B. M., Salakhutdinov, R., & Tenenbaum, J. B. (2015). Human-level concept learning through probabilistic program induction. Science, 350(6266), 1332-1338.
- Langley, P. (2022). The computational gauntlet of human-like learning. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 36, No. 11, pp. 12268-12273).
- Lyndgaard, S., Tidler, Z. R., Provine, L., & Varma, S. (2022). Catastrophic interference in neural network models is mitigated when the training data reflect a power-law environmental structure. In Proceedings of the Annual Meeting of the Cognitive Science Society (Vol. 44, No. 44).
- MacLellan, C.J., Harpstead, E., Patel, R., Koedinger, K.R. (2016). The Apprentice Learner Architecture: Closing the loop between learning theory and educational data. In Proceedings of the 9th International Conference on Educational Data Mining. Raleigh, NC: International Educational Data Mining Society.
- MacLellan, C.J., Matsakis, P., & Langley, P. (2022). Efficient Induction of Language Models via Probabilistic Concept Formation. In Proceedings of the Tenth Annual Conference on Advances in Cognitive Systems.
- McLure, M. D., Friedman, S. E., and Forbus, K. D. (2015). Extending analogical generalization with near-misses. In Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, 565–571. Austin, TX: AAAI Press.
- Mitchell, T., Cohen, W., Hruschka, E., Talukdar, P., Yang, B., Betteridge, J., … & Welling, J. (2018). Never-ending learning. Communications of the ACM, 61(5), 103-115.
- Mohan, S., Klenk, M., Shreve, M., Evans, K., Ang, A., & Maxwell, J. (2020). Characterizing an Analogical Concept Memory for Architectures Implementing the Common Model of Cognition. In Proceedings of the Annual Conference on Advances in Cognitive Systems 2020.
- Rumelhart, D.E., J.L. McClelland and the PDP Research Group (1986). Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Volume 1: Foundations, Cambridge, Massachusetts: MIT Press, ISBN 978-0262680530
- Ruoff, M., Myers, B. A., & Maedche, A. (2023). ONYX: Assisting Users in Teaching Natural Language Interfaces Through Multi-Modal Interactive Task Learning. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (pp. 1-16).
- VanLehn, K., Jones, R. M., & Chi, M. T. H. (1991). Modeling the self-explanation effect with Cascade 3. In Proceedings of the human factors in computing systems conference (pp. 132–137).
- Webber, B., Badler, N., Di Eugenio, B., Geib, C., Levison, L., & Moore, M. (1995). Instructions, intentions and expectations. Artificial Intelligence, 73(1-2), 253-269.
- Weitekamp, D., Ye, Z., Rachatasumrit, N., Harpstead, E., & Koedinger, K. (2020). Investigating differential error types between human and simulated learners. In Proceedings of the 21st International Conference on Artificial Intelligence in Education, 586-597. Ifrane, Morocco: Springer International Publishing.