Evolution of the lab's research program: A five-year reflection

On June 30, 2026, I will complete 5 years of running a human memory research lab. Here is a summary of our research efforts so far and vision for the upcoming 5 years.

Overview of our research program

Our main area of focus is human memory. However, we are investigating memory at multiple levels: 1) Memory at the stimulus level (averaged across many people)—memorability 2) Memory at the neurocognitive mechanistic level—temporal context and event segmentation 3) Neural propagating waves as the basis for multiscale neuronal communication (and therefore, cognition) in the brain.

1) Memorability

We have developed a body of work describing what makes experiences memorable. For instance, the word "PINEAPPLE" may be more memorable to people in general, regardless of their individual memory capacities, than a word like "FOOT." The key insight that has come out of our work is that memorability is not a single fixed concept, unlike what is often claimed or implied in the literature. Memorability is highly context-dependent. So PINEAPPLE is only memorable to a lot of people if you test them in a specific way. Yes, there are some aspects of face memorability, for example, that do exhibit a lot of stability across ages and other differences in populations, but again, that is only true when you test memory in a specific way (recognition memory). We have now described such context sensitivity by analyzing word memory datasets that have used different memory test formats [1,2] and by comparing the memorability of the SAME words (translated to Hindi) in a Hindi-speaking population [3]. Even the same concepts do not retain their memorability patterns when tested with the same memory task in a different language and population [3]. Taken together, our work suggests that there is no "intrinsic" memorability quality that can be attached to stimuli, as is often claimed or implied in the literature. One must always qualify it with various contextual factors such as the memory task and population being tested.

We are extending this work beyond isolated words to richer linguistic units such as sentences [4] and paragraphs. To this end, we have published work assessing paragraph-level measures of coherence and "narrative flow" between sentences and expect that such measures will be useful in determining whether a paragraph is memorable [5,6]. 

Through a collaboration with a computer vision scientist (who also bears the brunt of my venting sometimes about various faculty life matters), we have also investigated what spatiotemporal features humans and models pay attention to when memorizing videos for later recognition memory tasks [7]. Together, this body of work paints a clear picture of the memorability of experiences being sensitive to culture, language, and memory test formats. So papers that are being written currently about how to make ads memorable using insights from recognition-memorability models need to pause and think: Do I want people to remember the brand name when cued by the concept, or do I want them to just recognize an ad if we display it to them again? For example, when someone thinks of noodles (a memory cue), do I want them to think of Maggi (a memory target), or do I just want them to be able to recognize the ad if they see it again and say, "Yes, I've seen this ad before"? Right now, it is the latter that people are building towards, which is a bit misguided, IMO. 

In addition to using open datasets, we have run experiments and collected data from over 800 participants for this work on memorability. The memorability literature makes some leaps of faith about applications as mentioned above. In such cases, moving away from academic research to make efforts to translate findings to real-world use cases will reveal the weaknesses of research settings. Our work has also led to some interesting discussions with an industry partner on how to translate some of these insights to real-world experiences. I will have more to say on that in maybe a year from now. 

While this memorability work was something I initiated to keep ourselves busy and productive while I tried to build infrastructure to do some ambitious cognitive neuroscience work, it has turned out to be a fulfilling line of work. Finally, in doing this work, we faced challenges doing analogous analyses in Hindi when comparing English and Hindi memorability patterns. Specifically, we needed to use psycholinguistic features to build models to explain these memorability patterns. The dearth of good resources for Hindi has led us to enter into a few international collaborations with colleagues in the US and Australia to build human free association databases for Hindi (and soon, hopefully, other low-resource Indian languages). 

While we will continue our memorability work with paragraphs and more real-world experiences with our industry partners, the next phase in our lab will be focused on the cognitive neuroscience aspects described below since we have now managed to get funds to procure EEG equipment. Getting a permanent space for it will take time, but we should be able to start our EEG studies from our temp space soon.

2) Temporal context, event segmentation, and memory

We experience events that are ever-changing. No moment is exactly the same as the previous one due to changes in environmental input, internal states, and both random and non-random processes in neural systems in general. Estes' stimulus sampling theory [8] provided an early mathematical account of how such changing environmental states affect context and, in turn, memory. Context, btw, is any set of extra features (e.g., those of the office) that are associated with entities we focus on (e.g., this blog post as I write it, sitting in my office) such that later, when cued by context, we remember the associated entities. For example, if you celebrated your birthday last year at a pub a bit away from your home, the next time you pass by, seeing the pub may remind you of the events of the birthday. Temporal context is a dynamic version of context, and a formal mathematical model [9] proposed that it is made up of features of not only the currently experienced event but also those of recently experienced ones. This temporal context evolves with time. Therefore, there is a history to this context, and it is to this dynamic context that we associate ongoing experiences. Critically, later when remembering an entity, we also retrieve the state of the dynamic temporal context that was associated with that entity earlier in our experience. This framework provides elegant explanations of various memory phenomena such as the temporal contiguity effect: when you remember an event, you are also likely to remember other events that you experienced close to that event in time. 

However, such memory phenomena can be explained by other models as well. So what exactly is the mechanism underlying how we organize memories in time? Even though the temporal context model [9] is an elegant model, neural evidence for the specific mechanisms proposed by it comes primarily from free recall studies [10], which have certain confounds in them [11]. To really know whether the mechanisms proposed in the model are actually how we encode and retrieve memories, we need to obtain direct neural evidence for those mechanisms because behavioral data by themselves have not been able to tease apart different explanations for the same phenomena. We are currently analyzing neuroimaging data to test how contextual representations are reinstated during memory. In parallel, we are also designing EEG experiments to get time-resolved measures of neural temporal context representations. 

Temporal context is not just a smooth changing neural state. It is often disrupted at what are known as event boundaries. These are change points between events, and tracking these change points has been found to be important for various downstream memory tasks [12,13,14]. Research on event segmentation, however, has been somewhat unduly influenced by early experiments that were specifically designed for enabling fMRI analysis (e.g., the Ezzyat-duBrow-Davachi (EDD) paradigm [15]). These are simplistic experiments that show a bunch of faces and then a bunch of objects and alternate between them at predictable intervals. The switch is called the event boundary. We recently proposed that the results from those experiments are unlikely to generalize to more naturalistic settings due to additional psychological processes recruited in realistic experiences [16]. Our predictions were borne out in recent studies by a (the?) leading event cognition lab that found that patterns of temporal order memory found in the EDD paradigm reverse under some circumstances when you use naturalistic tasks where people have semantic knowledge about the events [17,18]. 

A major direction in the lab is to study event segmentation using more naturalistic and ecologically valid materials. While many other labs around the world are also working on the same problem, we believe that we have unique insights to provide because we have the advantage of being newcomers to the field. Some of our initial publications in this domain [19] have resulted from minor tweaks to experiments done by other labs to address confounds that were apparent to us, perhaps with our fresh/less biased eyes, and we have indeed found reliable differences (including meaningful null effects) in the pattern of results from the original studies. 

We will also start some new funded work on how richer real-world experiences shape event segmentation and memory. However, we are awaiting the sanction letter, and it has been two years since I proposed the ideas. We are getting slowed down due to such delays in fund disbursals. Studies involving human participants and EEG need significant funding. So we do what we can to stay productive and contribute what we can to this hot area of research as we await the approved funds. 

We have also initiated another exciting and novel line of work with a good friend and collaborator at IIT Hyderabad to use neuromodulation to complement our efforts under this theme. Yet another related line of work, again with my trusty computer vision colleague, is to understand how insights from human event cognition can help improve long-video understanding tasks in computer vision. 

So, determining how temporal context and breaks in temporal context work together to influence memory encoding and retrieval will advance our understanding of episodic memory, and we want to do it in a way that respects the dynamics of real-world experiences so that our findings are generalizable to contexts that matter. 

3) Propagating neural waves and cognition

During my postdoctoral research at the NIH, I built on methods based on the physics of waves to detect propagating waves of neural activity in the human cortex during a memory task. The main finding was that waves at different spatial scales (macroscopic: ~8 cm and mesoscopic: ~4 mm) tended to co-occur and that mesoscopic waves increase the likelihood of single-unit firing [20]. While the initial paper was preprinted in 2020, it got reviewed at a few glam journals and was rejected, as we did not find any relationship between wave propagation and memory performance in the task. One rejection even said that it was a "tour de force" analysis. However, when working with data collected from epilepsy patients, one concern that typically comes up is whether the phenomenon is epilepsy-specific. So it is often quite useful to demonstrate that neural phenomena are related to task performance, as that demonstrates two things: 1) The phenomenon is relevant for cognition and therefore interesting, and 2) the phenomenon is not an artifact or characteristic of epilepsy that is unlikely to generalize to the rest of the population. 

In the last five years, I have been so busy with setting up my own lab, teaching, other admin duties, etc. that this paper fell a bit by the wayside. I agreed to a lot of service requests at the institute (many without official credit, as I was not put on committees initially). This is entirely my fault because I also volunteered to fix several small issues I saw in the system. This is NOT advisable for starting faculty. In hindsight, I should have protected my time a bit more. Finishing your prior papers on time is really important. So I grabbed an opportunity to resume analyses of that dataset when I was recently invited to give a talk at the Wave Club, an online seminar series organized by some of the top researchers in the domain of traveling waves. To prepare for this talk, I had to carve out some time for analysis and cancelled some student meetings, consolidated them to a single day, and shut myself in my office with a DND sign for a few months. These recent reanalyses of this dataset have strengthened our confidence that propagating waves are behaviorally meaningful for memory. My immediate priority is to write up these newer results and submit that paper to a journal within a few months. This is probably one of the most important papers I will ever write. So, I have requested a teaching break for the upcoming semester without which it would be really hard to pull this off. 

How neurons communicate both within and across brain regions is a fundamental problem in neuroscience. Coordination of activity with time delays on the order of spike time-dependent plasticity (STDP) timescales is a promising candidate mechanism for brain-wide communication. See this recent Science paper, for example, about spiraling propagating waves. However, a lot of work remains to be done to elucidate the neuronal processes that are responsible for generating waves, reversing their directions or other wave features in a task-sensitive manner [21], coordinating waves at different scales, coordinating spiking activity, etc. To that end, over the next several years, the lab will develop toolboxes for robust measurement of waves in neural recordings, build neuronal models that can help us understand multiscale coordination behaviors, and use these models to make predictions that are testable using high-density EEG. A collaborator (who is an expert on neuronal modeling) and I have submitted a grant for this, and I am optimistic that it will be reviewed and scored well sooner or later. Once this line of work takes off, I would also like to build neuromodulation setups to further test the functional role of propagating waves in cognition. This really is a 5-10 year roadmap, but our initial efforts are well underway with a MATLAB toolbox that is almost fully ready for the detection of not just propagating planar waves but also spiral and other types of complex wave patterns. Before the grants come in, we will be able to make some progress with some open datasets as well. 

Miscellaneous projects and some use-inspired research problems

In addition to the main area of focus above, we also have a few curiosity-driven projects such as one on an ethological framework for understanding emotion perception and others on how to improve learning and memory with tweaks to established spaced retrieval practice and other methods. The ethology work, in particular, was close to my heart, as it was a project I had started as a graduate student with Dr. David Huron. With David's unfortunate passing in 2025, it was a personal goal to finish the project and publish it in his memory. Thanks to some brilliant students in the lab, we pulled off a 4-5 part preregistered study and collected data from ~650 participants across these studies. The simple question we asked was "Why is jealousy not easy to perceive in affective displays like music?" While many theories of emotion offer partial answers, we turned to the ethological theory of animal signaling to offer a different, complementary view. It is notoriously hard to falsify evolutionary theories, but we made a good attempt, IMO. This work will be submitted to a journal within a month or so. To be clear, we do not offer this as a complete explanation of emotion detectability. However, if supported, the ethological framework would offer one novel pathway (among many other sociocultural mechanisms) by which we come to express and understand emotions today in our communicative behaviors. 

We also have a keen interest in how we can learn effectively. After all, better learning leads to better memory, but a key question is what kind of memory? A lot of commercially successful platforms like Duo Lingo use spaced retrieval practice to help people learn and retain information. However, a popular complaint about such tools is that you don't really learn skills that matter in the relevant contexts. For example, you would probably want to be able to communicate with someone when you travel to Spain, even if it is in broken Spanish, rather than develop a perfect memory for "El oso bebe leche" ("The bear drinks milk"). Curiously, there are people who find this entertaining and therefore stick with the lessons. While it is important to motivate people to stay consistent with their learning plan, it is clear that the training protocol itself needs to be tailored to the contexts in which the skills are to be used. So, in collaboration with colleagues working on AI and education, we are exploring how retrieval-practice protocols can be adapted to better match the contexts in which learned skills are actually used.

We also have ongoing studies on conditions like SDAM, (c)PTSD, etc and how they influence various memory processes. 

Summary

Our work broadly covers human memory and learning. While we took some time to set up infrastructure for neuroscience work, during the next phase in our lab's journey, we will focus on testing the neural mechanisms underlying memory as well as more general communication in the brain, using open datasets and novel EEG experiments while continuing to develop a few of the most promising threads from the first five years of our research at IIITH on memorability and related topics.

References

*+ equal contribution

1. Agrawal, Y.*, Rathore, P.*, & Sreekumar, V. (drafted and in submission). Is Word Memorability Task-Invariant? A Replication and Extension of Aka et al. (2023)
2. Maity, A., & Sreekumar, V. (2026). Functional role and semantic structure shape word memorability: Cue–target asymmetries in cued recall. Cognition, 274, 106592. [PDF] 
   
3. Kumar, P., Xie, W., Bainbridge, W., & Sreekumar, V. (In prep, to be submitted within a month). Memorability Emerges from Language-Specific Semantic Structure.
4. Agrawal, Y.*, Rathore, P.*, & Sreekumar, V. (in progress). Sentence memorability studies. 
5. Sunny, A., Gupta, A., Chandak, Y., & Sreekumar, V. (2025). From Stories to Statistics: Methodological Biases in LLM-Based Narrative Flow Quantification. (oral presentation, best paper award, $500 sponsored by Google DeepMind). Proceedings of the 29th SIGNLL Conference on Computational Natural Language Learning (CoNLL 2025, Vienna, Austria). [PDF]
   
6. Sunny, A.*, Gupta, A.*, & Sreekumar, V. (2025). Context is Enough: Empirical Validation of Sequentiality on Essays. arXiv preprint arXiv:2511.09185.
7. Kumar, P.*, Khandelwal, E.*, Tapaswi, M.+, & Sreekumar, V.+ (2025). Eye vs. AI: Human Gaze and Model Attention in Video Memorability. In Winter Conference on Applications of Computer Vision (WACV), Mar 2025 (oral presentation). [PDF]
   
8. Estes, W.K. (1950). Towards a statistical theory of learning. Psychological Review, 57, 94-107.
9. Howard, M. W., & Kahana, M. J. (2002). A distributed representation of temporal context. Journal of mathematical psychology, 46(3), 269-299.
   
10. Manning, J. R., Polyn, S. M., Baltuch, G. H., Litt, B., & Kahana, M. J. (2011). Oscillatory patterns in temporal lobe reveal context reinstatement during memory search. Proceedings of the National Academy of Sciences, 108(31), 12893-12897.
    
11. Hintzman, D. L. (2016). Is memory organized by temporal contiguity?. Memory & cognition, 44(3), 365-375.
    
12. Kurby, C. A., & Zacks, J. M. (2008). Segmentation in the perception and memory of events. Trends in cognitive sciences, 12(2), 72-79.
    
13. Sargent, J. Q., Zacks, J. M., Hambrick, D. Z., Zacks, R. T., Kurby, C. A., Bailey, H. R., ... & Beck, T. M. (2013). Event segmentation ability uniquely predicts event memory. Cognition, 129(2), 241-255.
    
14. Shin, Y. S., & DuBrow, S. (2021). Structuring memory through inference‐based event segmentation. Topics in cognitive science, 13(1), 106-127.
    
15. Davachi, L., & Murty, V. P. (2024). Introduction to the Special Focus: Remembering Sarah DuBrow. Journal of Cognitive Neuroscience, 36(11), 2299-2301.
    
16. Pooja, R.*, Ghosh, P.*, & Sreekumar, V. (2024). Towards an ecologically valid naturalistic cognitive neuroscience of memory and event cognition. Neuropsychologia, 203, 108970.
    
17. Ding, Y., & Zacks, J. M. (2026). Semantic knowledge and hierarchical event structure can scaffold memory for temporal order. Journal of Experimental Psychology: Learning, Memory, and Cognition.
    
18. Ding, Y., Alperin, D. R., & Zacks, J. M. (2026). Aging and memory for temporal order in naturalistic events. Psychology and aging.
    
19. Ghosh, P., Pooja, R., & Sreekumar, V. (2026). Selective effects of task change-driven event boundaries on recognition memory. Proceedings of the Annual Meeting of the Cognitive Science Society, 48 (Accepted as full paper). An extended paper has been drafted for a journal.
20. Sreekumar, V., Wittig Jr, J. H., Chapeton, J., Inati, S. K., & Zaghloul, K. A. (2020). Low-frequency traveling waves in the human cortex coordinate neural activity across spatial scales. BioRxiv, 2020-03.
    
21. Mohan, U. R., Zhang, H., Ermentrout, B., & Jacobs, J. (2024). The direction of theta and alpha travelling waves modulates human memory processing. Nature Human Behaviour, 8(6), 1124-1135.