It’s been a while! I had a lot of grand ambitions for keeping up this blog consistently, but obviously that did not come to pass. The fact is that a lot has happened since mid-2022! The project I was working on at that time ended up being kind of a disappointment (I can blame the nature of human subjects experiments until I’m blue in the face… and I will). Since then I made a substantial pivot in my research direction, focusing less explicitly on education (though it remains a topic of great importance to me!) and more on how robots and AI can act as embedded cognitive supports in other environments. Accordingly, I transitioned away from my work with the inter-departmental institute I had been working with at the time, and suddenly found myself operating even more independently, with an overwhelming amount of freedom to pursue whatever I wanted.
I also was diagnosed with a chronic illness, took a leave of absence to get a major surgery, got COVID again, dealt with some family health problems, continued to struggle with my mental health and severe imposter syndrome, got diagnosed with ADHD and started medication, and generally did a lot of floundering when it came to not only my research direction but my life trajectory, passions, and values. Ah, the joys of young adulthood. I’m very grateful to my advisors for allowing me the space and time to take care of my physical & mental health, and helping me find a research thread that I’m passionate about.
Another big change that’s happened since 2022 is LLMs suddenly being everywhere and everything. The nature of HAI and HRI work has completely changed as a result. With LLMs and multi-modal language models (MLMs) suddenly on every app and website you can imagine, I think it’s become apparent that these models are extremely powerful – but also that they are not a magic bullet or a solve to general intelligence. This, to me, is more exciting than the alternative! We have this new powerful tool whose limitations we are still working to discover and mitigate. How do we wield these models to their fullest potential? As a spoiler, I think the answer lies in leaning into LLM’s strength – lingustic operations! – and combining them with other AI techniques to get both the usability of LLMs and the interpretability and consistency that we want in our AI (or at least I do). This is something we’ve already started to see from major LLMs like ChatGPT: by incorporating retrieval-augmented generation, GPT is now able to source its claims from a live knowledge base (i.e. the internet), rather than relying on implicit knowledge gleaned from outdated training data to answer user questions.
Right now, my research topic is in the form of a target application area – task handover (no, not literal object handover from a robot). The inspiration for this topic was actually patient handover in hospitals. This is a difficult cognitive task that nurses and other clinicians must perform at the end of every shift. The outgoing nurse must provide the incoming nurse with all of the relevant information that they need to ensure continuity of care for the patient, for each patient. This info can include aspects of the patient’s medical history, active conditions, the outgoing nurse’s action history, any tests the patient might be waiting on, conditional recommendations… the list goes on. Despite the proliferation of electronic health records and standardized handover protocols, existing systems still fail to offload the cognitive and temporal burden faced by clincians in maintaining and delivering this information.
So, that’s what my area exam (and probably, eventually, my dissertation!) is going to be about: how can AI support the handover process? What advances in things like language models, sensing, and model reconciliation can be applied to this challenging problem space?
Now I’m going to start writing about the process of doing this area exam. Wish me luck.