ISSP 2030 Spring AI Forum Week 1

22 Jan 2021 - Arun Balajiee

ISSP 2030 AI Forum Week 1

Talk Date: 2021-01-22

Speaker: Steven Small

Neurobiology of Language In this talk, Dr. Small specified the various misconeptions in research questions pursued under the name of “Neurobiology” when they in fact should be categorized as psycholinguistics or simply linguistics. Then Dr. Small spent some time stipulating the definitions of “Neurobiology of language” to further elicit the novelty of the research and developments in the field by citing different problems as examples for research questions. Specifically, the guiding principles for research questions in this field are,

Human language evolved from sensory, motor, emotional and social functions.
Human language performance based on spatial-temporal data on multi-system interactions
Spatial encodings for language not just on one region or few regions of the brain and temporal encodings not just on a single time scale.

Ecological langauge processing of the human brain involves several asepcts such as sounds, events, speech, words, emotions, irony, gesutre – aspects of Multi-modal inputs processing. Modelling the human language by neurobiology can be performed mathematically but not necessarily into an AI based solution.

A surprising aspect is that language processing regions in the brain are scattered across different regions and there are about 44 different regions in the human that support language processing, comprehension and behaviour. These different regions are centrally connected by a mesh-like structure formed by the neurons in the brain.

Finally, Dr. Small ended the talk by presenting his book that could be used for reference and understanding research in the field.

Speaker: Harry Hochheiser

Clinician-focused machine learning

In this talk, Dr. Hochheiser focused on the design studies, structure of experiments, collaborative design with end-users that are very relevant to researchers in Human-Computer Interaction. Specifically, he talked about not building models that outperform precedents but models that affect and assist the everyday challenges faced by clinicians in their field. The system of model development involves constant dialogue and inputs from clinicians and building dataset of different features that could assist in building a model that processes the datasets. This also helps in eliminate probable causes for biases in model and the models becoming incomplete solutions to the actual real-life problems the models try to solve.

Further, however, Dr. Hochheiser cautioned the audience that despite being an inclusive design, the model could fall short on numerous counts in including the different challenges either because of an incomplete data collection during the design inquiry process, model development or other factors inherently biased in a model such as overfitting to a particular usecase or dataset.

The takeaways from the talk would be that the while developing the inclusive design for building models, one should not ask questions that seek direct answers or responses on the challenges, but discern the problems and the different factors leading to them as well as the different possible solutions through obversational and survey-based studies conducted longitudinally by the research team.