Avinash Vipin

Central Institute of Psychiatry

Author Statistics

Papers

Citation

H-Index

i-10 index

Research Trends

Author Order

Document Type

Co-Authors

Manul Das

Central Institute of Psychiatry

Justin Raj

Central Institute of Psychiatry

Nishant Goyal

Central Institute of Psychiatry

Subham Samantaray

Central Institute of Psychiatry

S.Mukesh Chowdary

Central Institute of Psychiatry

Urvakhsh Meherwan Mehta

National Institute of Mental Health and Neurosciences

Neelabja Roy

National Institute of Mental Health and Neurosciences

Srijan Das

Central Institute of Psychiatry

Liza Daniel

Manipal College of Medical Sciences

Cooperative Institutions

Central Institute of Psychiatry

National Institute of Mental Health and Neurosciences

National Institute of Mental Health

Jawaharlal Institute of Post Graduate Medical Education and Research

Creative Commons

Siksha O Anusandhan University

Indian Institute of Technology Bombay

Siddhartha Medical College

National Institute of Advanced Studies

Harvard University

Author Statistics

Papers

Citation

H-Index

i-10 index

Research Field

How Do Large Language Models Perform on Social Cognition Tasks Compared to Humans? A SOCRATIS Based Evaluation of Social Intelligence of GPT-4 and PaLM 2 Based AI Chatbots

Manul Das Avinash Vipin Justin Raj Urvakhsh Meherwan Mehta Nishant Goyal

GPT-4 and PaLM 2 are advanced large language models (LLMs) that have made remarkable progress and are widely used in various human applications. The extent to which these artificial intelligence systems possess social intelligence is pivotal for their seamless integration into everyday human interactions. In this study, we used the SOCRATIS tool to measure and compare the social cognitive abilities in first-order theory of mind, second-order theory of mind, faux pas recognition, and attributional styles between GPT-4 based Microsoft Bing, PaLM 2 based Google Bard, and healthy humans. Our results showed a significant difference in second-order theory of mind, faux pas recognition, and attributional styles among the groups. Specifically, post-hoc comparisons revealed that Bing, based on GPT-4, performed significantly better than Bard, based on PaLM 2, and healthy humans in second-order theory of mind and faux pas recognition tasks. Regarding attributional styles, Bard exhibited a higher personalizing bias compared to both Bing and healthy humans. Microsoft Bing on the other hand had significantly lower externalizing bias score compared to other two groups.

Social intelligence

10.2139/ssrn.4689399

Cite

Citations (0)

Deep Learning Models for Classification of Human Figure Drawings into Psychiatric Diagnostic Labels: A Proof of Concept Study

Manul Das S.Mukesh Chowdary Srijan Das Nishant Goyal Liza Daniel

Objective: The objective of this study was to evaluate the feasibility of using transfer learning with state- of-the-art Convolutional Neural Networks (CNNs) to classify human figure drawings of patients diagnosed with Mania and Schizophrenia.Methods: We collected a total of 230 images (108 from patients with Mania and 122 from patients with Schizophrenia). We experimented with several CNN models, including ResNet-50, InceptionV3, DenseNet-121, VGG19, and Inception-ResNet-v2, by utilizing their pre-trained weights on the ImageNet dataset to extract features from the human figure images. These features were then used to train a dense neural network head for binary classification. We employed 15-fold cross-validation for internal validation.Results: The Inception-ResNet-v2 model demonstrated the best performance, achieving an accuracy of 73.6 ± 9.4%, an AUC score of 0.72 ± 0.09, and an F1 score of 73 ± 20% in classifying Schizophrenia and Mania. Additionally, qualitative analysis revealed that features associated with psychosis and mania, as documented in existing literature, were frequently identified in high-probability classifications.Conclusion: This study underscores the potential of CNNs in classifying human figure drawings of psychiatric patients and highlights the need for continued exploration in this domain with larger and more diverse datasets.

Proof of concept

10.2139/ssrn.4848637

Cite

Citations (0)