Leveraging Data to Say No: Memory Augmented Plug-and-Play Selective Prediction

Selective prediction aims to endow predictors with a reject option, to avoid low confidence predictions. However, existing literature has primarily focused on closed-set tasks, such as visual question answering with predefined options or fixed-category classification. This paper considers selective prediction for visual language foundation models, addressing a taxonomy of tasks ranging from closed to open set and from finite to unbounded vocabularies, as in image captioning. We seek training-free approaches of low-complexity, applicable to any foundation model and consider methods based on external vision-language model embeddings, like CLIP. This is denoted as Plug-and-Play Selective Prediction (PaPSP). We identify two key challenges: (1) instability of the visual-language representations, leading to high variance in image-text embeddings, and (2) poor calibration of similarity scores. To address these issues, we propose a memory augmented PaPSP (MA-PaPSP) model, which augments PaPSP with a retrieval dataset of image-text pairs. This is leveraged to reduce embedding variance by averaging retrieved nearest-neighbor pairs and is complemented by the use of contrastive normalization to improve score calibration. Through extensive experiments on multiple datasets, we show that MA-PaPSP outperforms PaPSP and other selective prediction baselines for selective captioning, image-text matching, and fine-grained classification. Code is publicly available at https://github.com/kingston-aditya/MA-PaPSP.

Virtual Screening of Pharmaceutical Compounds with hERG Inhibitory Activity (Cardiotoxicity) using Ensemble Learning

In silico prediction of cardiotoxicity with high sensitivity and specificity for potential drug molecules can be of immense value. Hence, building machine learning classification models, based on some features extracted from the molecular structure of drugs, which are capable of efficiently predicting cardiotoxicity is critical. In this paper, we consider the application of various machine learning approaches, and then propose an ensemble classifier for the prediction of molecular activity on a Drug Discovery Hackathon (DDH) (1st reference) dataset. We have used only 2-D descriptors of SMILE notations for our prediction. Our ensemble classification uses 5 classifiers (2 Random Forest Classifiers, 2 Support Vector Machines and a Dense Neural Network) and uses Max-Voting technique and Weighted-Average technique for final decision.