ASPIRE: Language-Guided Augmentation for Robust Image Classification

Neural image classifiers can often learn to make predictions by overly relying on non-predictive features that are spuriously correlated with the class labels in the training data. This leads to poor performance in real-world atypical scenarios where such features are absent. Supplementing the training dataset with images without such spurious features can aid robust learning against spurious correlations via better generalization. This paper presents ASPIRE (Language-guided data Augmentation for SPurIous correlation REmoval), a simple yet effective solution for expanding the training dataset with synthetic images without spurious features. ASPIRE, guided by language, generates these images without requiring any form of additional supervision or existing examples. Precisely, we employ LLMs to first extract foreground and background features from textual descriptions of an image, followed by advanced language-guided image editing to discover the features that are spuriously correlated with the class label. Finally, we personalize a text-to-image generation model to generate diverse in-domain images without spurious features. We demonstrate the effectiveness of ASPIRE on 4 datasets, including the very challenging Hard ImageNet dataset, and 9 baselines and show that ASPIRE improves the classification accuracy of prior methods by 1% - 38%. Code soon at: https://github.com/Sreyan88/ASPIRE.

Multiview Scattering Scanning Imaging Confocal Microscopy through a Multimode Fiber

Confocal and multiphoton microscopy are effective techniques to obtain high-contrast images of 2-D sections within bulk tissue. However, scattering limits their application to depths only up to ~1 millimeter. Multimode fibers make excellent ultrathin endoscopes that can penetrate deep inside the tissue with minimal damage. Here, we present Multiview Scattering Scanning Imaging Confocal (MUSSIC) Microscopy that enables high signal-to-noise ratio (SNR) imaging through a multimode fiber, hence combining the optical sectioning and resolution gain of confocal microscopy with the minimally invasive penetration capability of multimode fibers. The key advance presented here is the high SNR image reconstruction enabled by employing multiple coplanar virtual pinholes to capture multiple perspectives of the object, re-shifting them appropriately and combining them to obtain a high-contrast and high-resolution confocal image. We present the theory for the gain in contrast and resolution in MUSSIC microscopy and validate the concept through experimental results.

IITK at the FinSim Task: Hypernym Detection in Financial Domain via Context-Free and Contextualized Word Embeddings

In this paper, we present our approaches for the FinSim 2020 shared task on "Learning Semantic Representations for the Financial Domain". The goal of this task is to classify financial terms into the most relevant hypernym (or top-level) concept in an external ontology. We leverage both context-dependent and context-independent word embeddings in our analysis. Our systems deploy Word2vec embeddings trained from scratch on the corpus (Financial Prospectus in English) along with pre-trained BERT embeddings. We divide the test dataset into two subsets based on a domain rule. For one subset, we use unsupervised distance measures to classify the term. For the second subset, we use simple supervised classifiers like Naive Bayes, on top of the embeddings, to arrive at a final prediction. Finally, we combine both the results. Our system ranks 1st based on both the metrics, i.e., mean rank and accuracy.

IITK at SemEval-2020 Task 8: Unimodal and Bimodal Sentiment Analysis of Internet Memes

Social media is abundant in visual and textual information presented together or in isolation. Memes are the most popular form, belonging to the former class. In this paper, we present our approaches for the Memotion Analysis problem as posed in SemEval-2020 Task 8. The goal of this task is to classify memes based on their emotional content and sentiment. We leverage techniques from Natural Language Processing (NLP) and Computer Vision (CV) towards the sentiment classification of internet memes (Subtask A). We consider Bimodal (text and image) as well as Unimodal (text-only) techniques in our study ranging from the Na\"ive Bayes classifier to Transformer-based approaches. Our results show that a text-only approach, a simple Feed Forward Neural Network (FFNN) with Word2vec embeddings as input, performs superior to all the others. We stand first in the Sentiment analysis task with a relative improvement of 63% over the baseline macro-F1 score. Our work is relevant to any task concerned with the combination of different modalities.