Manju Bhardwaj

Visually impaired people (VIP) need assistive technology for smooth and confident navigation on pavements along the roads. VIPs find it hard to follow the speech generated through a text-to-audio converter due to the presence of advertisement text. Automating the process of audio conversion from the signboards by integrating the capabilities of emerging AI models in a fraction of a second will be useful for their confident navigation. In this context, SignSpeak, an AI-based assistive utility, is designed to convert signboard text to speech using visual inputs, which may be captured through smart sensors or canes. The proposed system follows a three-step (detection-extraction-conversion) pipeline. Initially, a fine-tuned lightweight YOLOv11s model is used to generate a bounded box for the relevant text region, which is input to Gemini model, a multimodal large language model for text extraction. Lastly, the extracted text is converted into audio using the Google Text-to-Speech tool. We fine-tuned the YOLOv11s model on 200 signboard images and tested its performance on an additional curated dataset of 42 real-world signboard images. Experimental findings indicate that the proposed integrated pipeline achieved superior performance compared to the solitary usage of the Gemini model for text extraction followed by its audio conversion.

Improved classification for imbalanced data using ensemble clustering

Predicting the class to which a given protein sequence belongs is a challenging research area in bioinformatics. Machine learning techniques have been successfully applied to protein prediction problems like allergen prediction, mitochondrial prediction and toxin prediction. Physicochemical properties derived from sequences of amino acids have been commonly used for this purpose. In this paper, we propose an SVM based ensemble method for classification of protein datasets. The constituent classifiers of the ensemble are generated in a sequential manner, each one attempting to rectify mistakes made by previous one. The ensemble is aptly called Self-Chastisting Ensemble (SCE) because of the iterative refinement each classifier carries out over the previous one. We present two versions of the algorithm: SCE-Bal for balanced datasets and SCE-Imbal for imbalanced datasets. Empirical results further demonstrate that the algorithm delivers superior performance using simple and computationally efficient features (amino acid composition and dipeptide composition) compared to other machine learning methods using complex feature sets.

Eubacterial genomes vary considerably in their nucleotide composition. The percentage of genetic material constituted by guanosine and cytosine (GC) nucleotides ranges from 20% to 70%. It has been posited that GC-poor organisms are more dependent on protein folding machinery. Previous studies have ascribed this to the accumulation of mildly deleterious mutations in these organisms due to population bottlenecks. This phenomenon has been supported by protein folding simulations, which showed that proteins encoded by GC-poor organisms are more prone to aggregation than proteins encoded by GC-rich organisms. To test this proposition using a genome-wide approach, we classified different eubacterial proteomes in terms of their aggregation propensity and chaperone-dependence using multiple machine learning models. In contrast to the expected decrease in protein aggregation with an increase in GC richness, we found that the aggregation propensity of proteomes increases with GC content. A similar and even more significant correlation was obtained with the GroEL-dependence of proteomes: GC-poor proteomes have evolved to be less dependent on GroEL than GC-rich proteomes. We thus propose that a decrease in eubacterial GC content may have been selected in organisms facing proteostasis problems.

Real life datasets often suffer from the problem of class imbalance, which thwarts supervised learning process. In such data sets examples of positive (minority) class are significantly less than those of negative (majority) class leading to severe class imbalance. Constructing high quality classifiers for such imbalanced training data sets is one of the major challenges in machine learning, since traditional classification algorithms tend to get biased towards majority class. In this paper, we compare three ensemble based approaches for handling imbalanced datasets. All the three approaches aim to overcome the under representation of minority class by learning from each of the minority class samples and a subset of majority class samples. The three approaches namely, PARTEN, UMjC and LFM were evaluated on several public datasets. Precision, recall, F- measure, g-mean and ROC space measures were used for comparison. Thread-bare discussion of the results is presented in the paper. Subsequently, we present an astronomy application, where the three methods are compared for prediction of class II, IIn and IIp supernovae.