Suresh Manandhar

Chaitanya Kaul, Nick Pears, Suresh Manandhar

Deep neural networks have established themselves as the state-of-the-art methodology in almost all computer vision tasks to date. But their application to processing data lying on non-Euclidean domains is still a very active area of research. One such area is the analysis of point cloud data which poses a challenge due to its lack of order. Many recent techniques have been proposed, spearheaded by the PointNet architecture. These techniques use either global or local information from the point clouds to extract a latent representation for the points, which is then used for the task at hand (classification/segmentation). In our work, we introduce a neural network layer that combines both global and local information to produce better embeddings of these points. We enhance our architecture with residual connections, to pass information between the layers, which also makes the network easier to train. We achieve state-of-the-art results on the ModelNet40 dataset with our architecture, and our results are also highly competitive with the state-of-the-art on the ShapeNet part segmentation dataset and the indoor scene segmentation dataset. We plan to open source our pre-trained models on github to encourage the research community to test our networks on their data, or simply use them for benchmarking purposes.

Sharad Duwal, Suraj Prasai, Suresh Manandhar

Continual learning has emerged as an important research direction due to the infeasibility of retraining large language models (LLMs) from scratch in the event of new data availability. Of great interest is the domain-adaptive pre-training (DAPT) paradigm, which focuses on continually training a pre-trained language model to adapt it to a domain it was not originally trained on. In this work, we evaluate the feasibility of DAPT in a low-resource setting, namely the Nepali language. We use synthetic data to continue training Llama 3 8B to adapt it to the Nepali language in a 4-bit QLoRA setting. We evaluate the adapted model on its performance, forgetting, and knowledge acquisition. We compare the base model and the final model on their Nepali generation abilities, their performance on popular benchmarks, and run case-studies to probe their linguistic knowledge in Nepali. We see some unsurprising forgetting in the final model, but also surprisingly find that increasing the number of shots during evaluation yields better percent increases in the final model (as high as 19.29% increase) compared to the base model (4.98%), suggesting latent retention. We also explore layer-head self-attention heatmaps to establish dependency resolution abilities of the final model in Nepali.

Love Panta, Suraj Prasai, Karishma Malla Vaidya et al.

Cervical cancer remains a significant health challenge, with high incidence and mortality rates, particularly in transitioning countries. Conventional Liquid-Based Cytology(LBC) is a labor-intensive process, requires expert pathologists and is highly prone to errors, highlighting the need for more efficient screening methods. This paper introduces an innovative approach that integrates low-cost biological microscopes with our simple and efficient AI algorithms for automated whole-slide analysis. Our system uses a motorized microscope to capture cytology images, which are then processed through an AI pipeline involving image stitching, cell segmentation, and classification. We utilize the lightweight UNet-based model involving human-in-the-loop approach to train our segmentation model with minimal ROIs. CvT-based classification model, trained on the SIPaKMeD dataset, accurately categorizes five cell types. Our framework offers enhanced accuracy and efficiency in cervical cancer screening compared to various state-of-art methods, as demonstrated by different evaluation metrics.

Chaitanya Kaul, Nick Pears, Suresh Manandhar

The application of deep learning to 3D point clouds is challenging due to its lack of order. Inspired by the point embeddings of PointNet and the edge embeddings of DGCNNs, we propose three improvements to the task of point cloud analysis. First, we introduce a novel feature-attentive neural network layer, a FAT layer, that combines both global point-based features and local edge-based features in order to generate better embeddings. Second, we find that applying the same attention mechanism across two different forms of feature map aggregation, max pooling and average pooling, gives better performance than either alone. Third, we observe that residual feature reuse in this setting propagates information more effectively between the layers, and makes the network easier to train. Our architecture achieves state-of-the-art results on the task of point cloud classification, as demonstrated on the ModelNet40 dataset, and an extremely competitive performance on the ShapeNet part segmentation challenge.

Chaitanya Kaul, Nick Pears, Hang Dai et al.

We propose a new residual block for convolutional neural networks and demonstrate its state-of-the-art performance in medical image segmentation. We combine attention mechanisms with group convolutions to create our group attention mechanism, which forms the fundamental building block of our network, FocusNet++. We employ a hybrid loss based on balanced cross entropy, Tversky loss and the adaptive logarithmic loss to enhance the performance along with fast convergence. Our results show that FocusNet++ achieves state-of-the-art results across various benchmark metrics for the ISIC 2018 melanoma segmentation and the cell nuclei segmentation datasets with fewer parameters and FLOPs.

Chaitanya Kaul, Nick Pears, Hang Dai et al.

Loss functions are error metrics that quantify the difference between a prediction and its corresponding ground truth. Fundamentally, they define a functional landscape for traversal by gradient descent. Although numerous loss functions have been proposed to date in order to handle various machine learning problems, little attention has been given to enhancing these functions to better traverse the loss landscape. In this paper, we simultaneously and significantly mitigate two prominent problems in medical image segmentation namely: i) class imbalance between foreground and background pixels and ii) poor loss function convergence. To this end, we propose an adaptive logarithmic loss function. We compare this loss function with the existing state-of-the-art on the ISIC 2018 dataset, the nuclei segmentation dataset as well as the DRIVE retinal vessel segmentation dataset. We measure the performance of our methodology on benchmark metrics and demonstrate state-of-the-art performance. More generally, we show that our system can be used as a framework for better training of deep neural networks.

Marcelo Sardelich, Suresh Manandhar

Stock market volatility forecasting is a task relevant to assessing market risk. We investigate the interaction between news and prices for the one-day-ahead volatility prediction using state-of-the-art deep learning approaches. The proposed models are trained either end-to-end or using sentence encoders transfered from other tasks. We evaluate a broad range of stock market sectors, namely Consumer Staples, Energy, Utilities, Heathcare, and Financials. Our experimental results show that adding news improves the volatility forecasting as compared to the mainstream models that rely only on price data. In particular, our model outperforms the widely-recognized GARCH(1,1) model for all sectors in terms of coefficient of determination $R^2$, $MSE$ and $MAE$, achieving the best performance when training from both news and price data.

Rohan Watkins, Nick Pears, Suresh Manandhar

We investigate whether ResNet architectures can outperform more traditional Convolutional Neural Networks on the task of fine-grained vehicle classification. We train and test ResNet-18, ResNet-34 and ResNet-50 on the Comprehensive Cars dataset without pre-training on other datasets. We then modify the networks to use Spatially Weighted Pooling. Finally, we add a localisation step before the classification process, using a network based on ResNet-50. We find that using Spatially Weighted Pooling and localisation both improve classification accuracy of ResNet50. Spatially Weighted Pooling increases accuracy by 1.5 percent points and localisation increases accuracy by 3.4 percent points. Using both increases accuracy by 3.7 percent points giving a top-1 accuracy of 96.351\% on the Comprehensive Cars dataset. Our method achieves higher accuracy than a range of methods including those that use traditional CNNs. However, our method does not perform quite as well as pre-trained networks that use Spatially Weighted Pooling.