Projects

Effect of perturbations in the images on Recurrent Visual Attention Model

Course Project: COMPSCI 691DD - Research Methods in Empirical Computer Science, UMass Amherst

ABSTRACT: In this study, I have examined the effect of three different perturbation - Gaussian Blur, Motion Blur and Reduced Contrast - on the test images to a Recurrent Visual Attention Model (RAM), a LSTM based classifier, trained using REINFORCE algorithm. While using noisy-MNIST dataset, it was observed that RAM was more resilient to motion blurred images compared to gaussian blur and contrast reduction. To find a causal reasoning for this set of experiments by varying the severity of perturbations were made to see the effect of such variations on the models performance. The results show that the performance of network is highly dependent on severity of perturbations in images. This raises questions about the way noisy-MNIST dataset was created and on its use by other researchers, as the level of severity in all the three different datasets is not similar.
LINK: Report

Lamprey Detection Using LampNet

Course Project: COMPSCI 682 - Neural Networks, UMass Amherst

ABSTRACT: Sea lampreys have become a major pest in the North American Great Lakes. They are considered as an invasive species that prey on many species of commercial value such as lake trouts, fishes and perches. To tackle this problem, we introduce LampNet, a Convolutional Neural Network(CNN) based classifier to detect lampreys in videos. We evaluate the efficacy of the LampNet on the dataset consisting of around 10,629 images with three classes (lamprey, fish and background). Our evaluations show that LampNet can identify and detect lampreys with an F1-score of 0.97 and an overall ROC-AUC score of 0.995. Further, it also shows better performance than previous work involving AlexNet on similar dataset. This framework could be extended to replace the human operating the fish elevator to prevent the lamprey from entering the water bodies
LINK: Report

Traffic Sign Prediction Using Modified Q- Learning

Guide: Prof. Srinivasa Chakravarthy, Computational Neuroscience Lab, IIT Madras

ABSTRACT: Traffic sign detection is an important component of Advanced Driver Assistance Systems (ADAS). Deep learning has contributed immensely in progress of research in this field. The current state-of-the-art methods for traffic sign detection like Faster R-CNN and YOLO9000 use convolutional neural networks (CNNs) and were developed to detect objects in an image. A more naturalistic approach to detect traffic signs is to use road video data to harness temporal information of traffic sign movement. We propose a Deep Q Neural Network for predicting the location of traffic signs in the next frame, using the current frame. The proposed network has two parts: feature extraction and “value” estimation using reinforcement learning. First, a pretrained Alexnet network is used to extract the features of the input frames. Secondly, a convolutional layer was added, that outputs a feature map of dimension 29 x 63, which computes the value estimation of the input frame. A reward of 1 at the location where next frame contains traffic signs is defined for each frame and is resized to 29 x 63. The weights of the network are fine-tuned for grayscale frames of road video data by back propagating the temporal difference error. The network was trained on 12,224 frames and tested on 715 frames. The model outputs an activity map, and after contour detection, the coordinates where the traffic sign might appear in the next frame are obtained. Our model achieved mean average precision (MAP) of 56.001% which beats both Faster R-CNN and YOLO9000 with MAP of 53.85% and 49.648% respectively. Our model took 18 hours for training on GPU compared to 36 and 83 hours for Faster R-CNN and YOLO9000 respectively. Thus our model not only outperforms both the state-of-the-art models in accuracy but also takes lesser time to train and generate predictions.
RESULTS: The video depicting the performance on Faster R-CNN, YOLO v2 and Prediction Network is shown below.LINKS: More details in the thesis, extended abstract and poster.

Prediction_Network.mp4

Prediction Network

Faster_RCNN_Triangle_Night.mp4

Faster RCNN

YOLO_Triangle_Night.mp4

YOLOv2

Biosensors

Biotech Club, Center for Innovation, IIT Madras

ABSTRACT: The contamination of the environment and its direct relation to human health has raised the standards pertaining to discharge of liquid effluents into water bodies. The dramatic increase in anthropogenic activity over the past decade has led to concerns about the sustenance of life on earth. Monitoring the pollution levels and making sustainable development is the only way towards maintaining stewardship of the environment. It is important to have methodologies to assess the quality of water. So, we have developed a novel biosensor that can assess the efficiency of wastewater treatment plants. Biochemical Oxygen Demand (BOD), is one of the primary parameters used to determine the level of toxicity of waste water and is defined as the amount of dissolved oxygen needed by aerobic biological organisms to metabolize the organic material present in the sample. The key sensing element of our device is the recombinant E. coli strain containing the lux genes A-E, which encodes for the luciferin-luciferase system. The genes are kept under the control of an arabinose based switch and characterized for optimal concentration of arabinose. The carbon source is metabolised via the glycolysis-TCA pathway to generate ATP and NADPH which regenerate the substrates required for the above operon. The presence of biodegradable organic pollutants would lead to bioluminescence and this would be correlated to the conventional BOD5 values. In order to improve the functionality of the device viz. stability and re-usability, the genetically engineered microbes have been immobilized in sodium alginate beads. This immobilized cell strategy would lead to increased stability as compared to the existing O2 electrode based sensing devices.

LINKS: More details in the conference paper and poster

Studying Life-Cycle of S.litura and H. Armigera

Guide: Prof. Radhika Venkatesan, National Center for Biological Sciences, TIFR

ABSTRACT: The study was done on S.litura (known as tobacco cutworm) and H.armigera (known as cotton bollworm), two different species of moths which collectively feed on more than 200 different species of agricultural crops and are biggest pest in many of the Asian countries including India. The main focus was to monitor their feeding behaviour on different diet. Larval weight was the parameter used to track the progress. Apart from feeding behaviour, there were experiments done to study the effect of two plant hormones: Methyl Jasmonic acid (MeJA) and Salicylic acid (SA); on the growth of S.litura larvae. The S.litura larvae were tested with two different antibiotics mixed with their diet, Tetracycline and Ampicillin mixed with Cloxacillin. The data from the experiments suggested that both S.litura and H.armigera feed more during night time. Their growth rate varied during day and night feeding, with night feeding larvae growing faster than larvae feeding during day time. The difference between the growth rates of S.litura larvae feeding during night and that of day was more in castor leaves as compared to the artificial diet. In a study conducted for 6 days, the S.litura larvae which were subjected to concentrations more than 0.5 mM of MeJA showed early and high mortality rate as compared to larvae subjected to SA for similar concentration. This experiment gave an insight on the minimum concentration of MeJA to be produced by plants, lethal to the larvae. The pupal weight of the larvae feeding on the castor leaves was found to be almost twice of the larvae feeding on the artificial diet. Also, night feeding larvae on artificial diet had more pupal weight compared to day feeding larvae. This shows that the larvae survive better if they have adequate amount of food and that they found more nutrients in the castor leaves than the artificial diet.
LINKS: More details in the report.