Overview

The research project focuses on developing an FFN (Feed-Forward Network) with a unique feature of "Artificial Curiosity". Artificial Curiosity in the context of this project is a process that involves applying affine transformations to input test data and using a model trained on regular; non-transformed data to predict on the transformed data. The model then filters the predictions based on the least entropy for log-softmax values, specifically softmax values. The choice of the word "curiosity" to describe the process of applying different transformations to test images makes sense, as it reflects the idea of exploring different possibilities to make a prediction. The model is "curiosity-driven" in the sense that it explores different possibilities of orientations and mirroring of the test image in order to make a prediction, rather than relying solely on what it was trained on. This demonstrates some level of creativity in the sense that the model is able to find new ways of solving the problem of image recognition, rather than just being limited to what it was trained on. While the number and types of transformations used in this project are limited for demonstration purposes, the concept of artificial curiosity can be extended to include a range of creative transformations. This project uses a feed-forward neural network (FFN) because FFNs lack spatial perception and are unable to recognize images that deviate significantly from the training data. This weakness of FFNs is exploited to demonstrate the concept of artificial curiosity. When the model is predicting with 'curious' mode, it can recognize manipulated images, while in 'standard' mode, it cannot. This distinction provides a clear demonstration of the advantages and limitations of the two modes.

Primary Objective

The primary objective of this project is to develop a Feed-Forward Network (FFN) model with a unique feature of "Artificial Curiosity". The project aims to show that the model is able to find new ways of solving the problem of image recognition, demonstrating some level of creativity. The project also aims to demonstrate the advantages and limitations of the "curious" and "standard" modes of the model in recognizing manipulated images.

Results

'Accuracy score' was used to calculate 'Versatility score', therefore in the experiment Versatility score is defined as the difference between accuracy of curious and non-curious modes of the model. In standard mode the model achieved an average accuracy of 0.25 while the curious mode achieved an average accuracy of 0.61. This resulted in a Versatility score of 0.36, this implies with the help of artificial curiosity the model had an improvement of 36% compared to its standard classification ability, which is quite interesting. The null hypothesis was rejected owing to the p-value being 9.50e-113 which is well below the common significance level of 0.05. The negative t-statistic value of -133.77 obtained in the test indicated that the curious mode (second group) had a significantly higher mean accuracy than the standard mode (first group).

Overview

This open-source project focuses on creating an AI solution for detecting retinal tissue abnormality using optical coherence tomography (OCT) images and deep neural networks. Fine-Tuned, validated and tested ImageNet pre-trained DenseNet-121 model on a large medical dataset of size 109K images. Hyperparameter Tuning, Layer Freezing, Custom Heads and alternate neural network architectures were also explored. The model is deployed as a web service on Streamlit Community Cloud.

Primary Objective

The primary objective of this project is to develop an AI model to detect abnormalities in retinal tissues through optical coherence tomography (OCT) images and deploy the model as a web sevice on Streamlit community cloud.

Results

DenseNet-121 showed promising performance on training, validation and test sets with metrics like accuracy, recall, precision, f1 and ROC-AUC ranging from 94% to 98% without any signs of overfitting.

An 80-foot coal seam at the North Antelope Rochelle opencut coal mine.
By Peabody Energy, Inc. - Provided by Peabody Energy, CC BY 3.0, https://commons.wikimedia.org/w/index.php?curid=36846291

Overview

This end-to-end machine learning project is focused on analyzing coal usage in the United States from 2001 to 2021 and classifying coal using data gathered from the U.S. Energy Information Administration (EIA) through their API.

1. The data was cleaned and prepared for analysis using Geospatial Analysis, Chemometrics and coal production Time Series Analysis (only Trend).
2. Advanced Custom Transformers (`StratifiedStatisticalImputer` and `MultivariateStratifiedOutlierRemover`) were built for Feature Engineering.
3. Specialized Transformation Pipelines were implemented for convenient preprocessing of data.
4. Machine Learning algorithms like Softmax Regression, Decision Tree Classifier, Random Forest Classifier and Feed Forward Network were implemented and cross-validation was used to evaluate their performance.
5. Hyperparameter tuning was applied to improve the performance of Feed Forward Network.
6. The Random Forest classification model was deployed using Flask on Render Cloud Hosting.

Primary Objective

To analyze coal data collected from U.S. Energy Information Administration, build machine learning models to classify coal based on parameters like heat content, ash content and sulphur content then deploy best performing model for educational purposes.

Results

With performance measures like accuracy, precision, recall, and F1 score all greater than 99%, the classification model demonstrated outstanding results in its ability to accurately classify the data.

Overview

Built an image compression solution from scratch using NumPy, Matplotlib, Streamlit and Singular Value Decomposition (SVD). The web service allows a user to upload an image, reduce file size without compromising on image clarity and visualize the data associated with different channels using singular value analysis. The web app is deployed with Streamlit Community Cloud and can be accessed for free through the link below. Please note that with our current resources, images less than 1MB work seamlessly.

Primary Objective

To build a simple and intuitive image compression solution by harnessing the power of singular value decomposition (SVD).

Results

The product is successfully created and deployed as a web service on Streamlit Community Cloud. Version 1.0.0 is accessible through the link below.

Designed by Vilmosvarga / Freepik

Overview

This end-to-end machine learning project is focused on predicting medical insurance price using regression. Data was collected from Kaggle and cleaned for Exploratory Data Analysis using Statistical Analysis and Feature Engineering. A Custom Transformer was also built for Feature Engineering. Transformation pipelines were implemented for convenient preprocessing of existing and new data. Machine Learning techniques such as Linear Regression, Polynomial Regression, Decision Tree Regression, Support Vector Regression and Random Forest Regression were applied for creation of models and cross-validation was used to evaluate their performance. Hyperparameter Tuning was applied using GridSearchCV to improve performance of some models. The Random Forest classification model was deployed using Flask on Render Cloud Hosting.

Primary Objective

To develop regression models that predict the medical insurance cost for an individual based on their personal information like age, sex, BMI, number of children, region and their lifestyle habits such as smoking using the dataset available on Kaggle. R^2 score will be used to evaluate model performance and the best performing model will be deployed for educational purposes.

Results

After Hyperparameter Tuning, Random Forest Classification model achieved the highest R^2 score (0.8737) on test set without any signs of overfitting on the training data, therefore it was chosen for deployment.

Overview

This is project focuses on fine-tuning a BERT Base Uncased model for sentiment snalysis of IMDB reviews using HuggingFace Framework. The workflow involves loading the IMDB dataset using datasets library, preprocessing the data, tokenization, dynamic padding with DataCollator, defining training arguments, defining the model and parameters, defining compute metrics, defining Trainer object and then fine-tuning the transformer. The fine-tuned model is then tested on sampled reviews.

Primary Objective

To fine-tune a BERT Base Uncased model for sentiment analysis of IMDB reviews.

Results

Achieved an accuracy score of 0.9352 after fine-tunining the model on IMDB dataset. The model can be accessed through the link below.

Information about this project will be updated soon.

Information about this project will be updated soon.

Overview

The two-point masses model serves as a foundational concept for a novel frictionless braking system based on the conservation of angular momentum. The model considers a system of two masses attached to strings of equal length 'r' and revolving around a central point. The objective is to harness the conservation of angular momentum to achieve controlled deceleration or acceleration by manipulating the radial velocities of the masses.

Mathematical Derivation Key Steps

1. Conservation of Angular Momentum
2. Net Angular Momentum
3. Deriving r^2 in terms of 'omega' and 'L'
4. Derivation of Radial Velocity (dr/dt)
5. Angular Acceleration (alpha)
6. Torque and Braking/Acceleration

Primary Objective

To study the feasibility of a frictionless braking system using the principle of conservation of angular momentum for rotating elements. Develop theoretical model and understand the differential equations governing the motion.

Results

- The mathematical derivation provides insight into how altering the radial velocity of masses can lead to controlled angular acceleration, which in turn affects the angular momentum of the system. This theoretical framework forms the basis for designing a frictionless braking system that utilizes fluid flow to generate momentum for braking or acceleration.

- While the two-point masses model offers a simplified representation, it highlights the potential of the conservation of angular momentum as a mechanism for braking and acceleration. The derived torque equation demonstrates how changing fluid flow or radial velocities can influence system behavior.

- The results of this model emphasize the importance of careful engineering, fluid dynamics analysis, and control mechanisms to translate the theoretical concept into practical applications. Further research, experimentation, and validation are essential to assess the feasibility, efficiency, and practicality of this frictionless braking system in real-world scenarios.

Overview

Hydraulic pump is a device which converts mechanical power into hydraulic energy. Pumps incorporated with active valves provide active control on both the inlet and outlet valves. Since controlled actuation is required, piezoelectric actuators are used as the actuation system. Piezo-actuators can produce higher force and can be operated at higher actuation frequencies. In this work active valves are designed for a high pressure piezo-hydraulic pump. Valve Sealing Plates for piezoelectric actuators are designed for the valve actuation. The Valve Sealing Plate is analysed using ANSYS analysis software. Suitable mathematical model is developed for valve actuator using MATLAB software. An alternate design for active valve actuation using Rotary Valve Actuator is proposed and analysed. MATLAB Simulink is used to develop the mathematical model which synchronizes the motions of pumping actuator and rotary valve. Also, simulations are carried out for the determination of pumped volume and flow rate of a particular cylinder.

Primary Objective

The purpose of this project is to design active valves for the piezo hydraulic pump. The main aim of the work is to optimise the design and the flow rate. The following were the primary objectives of the project:
1. Design of a valve sealing plate for piezo actuated active valve.
2. Design and development of ‘Rotary active valve’ for the piezo-hydraulic pump.

Results

In this work, design of a ‘Valve Sealing Plate’ incorporated with active valves was carried out. Two designs were developed, out of which one of the designs was fabricated due to the machining limitations in the other design. The analysis of both the designs have been carried out using Ansys. Mathematical modelling was carried out for the valve actuator with the valve sealing plate and the displacement of the piezo actuator achieved was 289.89 microns. The valve sealing plate was fabricated using spring steel. Another design is ‘Rotary Valve Actuator’. In this work, design of ‘Rotary Valve Actuator’ was done and multiple variants were explored. A stepper motor would be driving an internal cylinder with a slot on its lateral surface area. Mathematical modelling was done to synchronize the motion of the pumping actuator with that of the stepper motor.

Overview

The focus of this project is to demonstrate a prototype and propose a design of a mechanical system that rectifies direction of rotation. It is analogous to a full-wave electrical rectifier and therefore the name 'Mechanical Rectifier Mechanism'. This mechanical system is the result of an advanced application of modified bevel gears. The mechanism converts input rotation of any direction (clockwise or counterclockwise) to one desired direction of output rotation depending on the design. By its nature it also effectively locks the gear train for one input direction of rotation and facilitates freewheeling in the other direction thereby completely cutting off power transmission when used in reverse irrespective of the direction of rotation. Another important characteristic of this design is the continuous uninterrupted transmission of mechanical power from input to output even when the input rotation is bidirectional. This system was theorized to enable functioning of another complex yet compact automobile transmission named 'Concentric Gears Transmission'. However, the feasibility study of this transmission was discontinued later.

Overview

The focus of this project is to demonstrate a prototype of a electronic drawing tool that integrates the functionality of compass, divider and protractor. The selling point of this device is the fact that it doesn't require a ruler for setting radius rather it is achieved using electronics. A 10k ohm linear potentiometer is used for pivoting arms of the device. An Arduino Nano microcontroller board is used for under the hood computation. An OLED 0.96" display is used for displaying radius and angular measurements, all of which are powered by two 3V button cells in series. The device is engineered to convert analog voltage resulting from mechanical rotation of the arms into a digital signal using Arduino's ADC and then calibrated based on the design of the potentiometer to compute rotation in degrees. Once the rotational information is processed, the distance between the tips of the arms is calculated using cosine rule.

Overview

The focus of this project is to demonstrate a prototype of an electronic device that can measure distance between "two" distant objects in front of it. Consider a triangular setup ABC, where point C is the location of the device, points A and B are the locations of target objects. A basic version of this device is primarily designed to compute the distance 'AB'. The device consists of two ultrasonic distance measuring sensors fitted on its two arms, a 10k ohm potentiometer, an arduino microcontroller, an OLED 0.96" display and a battery. The two arms of this hand held device are directed to the two targets in front of it and a button is pressed. The two ultrasound sensors measure the distances CA and CB, the potentiometer measures the angle between the two arms simultaneously. This information is then used by the microcontroller to compute the distance AB using Trigonometry. An advanced version of this device which is out of the scope of this project, integrated with servo motors to control the movement of arms and real time video processing can provide information like relative velocity and relative acceleration of the targets by computing first and second time derivatives of displacement signal streams from electromagnetic radiation based distance sensors. This kind of systems may find potential applications in advanced weapon systems and space rovers.

Overview

This is a conceptual design project aimed at demonstrating a design of a spanner that uses a form of 'mechanical rectifier mechanism' as described in one of the other projects to increase operational efficiency. The spanner design proposed here is different from conventional spanners in the sense that it uses two ratchet-pawl mechanism to increasing efficiency by operating in opposite direction as well. With a non-ratcheted spanner one has to detach it from a work-piece at the end of every half cycle, return to an approximately original angular position, attach the spanner to the work piece and fasten again. This is the end of one cycle. Single ratcheted spanner solves the problem of 'detaching' by introducing freewheeling in one of the directions using a ratchet-pawl mechanism. However, in the direction of freewheeling there is no transmission of torque to the 'jaws' and ultimately to the work piece. On an introduction of another ratchet-pawl mechanism, torque transmission could be done in both directions in a cycle. The spanner design has two shafts, one to hold and another to rotate. Rectification of rotational motion is achieved using a number of internal and external spur gears having intrinsic ratchet-pawl mechanism.

Overview

The centrifugal declutching mechanism is a design project aimed at creating a mechanism that works in the opposite way of a centrifugal clutch. The mechanism automatically declutches two power transmitting units as it approaches a certain angular velocity. The system consists of six main components, including a frame, mass, rope, pulley, spring, and friction plate. The springs default the clutch's state to 'engaged' when the system is stationary. The 'masses' by design have greater inertial mass than friction plates and other displacing components of the system under the action of centrifugal force. As the system's rotational speed increases, the centrifugal force kicks in and the friction plates lean inwards as a result of greater centrifugal force acting on 'Masses' (blue). A mathematical derivation was penned down to arrive at an equation that explains the amount of mass required to keep the system in equilibrium.