Over the last few decades, research on handwriting recognition has made impressive progress. The research and development on handwritten word recognition are to a large degree motivated by many application areas, such as automated postal address, data acquisition in banks, text-voice conversion, security, etc. As the prices of scanners, computers and handwriting-input devices are falling steadily, we have seen an increased demand for handwriting recognition systems and software packages. Few commercial handwriting recognition systems are now available in the market. Current commercial systems have an impressive performance in recognizing machine-printed characters and neatly written texts. For instance, Xerox in the U.S. has developed TextBridge for converting hardcopy documents into electronic document files. Notwithstanding the impressive progress, there is still a significant performance gap between the human and the machine in recognizing off-line unconstrained hand- written words. The difficulties encountered in recognizing unconstrained handwritings are mainly caused by huge variations in writing styles and the overlapping and the interconnection of neighboring characters. Furthermore, many applications demand very high recognition accuracy and reliability. Sadly, in the banking sector, although ATMs and networked banking systems being widely available, many transactions are still carried out in the form of cheques. Google’s Tesseract, despite being developed by the greatest IT beast, its output will have very poor quality if the input images are not preprocessed to suit it. This motivates us to research and develop a handwritten sentence recognizer with much improved accuracy.

Stock prices are volatile. They are hard to predict based only on some expertise through previous trends. Stock price fluctuation represents the current market trends and business growth among other factors that could be considered to sell or buy stocks. The daily news represents factual information about the companies and affect the stock market. Hence, we wish to create a model that will use news articles along with the historic trend of closing price to predict the change in stock prices of five companies namely TCS, LT, Infosys, Yes Bank and Tata Steel.

Games and artificial intelligence have a long history together. Much research on Artificial Intelligence in games is concerned with constructing agents for playing games, with or without a learning component. In this project, we propose to use Deep Q- learning and online learning Multi Layer Perceptron (MLP) for learning to control the game agent in T-Rex, the game embedded in Chrome offline mode.

In this project we have built a simple textual chatbot , which is a sequence to sequence model with attention decoder. An encoder is an utterance by the user and the decoder is the response to that utterance. The bot requires dataset for training itself. We have used the Cornell Movie-Dialogs Corpus for training our bot. The training dataset has been pre-processed as it contains some data which is useless for the model.

Classification of birds by their sound has many potential ap- plications in ecological surveillance, conservation monitoring, taxonomy, and vocal communication studies. Using audio recordings rather than pictures is justifiable since bird calls and songs have proven to be easier to collect and to discriminate better between species. But for it to be useful for the layman, it should be scalable, should work in noisy environments and on different lengths.

Chord Recognition is the automated process of assigning mu- sical chordsto segments of a music piece.It is an important task in the analysis of western music and music transcription in general, and it can contribute to applications such as key detection, structural segmentation, music similarity measures, and other semantic analysis tasks. The first step of the chord recognition (CR) is the extraction of a meaningful descriptor that enhances the contribution of every note at each instant. One of the main strategies of doing Chord Recognition is based on the comparison of vectors of the extracted descriptor with a set of chords templates by means of a distance measure. This method, due to the interaction of many instruments with dif- ferent timbral characteristics, is very influenced by the noise of the descriptor itself and, for this reason, can be ineffective. In this project, we worked on a method for Chord Recognition based on machine learning techniques. These techniques are designed to automatically learn the complicated relations that link an input observation to the corresponding class making use of probabilistic theory.

Recommendation systems represent user preferences for the purpose of suggesting items to purchase or examine. They have become fundamental applications in electronic commerce and information access, providing suggestions that effectively prune large information spaces so that users are directed towards those items that best meet their needs and preferences. A variety of techniques have been proposed for perform- ing the task of recommendation, including content-based, collaborative, knowledge-based and other techniques. While collaborative filtering makes automatic predictions (filtering) about the interests of a user by collecting information regarding preferences or interests of other users (collaborating), content based filtering uses information about the description and attributes of the items that the user has previously consumed to model user’s preferences. However, Recent research has demonstrated that a hybrid of the two models could be more effective than the pure approaches since it combines the benefits of the two, overcoming some of their shortcomings. Our project aims at building one such hybrid recommendation system to suggest news articles to readers based on their prior interactions with previous reads in the form of views, likes, comments, bookmarks and followings.

Many times we find ourselves unable to recall the most ap- propriate word for describing the idea/concept we have in our minds. This might just be a recall problem or might be due to our lack on knowledge in that language. And it turns out to be a very common problem for people who produce language. In such cases, dictionaries might not be the most perfect solution as it is easy to find the meaning of a particular word but it is almost impossible to find a word given its meaning. Hence, to address this issue we propose the idea of building a reverse dictionary.

In this project we worked on how to recommend a user new movie based on its previously watched movies and ratings it had given to watched movies. Movie recommendation is used in many areas like Netflix(https://www.netflix.com/in/) uses it to recommend movies to its user, also Amazon Prime(https://www.primevideo.com/) for movie recommenda- tion. It gives user a complete new experience of watching movies. Thus it motivated us to do project on this topic.

Neural Style Transfer is basically transferring of the style from one image to the another image called content image.We have done it in two stages which are explained below.
Stage 1 : In this we implemented the naive method which transfers the style without using neural network but with the help of image clustering, feature extraction and blending of images.
Stage 2 : We style our content image with the help of neural network based on minimizing loss functions for content image and style image.

Facial Expression Recognition is something that humans can vividly do on a daily basis. On a personal level, if you know someone long enough, you know how they are feeling even without uttering a word - a true friend/partner/sibling/family member. There are a number of applications for this technol- ogy. Applications of Emotion Recognition are spread across different fields like Medicine, E-learning, monitoring, Market- ing, Entertainment and Law. One of the application is "ATM not dispensing money when the user is scared". This is a challenge for a computer to perform with accuracy. There are many AI methods like neural networks, SVMs which can help in solving this problem. Our motivation comes from the Kaggle fer2013 challenge. We look into a Naive Bayes` approach to solve this problem. Although this doesn‘t give a high accuracy for commercial use, it gives an idea of how tough it is to build an AI method with great performance standards. We aim to classify an image into any one of the seven expreesions (Anger, Disgust, Fear, Happiness, Sadness, Surprise, Fear).