Ankit Gupta's repositories
injecting_numeracy
The accompanying code for "Injecting Numerical Reasoning Skills into Language Models" (Mor Geva*, Ankit Gupta* and Jonathan Berant, ACL 2020).
top_k_attention
The accompanying code for "Memory-efficient Transformers via Top-k Attention" (Ankit Gupta, Guy Dar, Shaya Goodman, David Ciprut, Jonathan Berant. SustaiNLP 2021).
AI_Bot_for_playing_PacMan
A deep convolutional neural net (CNN) for playing the well-known arcade game ‘Ms. Pac-Man’ trained using Q-learning (reinforcement learning). Game data is obtained using the ‘OpenAI Gym’. The neural net significantly outperforms all tried uniform-random strategies. We do not inject any prior knowledge in the Q-nets. For instance, we do not explicitly use the distance from the ghosts - the system only has access to the game frames and has to learn everything by itself.
English_Vietnamese_Translator
A recurrent neural net (RNN) based Neural Machine Translator for translating English to Vietnamese built using an Encoder-Decoder model. Training was done using a small-scale parallel corpus of TED talks.
Sentiment_Analysis
A deep recurrent neural network (RNN) based classifier trained using the IMDb reviews dataset for detecting +ve/-ve sentiment in text. Text is encoded using trainable word embeddings and we get 90% test accuracy. Classifier can then be used to determine the sentiment score on Twitter data collected using the Twitter API.
Creative_Text_Generator
A deep recurrent neural net (RNN) for both character-level and word-level language modeling that, given any starting context, generates creative text depending on the training data. Training was done using OANC corpus. Even with probabilistic sampling of character sequences, the RNN almost always outputs valid dictionary words making meaningful sentences.
Snapchat_Filter
A Snapchat-like application based on the standard pipeline of 1) Face Detection using Histogram of Oriented Gradients (HOG) descriptor, 2) Facial Landmark Estimation using Kazemi-Sullivan and finally, 3) adding overlays/lenses (like funny glasses, etc) to the frames using the previously estimated landmarks.
value_aware_attn
The accompanying code for "Value-aware Approximate Attention" (Ankit Gupta and Jonathan Berant, EMNLP 2021).