Bleeding edge Deep Learning solutions

From NLP, image processing to anomaly detections, we have various experience in AI and Machine Learning.
These are some of our solutions.

QU

A chatbot prototype based on cutting edge Deep Learning technology that allows an algorithm to learn how to respond as a person. Self-trained only on subtitle dialogues (over 1M of them), the bot is able to understand basic human questions and respond accordingly. 

Lucid ART

Mobile app that paints pictures or videos into the style of famous artists (e.g. Picasso, Van Gogh, etc.). Scaled from 0 to 200,000 users in under 2 weeks, LUCID processes over 1M images per day. First to pioneer commercial artistic style transfer for videos.

Craze

An event app that uses Machine Learning for Natural Language classification. With over 3M events, craze can discover local event trends via anomaly detection algorithms and then recommend them to the users via proprietary Recommendation Engine.

Open-source projects

 

We strive to push the Deep Learning research forward, so that's why we constantly experiment with new models and deep architectures. Some of the work we've done can be found below.

 
https://github.com/randomrandom/deep-atrous-nerA Deep Atrous CNN architecture suitable for Named Entity Recognition on input with variable length, which achieves state of the art results. Up to 10x times faster during prediction time and state of th…

https://github.com/randomrandom/deep-atrous-ner

A Deep Atrous CNN architecture suitable for Named Entity Recognition on input with variable length, which achieves state of the art results. Up to 10x times faster during prediction time and state of the art results.

The architecture replaces the predominant LSTM-based architectures for Named Entity Recognition tasks. Instead it uses fully convolutional model with dilated convolutions, which are resolution preserving. 

Trained and benchmarked on the popular for the task CoNLL-2003 dataset. Pre-trained GloVe embeddings are used for the word vectors.

https://github.com/randomrandom/deep-atrous-cnn-sentimentA Deep Atrous CNN architecture suitable for text (sentiment) classification with variable length.The architecture substitutes the typical conv->pool->...->conv->pool->softmax ar…

https://github.com/randomrandom/deep-atrous-cnn-sentiment

A Deep Atrous CNN architecture suitable for text (sentiment) classification with variable length.

The architecture substitutes the typical conv->pool->...->conv->pool->softmax architectures, instead to speed up computations it uses atrous convolutions which are resolution preserving. Another great property of these type of networks is the short travel distance between the first and last words, where the path between them is bounded by C*log(d) steps, where C is a constant and d is the length of the input sequence.

Contact Us

 

Have a request or question? Shoot us an e-mail and we'll follow-up promptly.