Hi, I'm Rashed.

I'm a software and research engineer at Samsung R&D Institute, Bangladesh. I received my B.S in Mechanical Engineering from Bangladesh University of Engineering and Technology in September 2017. I'm interested in research involving Robotics, Machine Learning and Material Science.

Almost all my academic interests are motivated by the process of deriving mathematical models of complex systems to solve problems. As an undergraduate, I materialized this pursuit by exploring two of the most fun avenues I was familiar with- competitive programming and robotics. Currently, much of my work revolves around building intelligent systems that can learn from both large datasets i.e an existing knowledge base as well as by interacting with the environment surrounding it.

In February 2018, I founded Bengali.Ai with Imtiaz and Samiul. We felt a lot of the challenges we faced with scarce and substandard datasets could be attenuated significantly through collaboration between researchers.

Things I like to do for fun include blogging, hiking and being terrible at the guitar.

What's New


Research


I am primarily interested in designing systems that can learn from past experiences. Be it for a pattern recognizing supervised learning task or for a reward optimizing reinforcement learning agent its only reference for inferring the target distribution is the reward/loss it seeks to optimize, and that is solely a function of the data/experiences encountered for that specific task.

Humans on the other hand can abstract away similarities between seemingly divergent sets of tasks and exploit previous experiences to augment the present task at hand. I like to describe this as connecting the dots by intelligently guessing some of the missing dots to construct the target image. One of my key intuitions in trying to imitate this ability in an agent is to learn representations that generalize well across different tasks. For example in the case of image recognition, data augmentation schemes are traditionally used to learn different representations of a single data point. However for humans this process is straightforward because they can infer a large portion of the entire distribution of data than can be generated from that one sample. This drastically reduces human reliance on amount of images needed to properly recognize scenes and objects. In the case of an agent that interacts with its environment, exploiting previous experience has already been shown to exhibit improved convergence rates.

Currently, my pursuits in this regard have been motivated by the following fields-

  • Computer Vision

    It can be argued that the central problem of vision is all but solved and the field has almost been saturated with superhuman performance of deep learning systems in tasks such as classification, detection and segmentation. But improved performance on even simple tasks still require large amounts of data, the majority of which are deemed redundant to the average human eye. I seek to augment the spatial hierarchy of features a system learns to better generalize with limited data. Features are learned by the network to best solve the vision problem at hand, can it be influenced to learn features that extend to potentially unseen data as well?

  • Representation Learning

    Encodings and latent spaces serve to automate the feature engineering process by learning rich lower dimensional representations of the data. Training on these representations permits the network to infer a much larger portion of the target distribution than it can from a sample of raw data. Finding representations that are functions of not just the data but of prior representations learned is one of my concerns in this field.

  • Reinforcement Learning

    While supervised learning systems perform well, they require a dataset that contain ground truth labels. These can be useful to solve problems where the optimal solution is readily available from humans. But in problems where the examples are possibly suboptimal, reinforcement learning can help find policies that best solve them. However, current SOTA RL agents take significantly long to train and require exponentially more interactions with the environment compared to the average human even at simple tasks. I’m interested in studying RL agents that can learn policies that help it intelligently interact with a new environment to converge faster to an optimal policy for a new task. Thus in a way learning a hierarchy of abstract to concrete policies that can extend to a wide range of tasks.

Publications


  • Sharif Amit Kamran, Ahmed Imtiaz Humayun, Samiul Alam, Rashed Mohammad Doha, Manash Kumar Mandal, Tahsin Reasat, Fuad Rahman

    AI Learns to Recognize Bengali Handwritten Digits: Bengali.AI Computer Vision Challenge 2018

    arXiv preprint arXiv:1810.04452. 2018 Oct 10.

  • Samiul Alam, Tahsin Reasat, Rashed Mohammad Doha, Ahmed Imtiaz Humayun

    NumtaDB - Assembled Bengali Handwritten Digits

    arXiv preprint arXiv:1806.02452. 2018 Jun 6.

Awards And Honors


HULT Prize Dubai, Regional Finalist

March 2017

Presented a low cost device to help financially rehabilitate refugees.

Bangladesh Education Board Talentpool Scholarship

March 2013

Among the top 30% students in the University who were selected for the scholarship.

BUET Technical Scholarship

Feburary 2013

Awarded the scholarship to fund education for 8 semesters.