You can also read my Research Statement for more details on my current and future research plans.

Machine Learning for Streaming Data
Sliding Bidirectional Recurrent Neural Networks

In communication networks, data streams are transported from one location to another. Part of my research focuses on developing novel machine learning techniques for learning and inference on streaming data. For example, I have developed a novel sequence detector based on neural networks for streaming data. This algorithm which is called sliding bidirectional recurrent neural network (BRNN), slides across the incoming data stream. The outputs of the sliding BRNNs are then combined using a dynamic programming approach, where the estimate of the previously observed data is constantly refined. The computational complexity of this approach increases linearly with memory length, making it suitable for applications where the data stream has a long memory. I demonstrate that sliding BRNNs can achieve a detection performance that is close to the optimal Viterbi detector for certain communication systems with lower computational complexity. I have received a provisional patent on the devised algorithm through the Stanford Office of Technology Licensing.

Design of Detection Algorithms Using Machine Learning
Neural Networks for Detection in Communication Systems

In some systems, such as chemical communication with multiple chemicals that react with each other, the underlying channel models are unknown and complex. Moreover, even when the channel models are known, many detection algorithms rely on the instantaneous channel state information (CSI), i.e., the instantaneous parameters of the model, for detection. Therefore, the instantaneous CSI needs to be estimated by transmitting and receiving a predesigned pilot sequence, which is known by the receiver. However, this estimation process is an overhead that would decrease the data transmission rate. Moreover, the accuracy of the estimation may also affect the performance of the detection algorithm. I investigate how a data-driven approach can be used to train NN detectors that dynamically adapt to changing channel conditions without any knowledge of the underlying channel models. For example, I have shown that if the training dataset is rich enough with sample transmissions under various channel conditions, the trained NN detector performs well in changing channel conditions without the CSI, which results in higher data rates.

Joint Source-Channel Coding Using Deep Neural Networks
Joint Source-Channel Neural Network Encoder and Decoder

In digital communications, data transmission typically entails source coding and channel coding. In source coding, the data is mapped to a sequence of symbols where the sequence length is optimized. Many communication systems may benefit from designing the source/channel codes jointly. I develop novel joint source-channel encoders and decoders using neural networks for three classes of data: text, images, and video. An important motivation for using deep learning for the design is that in many applications instead of recovering the exact transmitted data, we are interested in recovering some information of interest from the data. For example for text data, instead of recovering the exact sentence at the receiver, we are interested in recovering the semantic information such as facts or imperatives of the sentence. For example, the phrase "the car stopped" and "the automobile stopped" convey the same information. For image and video data, the information of interest could be objects or moving objects, and any image that preserves this information is considered as an error-free output by the decoder. Since the encoder extracts and encodes the information of interest from the data, a fewer number of bits are used for transmission, which results in higher data rates.

Capacity Analysis and Design of Molecular Communication Channels
Molecular Communication

In molecular communication, the transmitter releases chemicals or molecules and encodes information on some property of these signals such as their type/structure, timing, or concentration. The signal then propagates through the medium between the transmitter and the receiver, via means such as diffusion, until it arrives at the receiver where the signal is detected and the information is decoded. This is a promising new field for networking devices with dimensions much less than a millimeter. One of the main motivations behind this technology is in-body communication. For example, bio-sensors, such as synthetic biological devices, can constantly monitor the individual for different biomarkers for diseases and chemically transmit their measurement to a device on or under the skin, enabling a real-time health monitoring system. Another motivation is on-chip communications where different components such as a DNA storage module and a molecular processing unit are connected via molecular signals. As part of my research, I model different molecular communication channels and analyze their capacity limits. The Shannon capacity of a communication channel characterizes the maximum theoretical achievable data rates for the channel. Understanding this theoretical limit provides intuitions that are useful for system design, and provides a new lens through which to study biology as many biological processes are regulated using chemical communication. Another aspect of my research focuses on receiver design, where I derive the optimal sequence detector for molecular communication systems under different modulations. The optimal detectors for molecular communication can be computationally complex because of intersymbol interference (ISI). In these scenarios, I also use a data-driven approach based on machine learning for designing the receiver or derive computationally efficient heuristic algorithms.

Optimal Design for Molecular Communication on Bio-chips
On-Chip Molecular Communication

In communication systems, optimal design improves the performance of the system and ensures that the best possible data rate and highest reliability is achieved. One aspect of my research has focused on optimal design for molecular communication in bio-chips, where different components within a chip, such as a DNA storage unit and a molecular processing unit, are connected using chemical signaling. The bio-chips use chemical and molecular properties and interactions for storage and processing of information. They are different from silicon-based chips that use electrical properties and electrical signals. In bio-chips, using molecular communication is simpler than converting chemical properties into electric signals and then converting them back to chemical properties. I use information theory, modeling, and optimization to provide design guidelines for optimal placement of the components on the bio-chip such that the data rate between the components are maximized

Experimental Demonstrators for Molecular Communication
Molecular Communication Experimental Platform

Since the field of molecular communication is still in its infancy, experimentally validated standardized models do not yet exist. Some of my research has been focused on bridging the big gap between theory and experimentation by developing novel experimental platforms for molecular communication systems. I have built the world’s first molecular communication platform. The goal of this system was to reliably transmit short messages over a distance of a few meters. This system used an electronically controlled spray as the transmitter, a metal oxide sensor that detects the concentration of alcohol as the receiver, and aerosolized alcohol as the information-carrying chemical. This work attracted media coverage by The Economist, The Royal Society of Chemistry, and the Wall Street Journal, as well as other national and local newspapers. Now I am working to develop new experimental platforms that simulate real chemical communication environments for two important applications of molecular communication: in-body and on-chip communication. In particular, my goal is to build an artificial network of veins and arteries. Using the platform, one can simulate a device injecting different chemicals such as drugs, sugars, or proteins into the bloodstream. The resulting chemical signal propagation can be measured by different detectors. Another area that I intend to explore is on-chip communication, especially for DNA storage and molecular computation.