jianghui

Hui Jiang

By |

My research focuses on statistical genomics and computational statistics. I am interested in developing statistical and computational methods for the analysis of large-scale biological datasets generated by modern high-throughput technologies such as next-generation sequencing. I implemented many of the methods that I developed as software tools and packages to be used by the research community. I am also interested in developing efficient algorithms and methods that deal with computational problems arising from statistics genomics. I have worked on and am working on problems including efficient algorithms for resampling-based hypothesis testing, penalized modeling and optimization algorithms for model fitting, as well as computational methods for density estimation and machine learning.

Nishil Talati

Nishil Talati

By |

I am a research faculty at the Computer Science and Engineering department at University of Michigan. I work with a group of talented PhD students on computer architecture, compiler techniques, and software engineering. My group focuses on developing novel software and hardware solutions to optimize large-scale data intensive worklods (e.g., graph traversals).

I earned my PhD degree in CSE from University of Michigan, Ann Arbor, USA, master’s degree in EE from Technion – Israel Institute of Technology, Haifa, Israel, and an undergraduate degree in EEE from BITS Pilani, Goa Campus, Goa, India.

Samet Oymak

Samet Oymak

By |

I am interested in principled approaches to machine learning with focus on data-driven decision making, deep learning foundations, and heterogeneous data. My research integrates optimization methods (specifically convex and first-order) and statistical learning theory to design efficient algorithms/architectures that address these data-science problems.

Additional Information

How did you end up where you are today?

I obtained my PhD degree from Caltech in 2015 where I received a Charles Wilts Prize for the best departmental thesis. During postdoc, I was at UC Berkeley as a Simons Fellow. After spending few years in industry, I joined UC Riverside where I received NSF CAREER and Google Research Scholar Awards. Starting Fall 2023, I will be joining EECS department at U-M.

An interesting fact: The elegance of mathematics has always amazed me and led me to participate in International Math Olympiad.
Also, my Erdos number is 3!


Accomplishments and Awards

Anne Draelos

Anne Draelos

By |

My lab’s research focuses on understanding computation in large-scale neural circuits through adaptive perturbations and real-time inference. We develop statistical machine learning algorithms to adaptively build models of neural and behavioral data online, and use them for understanding the mapping between multidimensional neural stimulations and complex behavioral outcomes. We emphasize data-driven Bayesian approaches suited for real-time prediction of latent neural and behavioral dynamics.

Additional Information

How did you end up where you are today?

My academic training began in both physics and computer science. My masters work in electrical and computer engineering focused on modeling interconnected resistive systems, and I did my doctoral work in experimental condensed matter physics, focused on the interplay between various quantum phenomena in two-dimensional materials. I am most interested in methods for dissecting complex networked systems, and I ultimately decided to make the switch to neuroscience to study arguably the most complex network around: the brain.v

My postdoctoral work focused on automated neural circuit dissection in larval zebrafish, where I designed both mathematical and software tools for experimental integration of machine learning algorithms. As an Assistant Professor, my lab is focused on leveraging increasingly rich behavioral data (in e.g., mice, monkeys) alongside neural data in streaming contexts to causally relate perturbations of neural function and resultant outcomes in latent behavior spaces. My ultimate goal is to bridge the gap between algorithm and implementation, using adaptive methods as a new design language for neuroscience experiments.

What makes you excited about your data science and AI research?

One of the reasons I decided to move into neuroscience as my application domain was the relatively recent data explosion in the field. We can now record from tens of thousands up to a million neurons simultaneously, at quite high-resolution in time. The field is also increasingly focused on naturalistic and effectively non-repeatable events as behavioral metrics. Both of these items make for incredibly complex and rich problems for neuroscience, which is why I think data science and AI have amazingly interesting roles to play there.


Accomplishments and Awards

Talia Moore

Talia Moore

By |

My research involves discovery, modeling, and design for comparative biomechanics and bio-inspired robotics. I develop new information-based methods to analyze the behavior and locomotion of animals in natural environments and use laboratory-based experiments to build detailed models of how animal motion is generated by muscles, tendons, and bones. I’ve developed machine learning methods to automatically segment non-standard animals in photographs for taxonomically-broad phylogenetic comparative analyses of color pattern and behavior. I also use Finite Element Analysis to understand how the evolution of 3-dimensional shape in animal bones and teeth have adapted to a variety of ecological situations, such as novel substrates or diet.

Because FEA and other biomechanical methods are so computationally intensive, I’ve adopted a surrogate model approach that uses Bayesian Optimization to perform stimulus selection and make accurate predictions of sample performance with the minimum number of model datapoints. I’ve also begun applying this surrogate model approach to explore the design space of bio-inspired grippers made from dielectric fluid actuators. Data-driven modeling also informs the design rigid legged robots in my lab, which we will use to test hypotheses regarding how limb shape affects overall locomotion.

Finite element modeling of snake fangs helps us understand how fang shape can be adapted for different types of loading conditions, and therefore prey types. We use Bayesian Optimization to select which species to analyze for our surrogate model, which minimizes the computation time while maximizing prediction accuracy.

Additional Information

What are some of your most interesting projects?

Jerboas are bipedal desert rodents that hop erratically with zig-zag trajectories when they are escaping from predators (some describe them as “ricochetal”). These animals are about the size of your fist, but they can jump over 3 feet straight up in the air or forwards. Little was known about their locomotion, and the majority of biomechanical locomotion research was performed in lab environments on treadmills. However, training an animal to run on a treadmill reduces the variability in direction and speed that the animals would need to survive in the wild. So I went out into the desert and filmed these animals jumping and running on their natural substrates.

To understand how these trajectories might confound predators, I measured the unpredictability (or entropy) and found that bipedal desert rodents are much less predictable than quadrupedal desert rodents. Then, taking a closer look using high-speed video cameras, I determined that they have at least 5 distinct bipedal gaits. Using both kinematic and dynamic data, I built a modified Spring-Loaded Inverted Pendulum model with a torsional spring to control the neutral leg swing angle. I then performed a numerical search using a continuing approach to test neighboring parameter values for viability to discover bifurcations in gait structure. I found that by decoupling the neutral leg swing angle between left and right legs, the model was capable of transitioning between gaits across the entire speed range, just as the real jerboas do. This research will be used to inform the design of controllers for legged robots to switch gaits smoothly across a wide range of speeds.

How did you end up where you are today?

I started off not knowing at all what I wanted to do, but enjoying martial arts. A friend told me to check out the biomechanics class in my last year in undergrad at UC Berkeley, and I was hooked! I joined the lab immediately and worked there for two years after graduating. I wanted to learn everything, so I worked on a different project with a different graduate student every day of the week. During that time, I worked with cockroaches, geckos, iguanas, and agama lizards, learning about how they generate and control their motions. I was also lucky to get hands-on experience with designing and building bio-inspired robots and using them to test biological hypotheses to reveal fundamental principles of animal locomotion.

After that, I went on to study biomechanics and evolutionary biology at Harvard, where I was introduced to jerboas for the first time. They are such strange and wonderful creatures that I knew I wanted to study them for the rest of my life. I came to UM in 2015 and worked as a postdoc in Ecology and Evolutionary Biology, where I developed a new modular ethogram system to analyze snake anti-predator behaviors and design snake-mimicking soft robots. Then I became a Research Scientist in the Robotics Institute, followed by being hired as an Assistant Professor in Mechanical Engineering in January of 2021. Now I am appointed in both Robotics and Mechanical Engineering and have affiliations with Ecology and Evolutionary Biology and the Museum of Zoology.

What is the most significant scientific contribution you would like to make?

I really want to build tools to bridge the gap between biology and engineering. So many biological questions are constrained by the technology we have available. By forming connections between these fields, I have already facilitated more quantitative study of non-steady-state locomotion in natural environments. There is also a big gap between what motions animals and robots are capable of performing. I hope to learn strategies from animals and design robots to succeed in unstructured and complex environments.

What makes you excited about your data science and AI research?

I’m extremely excited that data science is making it possible to analyze the types of large datasets that we can collect from animals. There is no limit to the amount of data you can collect about animal locomotion, behavior, appearance, or structure, and the types of studies that used to take decades can now be done in a semester thanks to data science and AI. This makes it possible to integrate information from multiple different data streams and understand more complex relationships between animals, their environments, and how these relationships change through space and time.

What are 1-3 interesting facts about yourself?

I’ve done fieldwork in Malaysia, China, Australia, California, the Bahamas and Peru. I think it’s incredibly important to examine animals in their natural habitats, because our assumptions about their behaviors might be totally wrong if we only see them in zoos or labs.
The first sentence I try to learn in every language is “Where is the bathroom?”
My two rules for fieldwork are: 1) Never stop moving and 2) Never sit down or lean on anything.

Vineet Kamat

Vineet Kamat

By |

My group conducts research in automation and robotics to improve work processes in the construction, operation, and maintenance of civil infrastructure and the built environment. Our research has developed several licensable technologies that include visualization, perception, and modeling techniques to help on-site construction robots with autonomous decision making. We are particularly interested in exploring new methods for enabling collaborative work strategies for human-robot teams jointly performing field construction work. In addition, we are also interested in exploring methods to integrate data to support semi-autonomous mobility for people with physical disabilities in the urban built environment.

Data-Driven Co-Robotic Field Construction Work

Michael Cianfrocco

Michael Cianfrocco

By |

Dr. Michael Cianfrocco uses cryo-electron microscopy (cryo-EM) to determine protein structures to understand how nanometer-sized molecular machines work. While a powerful technique, cryo-EM data collection and subsequent image analysis remain bespoke, clunky, and heuristic. Dr. Cianfrocco is coupling his 16+ years of experience with artificial intelligence to automate data collection and processing by capturing human expertise into AI-based algorithms. Recently, his laboratory implemented reinforcement learning to guide cryo-electron microscopes for data collection [1, 2]. This work combined real-world datasets and Dr. Cianfrocco’s expertise with AI-driven optimization algorithms to find the ‘best’ areas of cryo-EM samples for data collection.

cryoRL Distributed Data Collection process diagram

Human users must curate and select areas for subsequent analysis after data collection. Subjective decisions guide how to process the single particles and determine what constitutes ‘good’ data. To automate subsequent preprocessing, Dr. Cianfrocco’s lab built the first AI-backed data preprocessing in cryo-EM by training CNNs to recognize ‘good’ and ‘bad’ cryo-EM data [3]. This work enabled fully-automated cryo-EM data preprocessing, the first step in the processing pipeline of cryo-EM data. In the future, Dr. Cianfrocco wants to continue improving cryo-EM workflows to make them robust and automated, eventually surpassing human experts in the ability of algorithms to collect and analyze cryo-EM data. 1. Fan Q*, Li Y*, et al. “CryoRL: Reinforcement Learning Enables Efficient Cryo-EM Data Collection.” arXiv preprint arXiv:2204.07543 (2022). 2. Li Y*, Fan Q*, Optimized path planning surpasses human efficiency in cryo-EM imaging. bioRxiv 2022.06.17.496614 (2022). 3. Li Y, High-Throughput Cryo-EM Enabled by User-Free Preprocessing Routines. Structure. 2020 Jul 7;28(7):858-869.e3.

Sally Oey

Sally Oey

By |

Sally Oey’s group is studying massive star populations and the escape of ionizing radiation from starburst galaxies and super star clusters. The group is at the forefront of establishing a new paradigm for massive-star feedback, where superwinds from compact young star clusters fail to launch. Members have used numerical simulations and image processing techniques to investigate such conditions for allowing ionizing radiation to penetrate the dense gas in star-forming clouds and the interstellar medium in “green pea” galaxies and resolved nearby starbursts. The ionizing radiation may originate from massive binaries and their products, thus group members are carrying out data mining of observational surveys and binary population synthesis models to study how binarity manifests in stellar populations.

Leopoldo Pando Zayas

Leopoldo Pando Zayas

By |

My main research interest is in quantum gravity. Various aspects of quantum information and quantum chaotic systems have proven to be essential in recent developments.

Hun-Seok Kim

By | | No Comments

Hun-Seok Kim is an associate professor at the University of Michigan, Ann Arbor. His research focuses on system analysis, novel algorithms, and efficient VLSI architectures for low-power/high-performance wireless communication, signal processing, computer vision, and machine learning systems.


HTNN (Heterogeneous Transform Domains Neural Network) is a new class of transform domain deep neural networks, where convolution operations are replaced by element-wise multiplications in heterogeneous transform domains. To reduce the network complexity, this framework learns sparse-orthogonal weights in heterogeneous transform domains co-optimized with a hardware-efficient accelerator architecture to minimize the overhead of handling sparse weights. Furthermore, sparse-orthogonal weights are non-uniformly quantized with canonical-signed-digit (CSD) representations to substitute multiplications with simpler additions. The proposed approach reduces the complexity by a factor of 4.9 – 6.8 × without compromising the DNN accuracy compared to equivalent CNNs that employ sparse (pruned) weights.