I love tackling tough research problems and I am always open to new challenges. This has led me to work on problems in physics, biology, and machine learning. Please check out the following sections for more details of my past projects.

Neural Boltzmann Machines and Digital Twin Generators

Alex H. Lang, Anton D. Loukianov, Charles K Fisher


Conditional generative models are capable of using contextual information as input to create new imaginative outputs. Conditional Restricted Boltzmann Machines (CRBMs) are one class of conditional generative models that have proven to be especially adept at modeling noisy discrete or continuous data, but the lack of expressivity in CRBMs have limited their widespread adoption. We introduced Neural Boltzmann Machines (NBMs) which generalize CRBMs by converting each of the CRBM parameters to their own neural networks that are allowed to be functions of the conditional inputs. NBMs are highly flexible conditional generative models that can be trained via stochastic gradient descent to approximately maximize the log-likelihood of the data.

We demonstrate the utility of Neural Boltzmann Machines by using them to create digital twins of patients that can describe the evolution of their health over time. Digital twins have the potential to revolutionize medicine by enabling individual-level computer simulations of human health, which can be used to conduct more efficient clinical trials or to recommend personalized treatment options. Due to the overwhelming complexity of human biology, machine learning approaches that leverage large datasets of historical patients’ longitudinal health records to generate patients’ digital twins are more tractable than potential mechanistic models. We invented a neural network architecture, Digital Twin Generators or DTGs, that can create digital twins of individual patients and show that the same neural network architecture can be trained to generate accurate digital twins for patients across 13 different indications simply by changing the training set and tuning hyperparameters. By introducing a general purpose architecture, we aim to unlock the ability to scale machine learning approaches to larger datasets and across more indications so that a digital twin could be created for any patient in the world.

Perception for Autonomous Vehicles

Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, Oscar Beijbom

Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, Oscar Beijbom

Sourabh Vora, Alex H. Lang, Bassam Helou, Oscar Beijbom

I worked at Motional where our goal was to create the software stack to run a self driving taxi. Specifically, the machine learning team’s charter was to tackle the problems that are too tough to model explicitly. Therefore, our method of choice is deep learning, and we usually work closely with the raw sensor data. Our cars have a 360 degree coverage through multiple lidars, cameras, and radars (check out nuScenes for our actual data!), but of these sensors  lidar is the most important. Lidar is a laser ranging sensor that provides sparse, yet accurate, points in the 3D world. These point clouds are the key inputs for 3D object detection since they allow precise localization in the real world.

The ideal deep learning model would incorporate all sensor modalities (lidar, cameras, and radar), but a first step is to separately model each sensor. Images are relatively easy since a multitude of methods exist in the literature, so our research has focused on how to do object detector for lidar and radar. I’m going to keep focusing on lidar since it is the main sensor, but everything that follows about PointPillars could equally well be used on radar after a few minor changes.

So, what was the state of the art for lidar only object detection when we started our research? There were two main schools of thought that are best represented by PIXOR and VoxelNet. The fundamental difference is how to represent the sparse lidar point cloud. One school of thought (PIXOR, MV3D, …) is to create a set of fixed, hand crafted features. The other school (PointNet, Frustum PointNet, VoxelNet, SECOND) believes in end to end learning and just lets the network learn directly from the point cloud. From a performance and engineering perspective, end to end learning is always better because (1) the network should always be able to match (and usually far exceed) fixed encodings and (2) we let the network do the hard work of finding the encoder, rather than having to devote engineer’s time to discover the right encoding. So we should all do end to end learning!

But there is always a catch. The issue with VoxelNet is that it is too slow to run in realtime. The central problem is that they chose to do end to end learning on voxels. This forces them to use 3D convolutions which are extremely slow. In contrast, PIXOR can just use 2D convolutions which are well optimized for GPU computing. If only there was a way to blend the performance of end to end learning with the speed of fixed encoders.

It turns out, we found a method to do so: PointPillars. The fundamental realization was that pillars are the best representation. A pillar is a vertical column that can extend infinitely up and down. By learning end to end on pillars, we achieved state of the art detection performance on the KITTI leaderboard at blazing fast speeds (60 to >100 Hz), for a 2-4 fold improvement in runtime.

So where do we go from here? The next step is to work on sensor fusion. Camera and lidar are important sensor modalities for robotics in general and self-driving cars in particular. The sensors provide complementary information offering an opportunity for tight sensor-fusion. Surprisingly, when we started lidar-only methods outperformed fusion methods on the main benchmark datasets, suggesting a gap in the literature. We proposed PointPainting, a sequential fusion method to fill this gap. PointPainting works by projecting lidar points into the output of an image-only semantic segmentation network and appending the class scores to each point. The appended (painted) point cloud can then be fed to any lidar-only method. Experiments show large improvements on three different state-of-the art methods, Point-RCNN, VoxelNet and PointPillars on the KITTI and nuScenes datasets. Additionally, we demonstrated that the latency of PointPainting can be minimized through pipeline, making this a practical method for realtime applications.

Epigentic Landscapes and Cellular Reprogramming

Alex H. Lang, Hu Li, James J. Collins, and Pankaj Mehta

Sai Teja Pusuluri*, Alex H. Lang* (co-first authors), Pankaj Mehta, and Horacio E. Castillo

Sai Teja Pusuluri, Alex H. Lang, Pankaj Mehta, and Horacio E. Castillo

I think one of the most surprising recent scientific experiments was Takahashi and Yamanaka's reprogramming of a skin cell to something resembling an embryonic stem cell (ESC), dubbed induced pluripotent stem cells (iPSCs). Following this groundbreaking experiment, other reprogramming protocols have been found so now scientists can switch between a variety of cell types such as ESC, skin, liver, neurons, and cardiomyocytes (heart muscle). This has already revolutionized the understanding of biology and could drastically change the future of medicine. 

I am very interested in the answer to two different questions. First, what does it mean to be a cell type if minor perturbations (manipulating only a few genes, specifically 4 transcription factors out of thousands) can result in a completely different cell type? Second, how do we rationally design new reprogramming protocols? If we could make any desired cell type from an individual's own skin, this could eliminate the need for organ transplants.

So why am I, as a physicist, researching this? Well, biologists currently use Waddington's Landscape as an analogy to describe cellular development and cellular reprogramming. As a physicist, the word landscape has a specific mathematical meaning, so my advisor Pankaj Mehta thought it would be a great first project for us to work together to see if we could make the analogy more precise. This project took off and now, several years later, our model describes the biology much better than we had any right to expect. 

We developed a model based on spin-glass physics to construct Waddington's landscape. Using real biological data (publically available microarrays), and assuming that the landscape is well described by a Hopfield neural network, we can construct a complete (i.e. all cell types for which we have data) Waddington epigenetic landscape. We believe that, using our model, we can rationally design cellular reprogramming protocols but are still searching for the best experimental test of our model with an awesome group of experimentalists at the BU CReM.

In collaboration with Sai Teja Pusuluri and Horacio E. Castillo, we are working to answer interesting questions on how to control the dynamics of cellular reprogramming. First, we reanalyzed time-series data on cellular reprogramming from differentiated cell types to induced pluripotent stem cells (iPSCs) and showed that gene expression dynamics during reprogramming follow a simple one-dimensional reaction coordinate. Interestingly, this reaction coordinate is independent of both the time it takes to reach the iPSC state as well as the details of experimental protocol used and therefore seems to be a universal characteristic of reprogramming. In a second paper, we examine the sizes of the basins of attractions and show that they are highly depending on the specific correlation structure between cell types. We are currently working on ways to exploit these findings to design new reprogramming methods.

Thermodynamics of Cellular Information Processing

Alex H. Lang, Charles K. Fisher, Thierry Mora, Pankaj Mehta.

Pankaj Mehta, Alex H. Lang, David J Schwab.

An important task of cells is to perform complex computations in response to external signals. Intricate networks are required to sense and process signals, and since cells are inherently non-equilibrium systems, these networks naturally consume energy. Since there is a deep connection between thermodynamics, computation, and information (see Maxwell’s Demon and Landauer’s principle for examples), a natural question is what constraints does thermodynamics place on statistical estimation and learning. In our research, we modeled a single chemical receptor and established a fundamental relationship between the energy consumption and statistical accuracy. Recent advances in non-equilibrium thermodynamics have yet to be fully applied to biological problems, so this remains an exciting field with a multitude of potential problems.

Additionally, I was a part of a review paper which tried to bridge the gap between the thermodynamics of information processing and synthetic biology. Current experiments in synthetic biology are performing complex computations that are beginning to implement unique cellular information processing tasks. However, these biological circuits are still subject to the same general principles laid out by Landauer and others. We summarized recent theoretical work and found that energy consumption in cellular circuits usually serves five basic purposes: (1) increasing specificity, (2) manipulating dynamics, (3) reducing variability, (4) amplifying signal, and (5) erasing memory.

Bioinformatics and Biological Classification

A.A. Wilson, L. Ying, M. Liesa, C.P. Segeritz, J.A. Mills, S.S. Shen, J.C. Jean, G.C. Lonza, Alex H. Lang, J. Nazaire, A.C. Gower, F.J. Mueller, P. Mehta, A. Ordonez, D. Lomas, L. Vallier, G.J. Murphy, G. Mostoslavsky, A. Spira, O. Shirihai, M. Ramirez, P. Gadue, D.N. Kotton.

K. Dame, S. Cincotta, Alex H. Lang, R. Sanghrajka, R. Zhang, L. Kwok, T. Wilson, S. Monti, P. Mehta, D.N. Kotton, L. Ikonomou.

Biology is now rapidly producing large-scale quantitative data that needs to be analyzed and bioinformatics provides the necessary tools and techniques to handle this big data. In collaboration with researchers at Boston University’s Center for Regenerative Medicine (CREM), I have performed standard statistical analysis of microarrays. 

Additionally, we are working on developing new classification methods. Inspired by our work on epigenetic landscapes, we found a new technique based on linear algebra projections that provides an accurate and sensitive tool for discerning differences amongst biological samples. Our analysis method was used by the CREM to provide additional support for their technique to create thyroid progenitors.

Noise of microRNA Regulation and the ceRNA Hypothesis

Javad Noorbakhsh, Alex H. Lang, and Pankaj Mehta.

MicroRNAs are short sequences of RNA that can regulate gene expression by binding to mRNA. The ceRNA hypothesis recently proposed that microRNAs play an important role in gene regulation networks by inducing correlations between mRNAs competing for the same pool of microRNAs.

We mathematically modeled microRNAs using a set of stochastic differential equations. We found that microRNA networks are insensitive to many of the underlying biological details, but are responsive to the binding strength between the microRNA and mRNA. If the ceRNA hypothesis is true, all the microRNA and mRNA interaction rates need to be finely tuned to a similar value. Our work suggests the ceRNA hypothesis is unlikely to play a major role in gene regulation, but specific microRNAs can play an important, albeit limited, role in regulation.

COMETS: Flux Balance Analysis and Spatial Patterning

W.R. Harcombe, W. Riehl, I. Dukovski, B.R. Granger, A. Betts, Alex H. Lang, G. Bonilla, A. Kar, N. Leiby, P. Mehta, C.J. Marx, and D. Segre.

Flux balance analysis (FBA) is a technique to mathematically model metabolic networks. Daniel Segre and his group developed software to implement FBA in extended spatial patterns. This will allow labs to investigate the interactions between various bacteria strains under multiple growing conditions and patterns, mimicking bacterial biofilms. Listen to Daniel explain the project in his video. We helped implement diffusion in this program.

Disentanglement and Decoherence of Qubits

Dong Zhou, Alex H. Lang, and Robert Joynt

Amrit De, Alex H. Lang, Dong Zhou, and Robert Joynt

Quantum computers are a new type of computer built out of quantum based bits, dubbed qubits. This allows large computational speedup for certain types of problems, such as integer factorization. However, noise leads to the quantum qubits becoming classical, which is known as decoherence. In our research, we studied the effect of random telegraph noise and found that certain interactions could delay decoherence.