2020 CMP Academic Project Proposals

https://rstudio-pubs-static.s3.amazonaws.com/309220_27a809ae27b3436d8940...

https://evs.nci.nih.gov/ftp1/CTCAE/CTCAE_4.03/CTCAE_4.03_2010-06-14_Quic... https://www.meddra.org/

[Bosch2019] Bosch Sensortec, BME680 Low power gas, pressure, temperature & humidity sensor

[Herod2018] Kristen Karl Herod, Analyzing and Optimizing an Array of Low-Cost Gas Sensors for use in an Air Quality Measurement Device with Machine Learning

[WWTSM2019]Y. Wang, S. Willis, V. Tsoutsouras, and P. Stanley-Marbell. 2019. Deriving Equations from Sensor Data Using Dimensional Function Synthesis. ACM Trans. Embedd. Comput. Syst. 1, 1 (September 2019), 22 pages.

This project will build on recent research results [WWTSM2019] on exploiting the principle of dimensional homogeneity to learn functions that relate sensor signals in a multi-modal sensor system. The project will use a state-of-the-art multi-modal sensor system. In this project, you will:

(1) Investigate extending dimensional function synthesis for sensor data with the concept of conditional proportionality relationships, based on identifying model parameters which are constant.

(2) Evaluate first-principles methods for replacing energy-costly sensor types (e.g., gyroscopes) with energy-efficient sensor types (e.g., accelerometers) for two sensor-augmented mechanical systems which will be made available to the student.

(3) Apply the technique of dimensional function synthesis for sensor data to two mechanical systems provided and investigate new methods for replacing energy-costly sensor types (e.g., gyroscopes) with energy-efficient sensor types (e.g., accelerometers) for the specific mechanical systems in question.

(4) Investigate and evaluate the benefit of including information about prior statistics into the methods above and applying concepts from Bayesian statistics to obtain probabilistic synthetic sensor models.

(5) Evaluate the accuracy, result uncertainty, performance, and power dissipation tradeoffs in performing sensor substitution using both first-principles and dimensional-function-synthesized approaches for the two sensor-augmented mechanical systems.

Algorithms which compete for scarce energy, computing, and sensor access resources need to be able to satisfy their sensor data sampling and quality needs without depleting sensing energy resources. The quality of data from all modern integrated sensors depends on their operating configurations [SMR16]. These configurations include their operating voltage, sampling rate, sample precision, and dynamic range of their signal input interfaces. In practice, these configuration parameters also have a significant effect on the energy used by sensors for generating each sample. Because the power dissipated by sensors can equal or exceed the power used in optimized state-of-the-art microcontrollers, reducing sensor energy usage can have a significant impact on whole-system energy use in sensing platforms. The aim of this project is to investigate efficient sensor access schedules under fidelity requirements. The project will investigate new algorithms to schedule accesses to sensors given a set of sensor access jobs. Let a sensor access schedule be an ordering for carrying out sensing operations by activating sensors with given sensor parameter configurations. Let a constraint on the tail distribution of tolerable sample precisions, accuracies, erasures, and latencies of the nth sensor access be Pn, An, En, and Ln, respectively. To formulate the problem of sensor access scheduling when schedules must describe not just time of access, but also required sensor fidelity, let j = ( j1 , j2 , . . . , jk ) be the sequence of k sensor access events visible to a scheduler at a given point in time. We can then define a schedule for a set of k accesses across Ns sensors as a pair of functions (C(j, t), J(j, t)). C(j, t) is a function from the vector of sensor access jobs and time steps to the configurations of the Ns sensors. J(j, t) is a function from the vector of sensor access jobs and time steps to a sequence of vectors indicating the activation state of each of the Ns sensors: The kth element in the tth vector is 1 if sensor k is accessed by the jobs serviced at time t and is 0 otherwise. The challenge is to find the pair of functions (C(j, t), J(j, t)) for a given Pn, An, En, and Ln. In this project, you will:

(1) investigate the fundamental challenges and algorithms for computing schedules incrementally with the arrival of each sensor access request (dynamic online scheduling).

(2) Investigate algorithms to compute a schedule with complete knowledge of all sensor accesses that will comprise the schedule (static offline scheduling)

(3) Evaluate the effectiveness of the static-offline and dynamic-online schedulers using real-world sensor access activity traces which will be provided to you.

[SMR16] P. Stanley-Marbell and M. Rinard. Lax: Driver interfaces for approximate sensor device access. In HotOS XV, 2015.

[M81] R. B. Myerson. design. Mathematics of operations research, 6(1):58–73, 1981
[ FD91] D. Fudenberg and J. Tirole. Game Theory. MIT Press, 1991.

Dimensional function synthesis [WWTSM2019] is a new method to improve effectiveness of machine learning from sensor data, by performing dimensionality reduction on multi-modal sensor data to make subsequent machine learning steps more effective. It can improve the speed of training neural networks on sensor data by many orders of magnitude for some applications and can also improve the performance of inference from multi-modal sensor data many-fold. This project will investigate feasibility of an end-to-end integration of dimensional function synthesis on multi-modal sensor data with an open-source neural network accelerator [Marlann 2019] running on a miniature low-power field-programmable gate array (FPGA) [iCE40 2019]. The investigation will evaluate the circuit size, power dissipation, and performance of hardware implementations of circuits generated by dimensional function synthesis for two simple mechanical systems. The project will evaluate the circuit size, power dissipation, and performance of the Marlann neural network accelerator and will then investigate coupling the output of dimensional function synthesis to Marlann. In this project, you will:

(1) Investigate the circuit size, power dissipation, and performance of hardware implementations of circuits generated by dimensional function synthesis for two sensor-augmented mechanical systems, on the iCE40 FPGA (using the provided dimensional function synthesis implementation). Investigate the circuit size, power dissipation, and performance of the Marlann neural network accelerator on the iCE40 FPGA.

(2) Investigate alternative methods for efficient machine learning from sensor data. This could involve either a literature (meta-) survey or lab-based evaluation of techniques from the research literature for head-to-head comparisons.

(3) Propose implementation approaches for integrating the dimensionality reduction of dimensional function synthesis with hardware implementations of neural networks.

[WWTSM2019] Y. Wang, S. Willis, V. Tsoutsouras, and P. Stanley-Marbell. 2019. Deriving Equations from Sensor Data Using Dimensional Function Synthesis. ACM Trans. Embedd. Comput. Syst. 1, 1 (September 2019), 22 pages.

[Marlann2019] SymbioticEDA: MARLANN -- A simple FPGA Machine Learning Accelerator, https://github.com/SymbioticEDA/MARLANN [iCE40 2019] Lattice Semiconductor, iCE40 FPGA.

- Gauge Fields, Knots and Gravity, by John Baez and Javier P Muniain

- The webpage of the proof assistant Lean https://leanprover.github.io/

- The public Zulip chat for Lean

Friend AD, Eckes-Shephard AH, Fonti P, Rademacher TT, Rathgeber C, Richardson AD, Turton RH. 2019. On the need to consider wood formation processes in global vegetation models and a suggested approach. Annals of Forest Science, 76:49.

Hayat A, Hacket-Pain AJH, Pretzsch H, Rademacher TT, Friend AD. 2017. Modeling tree growth taking into account carbon source and sink limitations. Frontiers in Plant Science 8, 182. Friend AD, 22 others. 2014. Carbon residence time dominates uncertainty in terrestrial vegetation responses to future climate and atmospheric CO2. PNAS 111, 3280-3285.

We can offer many possible paths for a project, depending on the student's specific interests and skills. Some of the projects focus mostly on mathematical aspects of models: can we improve current models of DNA change by better predicting patterns in DNA sequence evolution, without compromising their computational efficiency? Are these models statistically identifiable? These projects require at least a basic understanding of probability and in particular of Markov models; some could require advanced probability and statistics.

Other projects are more computational, and require the student to think about how to improve certain algorithms and be able to code (preferentially in Python or C++), but still require the understanding and development of mathematical models. For example, in one project we want to improve the efficiency of certain sequencing technologies. A DNA sequencing machine reads the content of DNA samples, telling a computer the sequence of letters in the DNA of the sample. However, this information comes in small blocks, called reads, from random parts of the original DNA sequence. Our approach is to tell the machine which blocks are interesting and which are not, so that interesting blocks can be wholly read, while non-interesting blocks can be rejected, saving time and resources. The approach involves improving the computational efficiency of the current methods developed in our lab, and developing new methods based on alternative models of efficiency.

A further project, again requiring both modeling and computational skills, involves the genetic code: the set of rules that determine how DNA information is interpreted by living organisms that contain it. We want to investigate if the genetic code evolved to be robust to certain mutational events, and if the genetic code could be improved in this respect. This project will require the mathematical formalization of the properties of the genetic code, and the computational skills to write code to explore, assess and optimize large numbers of alternative genetic codes.

The project aims to develop general methods for sampling from a given probability distribution on a manifold to provide theoretical guarantees for the performance of these methods, and to explore their practical implementation and concrete applications. Sampling from the distribution means generating samples (random points) that satisfy the law of large numbers (LLN), which is at the heart of any Monte Carlo method (MC). Monte Carlo methods are fundamental to stochastic optimisation, Bayesian inference, and many other computational paradigms which are of fundamental importance in information engineering.

In order of increasing difficulty, the project may consider the following problems, (a) sampling on a compact homogeneous space from a uniform distribution, (b) sampling on a homogeneous space from an isotropy invariant distribution, (c) sampling on any suitable Riemannian manifold from a partially unknown distribution. Problem (a) has been considered extensively, in connection with random matrix theory. On the other hand, problem (b) is a fairly open challenge, at the heart of ongoing research in non-Euclidean machine learning. Compact Riemannian homogeneous manifolds include compact Lie groups, and Stiefel and Grassmann manifolds, which are of fundamental importance in robotics, telecommunications, and other fields. Among general (non-compact) homogeneous Riemannian manifolds, one finds hyperbolic spaces and various spaces of covariance matrices, which play a central role in brain-computer interface analysis (in addition to their great intrinsic mathematical interest).

Problem (c) is also a cutting-edge problem, treated only in very few recent papers, which deal with Markov Chain Monte Carlo methods, in Riemannian manifolds. The issue of an unknown target distribution often arises in Bayesian inference, where the probability density is known only up to a normalising factor. The development of sampling strategies on manifolds equipped with a variety of different metric structures such as Finsler manifolds is another possible research direction for the project. The resulting algorithms can be tested on real problems and data with the aim of systematically identifying the most appropriate metric structures for use in analysis and statistics for specific applications. Special consideration will be given to the space of SPDs of a fixed dimension. 
The project contains a wide choice of applied (ranging from building software toolboxes to proving theoretical performance guarantees) and pure problems (which concern the interplay of probability, geometry and harmonic analysis), and can therefore be adapted to the student's preference and motivation.

1. Absil, P.-A. , Mahony, R. and Sepulchre, R.: Optimization algorithms on matrix manifolds. Princeton University Press (2008).

2. Meckes, E. S.: Random matrix theory of the classical compact groups. Cambridge University Press (2019).

3. Said, S., Hajri, H., Bombrun, L. and Vemuri, B. C.: Gaussian distributions on Riemannian symmetric spaces. IEEE Trans. Inf. Theory 64(2) (2018).

4. Said, S. : On the Riemannian barycentre of a Markov chain. (arxiv:1908.08912, pre-print) (2019).

5. Diaconis, P., Holmes, S., Shahshahani, M.: Sampling from a manifold. (arXiv:1206.6913) (2012).

 1. Sozen B, Amadei G, Cox A, Wang R, Na E, Czukiewska S, Chappell L, Voet T, Michel G, Jing N, Glover DM, Zernicka-Goetz M (2018) Self-assembly of embryonic and two extraembryonic stem cell types into gastrulating embryo-like structures. Nature Cell Biology. 20: 979-989

2. LF Lang - A Numerical Framework for Efficient Motion Estimation on Evolving Sphere-Like Surfaces Based on Brightness and Mass Conservation Laws, SIAM Journal on Imaging Sciences 12 (1), 459-491

[1] Nicholas Boyd, Geoffrey Schiebinger, and Benjamin Recht. The alternating descent conditional gradient method for sparse inverse problems. SIAM Journal on Optimization, 27(2):616–639, 2017. [

2] Quentin Denoyelle, Vincent Duval, Gabriel Peyr ́e, and Emmanuel Soubies. The sliding Frank-Wolfe algorithm and its application to super-resolution microscopy. Inverse Problems, 36(1), 2019.

R. J. Shipley, A. F. Smith, P. W. Sweeney, A. R. Pries, and T. W. Secomb. 2019. A hybrid discrete-continuum approach for modelling microcirculatory blood flow. Mathematical Medicine & Biology, dqz006:1-18. DOI: https://doi.org/10.1093/imammb/dqz006.

P. W. Sweeney. 2018. Realistic numerical image-based modelling of biological tissue substrates. Doctoral thesis (PhD), University College London.

R. J. Shipley & S. Chapman. 2010. Multiscale modelling of fluid and drug transport in vascular tumours. Bulletin of Mathematical Biology, 72(6):1464-91.

The project will build on recent work done in our lab to unravel the earliest events that led to canine transmissible venereal tumour (CTVT) becoming an infectious cancer, and globally endemic. We have high coverage DNA sequencing information from several CTVT samples, as well as their hosts, which we have previously used to estimate the evolutionary tree underlying these tumours. This previous work has identified a new mutational signature (‘signature A’) that was active during the early evolution of CTVT, but later switched off.

In this project we will aim to use tumour copy number estimation to identify ‘frozen time points’: fragments of DNA sequence that were gained during the early evolution of the cancer. The relative ages of these fragments will be estimated by the degree to which they have accumulated somatic mutations. Combined with this timing information, by examining the fragments for the presence of signature A we will be able to determine whether signature A occurred continuously, or in bursts. Additionally, the earliest fragments will be highly representative of the genotype of the animal in which CTVT arose, illuminating perhaps the characteristics of the earliest domesticated dogs.

While we have a fixed goal in mind for the project, namely inferring the nature of the early mutations that occurred in CTVT, there is also plenty of scope for a student to pursue their own interests within the subject area, depending on how the project progresses.

1: Strakova, Andrea, and Elizabeth P. Murchison. 2015. “The Cancer Which Survived: Insights from the Genome of an 11000 Year-Old Cancer.” Current Opinion in Genetics & Development 30: 49–55.

2: Leathlobhair, Máire Ní, Angela R. Perri, Evan K. Irving-Pease, Kelsey E. Witt, Anna Linderholm, James Haile, Ophelie Lebrasseur, et al. 2018. “The Evolutionary History of Dogs in the Americas.” Science 361 (6397): 81–85.

3: Baez-Ortega, Adrian, Kevin Gori, Andrea Strakova, Janice L. Allen, Karen M. Allum, Leontine Bansse-Issa, Thinlay N. Bhutia, et al. 2019. “Somatic Evolution and Global Expansion of an Ancient Transmissible Cancer Lineage.” Science 365 (6452): eaau9923.