Search

The development and deployment of machine learning (ML) systems can be executed easily with modern tools, but the process is typically rushed and means-to-an-end. The lack of diligence can lead to technical debt, scope creep and misaligned objectives, model misuse and failures, and expensive consequences. Engineering systems, on the other hand, follow well-defined processes and testing standards to streamline development for high-quality, reliable results. The extreme is spacecraft systems, where mission critical measures and robustness are ingrained in the development process. Drawing on experience in both spacecraft engineering and ML (from research through product across domain areas), we have developed a proven systems engineering approach for machine learning development and deployment. Our Machine Learning Technology Readiness Levels (MLTRL) framework defines a principled process to ensure robust, reliable, and responsible systems while being streamlined for ML workflows, including key distinctions from traditional software engineering. Even more, MLTRL defines a lingua franca for people across teams and organizations to work collaboratively on artificial intelligence and machine learning technologies. Here we describe the framework and elucidate it with several real world use-cases of developing ML methods from basic research through productization and deployment, in areas such as medical diagnostics, consumer computer vision, satellite imagery, and particle physics.

The increasing number of vision-based tracking systems deployed in production have necessitated fast, robust camera calibration. In the domain of sport, the majority of current work focuses on sports where lines and intersections are easy to extract, and appearance is relatively consistent across venues. However, for more challenging sports like basketball, those techniques are not sufficient. In this paper, we propose an end-to-end approach for single moving camera calibration across challenging scenarios in sports. Our method contains three key modules: 1) area-based court segmentation, 2) camera pose estimation with embedded templates, 3) homography prediction via a spatial transform network (STN). All three modules are connected, enabling end-to-end training. We evaluate our method on a new college basketball dataset and demonstrate state of the art performance in variable and dynamic environments. We also validate our method on the World Cup 2014 dataset to show its competitive performance against the state-of-the-art methods. Lastly, we show that our method is two orders of magnitude faster than the previous state of the art on both datasets.

The major motor Kinesin-1 provides a key pathway for cell polarization through intracellular transport. Little is known about how Kinesin works in complex cellular surroundings. Several cargos associate with Kinesin via Kinesin light chain (KLC). However, KLC is not required for all Kinesin transport. A putative cargo-binding domain was identified in the C-terminal tail of fungal Kinesin heavy chain (KHC). The tail is conserved in animal KHCs and might therefore represent an alternative KLC-independent cargo-interacting region. By comprehensive functional analysis of the tail during Drosophila oogenesis we have gained an understanding of how KHC achieves specificity in its transport and how it is regulated. This is, to our knowledge, the first in vivo structural/functional analysis of the tail in animal Kinesins. We show that the tail is essential for all functions of KHC except Dynein transport, which is KLC dependent. These tail-dependent KHC activities can be functionally separated from one another by further characterizing domains within the tail. In particular, our data show the following. First, KHC is temporally regulated during oogenesis. Second, the IAK domain has an essential role distinct from its auto-inhibitory function. Third, lack of auto-inhibition in itself is not necessarily detrimental to KHC function. Finally, the ATP-independent microtubule-binding motif is required for cargo localization. These results stress that two unexpected highly conserved domains, namely the auto-inhibitory IAK and the auxiliary microtubule-binding motifs, are crucial for transport by Kinesin-1 and that, although not all cargos are conserved, their transport involves the most conserved domains of animal KHCs.

Bacterial processes ranging from gene expression to motility and biofilm formation are constantly challenged by internal and external noise. While the importance of stochastic fluctuations has been appreciated for chemotaxis, it is currently believed that deterministic long-range fluid dynamical effects govern cell-cell and cell-surface scattering - the elementary events that lead to swarming and collective swimming in active suspensions and to the formation of biofilms. Here, we report the first direct measurements of the bacterial flow field generated by individual swimming Escherichia coli both far from and near to a solid surface. These experiments allowed us to examine the relative importance of fluid dynamics and rotational diffusion for bacteria. For cell-cell interactions it is shown that thermal and intrinsic stochasticity drown the effects of long-range fluid dynamics, implying that physical interactions between bacteria are determined by steric collisions and near-field lubrication forces. This dominance of short-range forces closely links collective motion in bacterial suspensions to self-organization in driven granular systems, assemblages of biofilaments, and animal flocks. For the scattering of bacteria with surfaces, long-range fluid dynamical interactions are also shown to be negligible before collisions; however, once the bacterium swims along the surface within a few microns after an aligning collision, hydrodynamic effects can contribute to the experimentally observed, long residence times. As these results are based on purely mechanical properties, they apply to a wide range of microorganisms., 9 pages, 2 figures, http://www.pnas.org/content/108/27/10940

The morphology of a neuron is key to its function, but the principles that govern neuronal morphogenesis are not clear. To investigate these principles, we used the larval class IV sensory neuron in Drosophila as a model cell. Class IV neurons have highly branched dendritic morphology. Our specific question is whether this branching morphology arises from purely random processes or whether there exist non-random constraints on morphological parameters such as segment lengths and branching angles.To measure the statistical characteristics of the dendritic arbors, we imaged class IV neurons by confocal microscopy and analyzed their skeletons using Fiji and Matlab. First, we found that the lengths of dendritic segments, both terminal and non-terminal, followed exponential distributions. Given that the lengths of the dendritic segments are defined by consecutive branch points, this observation suggests that branching events follow a spatial Poisson process. Second, we found that the angles between two daughter segments follow a normal distribution with a mean of 96 degrees and a standard deviation of 31 degrees (n = 465). Because the mean differs from 180 degrees, we conclude that the branching angles are not uniformly distributed. These properties, namely the distributions of segment lengths and angles, were observed throughout morphogenesis.Our results indicate that there are morphological properties of class IV neurons which are not determined by purely random processes.

Spatial order and fast collective coherent dynamics of populations of the swimming bacteria Bacillus subtilis emerges from local interactions and from flows generated by the organisms’ locomotion. The transition from dilute, to intermediate, to high concentrations of cells is analyzed and presented as probability density functions for swimming velocity. The low concentration phase, which exhibits swimming speeds characteristic of individual bacteria, arrives at the anomalously high speed phase,the ZoomingBioNematic (ZBN), via an intermediate phase that exhibits surprisingly low mean speeds. We show that these low speeds at intermediate concentrations are due to transitional speeds that occur after collisions of the organisms, while the flagella that propel the bacteria re-form into a “bundle”. Measurement of individual and collective velocities, as well as correlation of speeds with alignment of velocity directions, within and adjacent to coherent patches, were found by Particle Imaging Velocimetry (PIV). The significance for mixing of the ZBN dynamic is demonstrated.

The persistent circulation of the cytoplasm, called cytoplasmic streaming, occurs in a variety of eukaryotic cells. One context in which streaming occurs is during the establishment of Drosophila body axes, when Kinesin-1 transports the axes determinants and drives cytoplasmic streaming. Although Kinesin is essential for flows, neither the mechanism by which Kinesin induces streaming nor the impact of these flows on transport are known. We have succeeded in a precise quantitative measurement of the statistical properties of streaming by Particle Image Velocimetry. We have combined these measurements with an in vivo study of the cytoplasm rheology, to calculate the energy dissipation due to streaming. Since Kinesin is required for flows we can relate the energy dissipated to the work done on the fluid by Kinesin and determine the minimum number of motors necessary to drive streaming. Furthermore we have performed these measurements on mutants that effect Kinesin-1 motor activity and found remarkable agreement between our in vivo measurements and in vitro studies.

Flagella-generated fluid stirring has been suggested to enhance nutrient uptake for sufficiently large micro-organisms, and to have played a role in evolutionary transitions to multicellularity. A corollary to this predicted size-dependent benefit is a propensity for phenotypic plasticity in the flow-generating mechanism to appear in large species under nutrient deprivation. We examined four species of volvocalean algae whose radii and flow speeds differ greatly, with Péclet numbers ( Pe ) separated by several orders of magnitude. Populations of unicellular Chlamydomonas reinhardtii and one- to eight-celled Gonium pectorale ( Pe ∼ 0.1–1) and multicellular Volvox carteri and Volvox barberi ( Pe ∼ 100) were grown in diluted and undiluted media. For C. reinhardtii and G. pectorale , decreasing the nutrient concentration resulted in smaller cells, but had no effect on flagellar length and propulsion force. In contrast, these conditions induced Volvox colonies to grow larger and increase their flagellar length, separating the somatic cells further. Detailed studies on V. carteri found that the opposing effects of increasing beating force and flagellar spacing balance, so the fluid speed across the colony surface remains unchanged between nutrient conditions. These results lend further support to the hypothesized link between the Péclet number, nutrient uptake and the evolution of biological complexity in the Volvocales.

At high cell concentrations, bacterial suspensions are known to develop a state of collective swimming (the "zooming bionematic phase," or ZBN) characterized by transient, recurring regions of coordinated motion greatly exceeding the size of individual cells. Recent theoretical studies of semidilute suspensions have suggested that long-range hydrodynamic interactions between swimming cells are responsible for long-wavelength instabilities that lead to these patterns, while models appropriate for higher concentrations have suggested that steric interactions between elongated cells play an important role in the self-organization. Using particle imaging velocimetry in well-defined microgeometries, we examine the statistical properties of the transition to the ZBN in suspensions of Bacillus subtilis, with particular emphasis on the distribution of cell swimming speeds and its correlation with orientational order. This analysis reveals a nonmonotonic relationship between mean cell swimming speed and cell concentration, with a minimum occurring near the transition to the ZBN. Regions of high orientational order in the ZBN phase have locally high swimming speeds, while orientationally disordered regions have lower speeds. A model for steric interactions in concentrated suspensions and previous observations on the kinetics of flagellar rebundling associated with changes in swimming direction are used to explain this observation. The necessity of incorporating steric effects on cell swimming in theoretical models is emphasized.

Evolution from unicellular organisms to larger multicellular ones requires matching their needs to the rate of exchange of molecular nutrients with the environment. This logistic problem poses a severe constraint on development. For organisms whose body plan is a spherical shell, such as the volvocine green algae, the current (molecules per second) of needed nutrients grows quadratically with radius, whereas the rate at which diffusion alone exchanges molecules grows linearly, leading to a bottleneck radius beyond which the diffusive current cannot meet metabolic demands. By usingVolvox carteri, we examine the role that advection of fluid by the coordinated beating of surface-mounted flagella plays in enhancing nutrient uptake and show that it generates a boundary layer of concentration of the diffusing solute. That concentration gradient produces an exchange rate that is quadratic in the radius, as required, thus circumventing the bottleneck and facilitating evolutionary transitions to multicellularity and germ–soma differentiation in the volvocalean green algae.

Benefits, costs, and requirements accompany the transition from motile totipotent unicellular organisms to multicellular organisms having cells specialized into reproductive (germ) and vegetative (sterile soma) functions such as motility. In flagellated colonial organisms such as the volvocalean green algae, organized beating by the somatic cells' flagella yields propulsion important in phototaxis and chemotaxis. It has not been generally appreciated that for the larger colonies flagellar stirring of boundary layers and remote transport are fundamental for maintaining a sufficient rate of metabolite turnover, one not attainable by diffusive transport alone. Here, we describe experiments that quantify the role of advective dynamics in enhancing productivity in germ soma-differentiated colonies. First, experiments with suspended deflagellated colonies ofVolvox carterishow that forced advection improves productivity. Second, particle imaging velocimetry of fluid motion around colonies immobilized by micropipette aspiration reveals flow fields with very large characteristic velocitiesUextending to length scales exceeding the colony radiusR. For a typical metabolite diffusion constantD, the associated Peclet numberPe= 2UR/D>> 1, indicative of the dominance of advection over diffusion, with striking augmentation at the cell division stage. Near the colony surface, flows generated by flagella can be chaotic, exhibiting mixing due to stretching and folding. These results imply that hydrodynamic transport external to colonies provides a crucial boundary condition, a source for supplying internal diffusional dynamics.

One key to the function of a neuron is the morphology of its dendritic arbors. In this work, we address the question of which branching rules are necessary to reproduce the complexity of arborization. We used the larval class IV sensory neuron in Drosophila as the model cell to approach this question.As class IV neurons display self-similarity over a range scales, the first key morphological parameter we use to study them is their fractal dimension. The fractal dimension of a neuron is a measure of its complexity and has been used to distinguish between classes of neurons. The second morphological parameter of a neuron involves realizing that such a branching structure can be viewed as a binary tree in which neuronal branching points are the nodes. The structure of interest here is the distribution of node depths, where the depth of a node is the number of other nodes between it and the root (i.e., the cell body) on the tree.Using both analytical techniques and in silico simulations, we made three findings. 1) The fractal dimension was always a monotonically increasing function of the neuron's maximal depth. 2) The observed Gaussian node-depth distributions are achievable via a termination rule in which the probability of branch termination is a sigmoidal function of node depth. 3) The observed node-depth distributions can be qualitatively accounted for by an "inheritance rule", whereby each daughter segment inherits morphological information from its mother segment.In conclusion, we demonstrate that a set of statistical rules accounts for the fractal dimension and node-depth distribution of class IV neurons.