Fruit flies (Drosophila melanogaster) are small insects, with correspondingly small power budgets. Despite this, they perform sophisticated neural computations in real time. Careful study of these insects is revealing how some of these circuits work. Insights from these systems might be helpful in designing other low power circuits.

Natural neural circuits, optimized by millions of years of evolution, are fast, low power, robust, and adapt in response to experience, all characteristics we would love to have in systems we ourselves design. Recently there have been enormous advances in understanding how neurons implement computations within the brain of living creatures. Can we use this new-found knowledge to create better artificial system? What lessons can we learn from the neurons themselves, that can help us create better neuromorphic circuits?

The ability to automatically segment an image into distinct regions is a critical aspect in many visual processing applications. Because inaccuracies often exist in automatic segmentation, manual segmentation is necessary in some application domains to correct mistakes, such as required in the reconstruction of neuronal processes from microscopic images. The goal of the automated segmentation tool is traditionally to produce the highest-quality segmentation, where quality is measured by the similarity to actual ground truth, so as to minimize the volume of manual correction necessary. Manual correction is generally orders-of-magnitude more time consuming than automated segmentation, often making handling large images intractable. Therefore, we propose a more relevant goal: minimizing the turn-around time of automated/manual segmentation while attaining a level of similarity with ground truth. It is not always necessary to inspect every aspect of an image to generate a useful segmentation. As such, we propose a strategy to guide manual segmentation to the most uncertain parts of segmentation. Our contributions include 1) a probabilistic measure that evaluates segmentation without ground truth and 2) a methodology that leverages these probabilistic measures to significantly reduce manual correction while maintaining segmentation quality.

Due to advances in electron microscopy and deep learning, it is now practical to reconstruct a connectome, a description of neurons and the chemical synapses between them, for significant volumes of neural tissue. Smaller past reconstructions were primarily used by domain experts, could be handled by downloading data, and performance was not a serious problem. But new and much larger reconstructions upend these assumptions. These networks now contain tens of thousands of neurons and tens of millions of connections, with yet larger reconstructions pending, and are of interest to a large community of non-specialists. Allowing other scientists to make use of this data needs more than publication-it requires new tools that are publicly available, easy to use, and efficiently handle large data. We introduce neuPrint to address these data analysis challenges. Neuprint contains two major components-a web interface and programmer APIs. The web interface is designed to allow any scientist worldwide, using only a browser, to quickly ask and answer typical biological queries about a connectome. The neuPrint APIs allow more computer-savvy scientists to make more complex or higher volume queries. NeuPrint also provides features for assessing reconstruction quality. Internally, neuPrint organizes connectome data as a graph stored in a neo4j database. This gives high performance for typical queries, provides access though a public and well documented query language Cypher, and will extend well to future larger connectomics databases. Our experience is also an experiment in open science. We find a significant fraction of the readers of the article proceed to examine the data directly. In our case preprints worked exactly as intended, with data inquiries and PDF downloads starting immediately after pre-print publication, and little affected by formal publication later. From this we deduce that many readers are more interested in our data than in our analysis of our data, suggesting that data-only papers can be well appreciated and that public data release can speed up the propagation of scientific results by many months. We also find that providing, and keeping, the data available for online access imposes substantial additional costs to connectomics research.

Reconstructing neuronal circuits at the level of synapses is a central problem in neuroscience, and the focus of the nascent field of connectomics. Previously used to reconstruct the C. elegans wiring diagram, serial-section transmission electron microscopy (ssTEM) is a proven technique for the task. However, to reconstruct more complex circuits, ssTEM will require the automation of image processing. We review progress in the processing of electron microscopy images and, in particular, a semi-automated reconstruction pipeline deployed at Janelia. Drosophila circuits underlying identified behaviors are being reconstructed in the pipeline with the goal of generating a complete Drosophila connectome.

Connectomes derived from volume EM imaging of the brain can generate detailed physical models of every neuron, and simulators such as NEURON or GENESIS are designed to work with such models. In principal, combining these technologies, plus transmitter and channel models, should allow detailed and accurate simulation of real neural circuits. Here we experiment with this combination, using a well-studied system (motion detection in Drosophila. Since simulation requires both the physical geometry (which we have) and the models of the synapses (which are not currently available), we built approximate synapses corresponding to their known and estimated function. Once we did so, we reproduced direction selectivity in T4 cells, one of the main functions of this neural circuit. This verified the basic functionality of both extraction and simulations, and provided a biologically relevant computation we could use in further experiments. We then compared models with different degrees of physical realism, from full detailed models down to models consisting of a single node, to examine the tradeoff of simulation resources required versus accuracy achieved. Our results show that much simpler models may be adequate, at least in the case of medulla neurons in Drosophila. Such models can be easily derived from fully detailed models, and result in simulations that are much smaller, much faster, and accurate enough for many purposes. Biologically, we show that a lumped neuron model reproduces the main motion detector operation, confirming the result of Gruntman, that dendritic compution is not required for this function.

Pixel and superpixel classifiers have become essential tools for EM segmentation algorithms. Training these classifiers remains a major bottleneck primarily due to the requirement of completely annotating the dataset which is tedious, error-prone and costly. In this paper, we propose an interactive learning scheme for the superpixel classifier for EM segmentation. Our algorithm is "active semi-supervised" because it requests the labels of a small number of examples from user and applies label propagation technique to generate these queries. Using only a small set (<20%) of all datapoints, the proposed algorithm consistently generates a classifier almost as accurate as that estimated from a complete groundtruth. We provide segmentation results on multiple datasets to show the strength of these classifiers.

A central problem in neuroscience is reconstructing neuronal circuits on the synapse level. Due to a wide range of scales in brain architecture such reconstruction requires imaging that is both high-resolution and high-throughput. Existing electron microscopy (EM) techniques possess required resolution in the lateral plane and either high-throughput or high depth resolution but not both. Here, we exploit recent advances in unsupervised learning and signal processing to obtain high depth-resolution EM images computationally without sacrificing throughput. First, we show that the brain tissue can be represented as a sparse linear combination of localized basis functions that are learned using high-resolution datasets. We then develop compressive sensing-inspired techniques that can reconstruct the brain tissue from very few (typically 5) tomographic views of each section. This enables tracing of neuronal processes and, hence, high throughput reconstruction of neural circuits on the level of individual synapses.