Topology of synaptic connectivity constrains neuronal stimulus representation, predicting two complementary coding strategies

In motor-related brain regions, movement intention has been successfully decoded from in-vivo spike train by isolating a lower-dimension manifold that the high-dimensional spiking activity is constrained to. The mechanism enforcing this constraint remains unclear, although it has been hypothesized to be implemented by the connectivity of the sampled neurons. We test this idea and explore the interactions between local synaptic connectivity and its ability to encode information in a lower dimensional manifold through simulations of a detailed microcircuit model with realistic sources of noise. We confirm that even in isolation such a model can encode the identity of different stimuli in a lower-dimensional space. We then demonstrate that the reliability of the encoding depends on the connectivity between the sampled neurons by specifically sampling populations whose connectivity maximizes certain topological metrics. Finally, we developed an alternative method for determining stimulus identity from the activity of neurons by combining their spike trains with their recurrent connectivity. We found that this method performs better for sampled groups of neurons that perform worse under the classical approach, predicting the possibility of two separate encoding strategies in a single microcircuit.

Projection Path Explorer: Exploring Visual Patterns in Projected Decision-making Paths

In problem-solving, a path towards a solutions can be viewed as a sequence of decisions. The decisions, made by humans or computers, describe a trajectory through a high-dimensional representation space of the problem. By means of dimensionality reduction, these trajectories can be visualized in lower-dimensional space. Such embedded trajectories have previously been applied to a wide variety of data, but analysis has focused almost exclusively on the self-similarity of single trajectories. In contrast, we describe patterns emerging from drawing many trajectories—for different initial conditions, end states, and solution strategies—in the same embedding space. We argue that general statements about the problem-solving tasks and solving strategies can be made by interpreting these patterns. We explore and characterize such patterns in trajectories resulting from human and machine-made decisions in a variety of application domains: logic puzzles (Rubik’s cube), strategy games (chess), and optimization problems (neural network training). We also discuss the importance of suitably chosen representation spaces and similarity metrics for the embedding.

Self-Organising Map Based Framework for Investigating Accounts Suspected of Money Laundering

There has been an emerging interest by financial institutions to develop advanced systems that can help enhance their anti-money laundering (AML) programmes. In this study, we present a self-organising map (SOM) based approach to predict which bank accounts are possibly involved in money laundering cases, given their financial transaction histories. Our method takes advantage of the competitive and adaptive properties of SOM to represent the accounts in a lower-dimensional space. Subsequently, categorising the SOM and the accounts into money laundering risk levels and proposing investigative strategies enables us to measure the classification performance. Our results indicate that our framework is well capable of identifying suspicious accounts already investigated by our partner bank, using both proposed investigation strategies. We further validate our model by analysing the performance when modifying different parameters in our dataset.

Deep Clustering with Self-supervision using Pairwise Data Similarities

<div>Deep clustering incorporates embedding into clustering to find a lower-dimensional space appropriate for clustering. In this paper we propose a novel deep clustering framework with self-supervision using pairwise data similarities (DCSS). The proposed method consists of two successive phases. In the first phase we propose to form hypersphere-like groups of similar data points, i.e. one hypersphere per cluster, employing an autoencoder which is trained using cluster-specific losses. The hyper-spheres are formed in the autoencoder’s latent space. In the second phase, we propose to employ pairwise data similarities to create a K-dimensional space that is capable of accommodating more complex cluster distributions; hence, providing more accurate clustering performance. K is the number of clusters. The autoencoder’s latent space obtained in the first phase is used as the input of the second phase. Effectiveness of both phases are demonstrated on seven benchmark datasets through conducting a rigorous set of experiments.</div>

Harnessing Matrix Completion to Unify and Extend Viral Serology Studies

SummaryCharacterizing the antibody response against large panels of viral variants provides unique insight into key processes that shape viral evolution and host antibody repertoires, and has become critical to the development of new vaccine strategies. Given the enormous diversity of circulating virus strains and antibody responses, exhaustive testing of all antibody-virus interactions is unfeasible. However, prior studies have demonstrated that, despite the complexity of these interactions, their functional phenotypes can be characterized in a vastly simpler and lower-dimensional space, suggesting that matrix completion of relatively few measurements could accurately predict unmeasured antibody-virus interactions. Here, we combine available data from several of the largest-scale studies for both influenza and HIV-1 and demonstrate how matrix completion can substantially expedite experiments. We explore how prediction accuracy evolves as the number of available measurements changes and approximate the number of additional measurements necessary in several highly incomplete datasets (suggesting ∼250,000 measurements could be saved). In addition, we show how the method can be used to combine disparate datasets, even when the number of available measurements is below the theoretical limit for successful prediction. Our results suggest new approaches to improve ongoing experimental design, and could be readily generalized to other viruses or more broadly to other low-dimensional biological datasets.

Efficient High-Dimensional Kernel k-Means with Random Projection++

Using random projection, a method to speed up both kernel k-means and centroid initialization with k-means++ is proposed. We approximate the kernel matrix and distances in a lower-dimensional space Rd before the kernel k-means clustering motivated by upper error bounds. With random projections, previous work on bounds for dot products and an improved bound for kernel methods are considered for kernel k-means. The complexities for both kernel k-means with Lloyd’s algorithm and centroid initialization with k-means++ are known to be O(nkD) and Θ(nkD), respectively, with n being the number of data points, the dimensionality of input feature vectors D and the number of clusters k. The proposed method reduces the computational complexity for the kernel computation of kernel k-means from O(n2D) to O(n2d) and the subsequent computation for k-means with Lloyd’s algorithm and centroid initialization from O(nkD) to O(nkd). Our experiments demonstrate that the speed-up of the clustering method with reduced dimensionality d=200 is 2 to 26 times with very little performance degradation (less than one percent) in general.

Treemap-Based Cluster Visualization and its Application to Text Data Analysis

This paper proposes Treemap-based visualization for supporting cluster analysis of multi-dimensional data. It is important to grasp data distribution in a target dataset for such tasks as machine learning and cluster analysis. When dealing with multi-dimensional data such as statistical data and document datasets, dimensionality reduction algorithms are usually applied to project original data to lower-dimensional space. However, dimensionality reduction tends to lose the characteristics of data in the original space. In particular, the border between different data groups could not be represented correctly in lower-dimensional space. To overcome this problem, the proposed visualization method applies Fuzzy c-Means to target data and visualizes the result on the basis of the highest and the second-highest membership values with Treemap. Visualizing the information about not only the closest clusters but also the second closest ones is expected to be useful for identifying objects around the border between different clusters, as well as for understanding the relationship between different clusters. A prototype interface is implemented, of which the effectiveness is investigated with a user experiment on a news articles dataset. As another kind of text data, a case study of applying it to a word embedding space is also shown.

High Resolution, High Contrast Beamformer Using Minimum Variance and Plane Wave Nonlinear Compounding with Low Complexity

The plane wave compounding (PWC) is a promising modality to improve the imaging quality and maintain the high frame rate for ultrafast ultrasound imaging. In this paper, a novel beamforming method is proposed to achieve higher resolution and contrast with low complexity. A minimum variance (MV) weight calculated by the partial generalized sidelobe canceler is adopted to beamform the receiving array signals. The dimension reduction technique is introduced to project the data into lower dimensional space, which also contributes to a large subarray length. Estimation of multi-wave receiving covariance matrix is performed and then utilized to determine only one weight. Afterwards, a fast second-order reformulation of the delay multiply and sum (DMAS) is developed as nonlinear compounding to composite the beamforming output of multiple transmissions. Simulations, phantom, in vivo, and robustness experiments were carried out to evaluate the performance of the proposed method. Compared with the delay and sum (DAS) beamformer, the proposed method achieved 86.3% narrower main lobe width and 112% higher contrast ratio in simulations. The robustness to the channel noise of the proposed method is effectively enhanced at the same time. Furthermore, it maintains a linear computational complexity, which means that it has the potential to be implemented for real-time response.

scJoint: transfer learning for data integration of single-cell RNA-seq and ATAC-seq

Single-cell multi-omics data continues to grow at an unprecedented pace, and while integrating different modalities holds the promise for better characterisation of cell identities, it remains a significant computational challenge. In particular, extreme sparsity is a hallmark in many modalities such as scATAC-seq data and often limits their power in cell type identification. Here we present scJoint, a transfer learning method to integrate heterogeneous collections of scRNA-seq and scATAC-seq data. scJoint uses a neural network to simultaneously train labelled and unlabelled data and embed cells from both modalities in a common lower dimensional space, enabling label transfer and joint visualisation in an integrative framework. We demonstrate scJoint consistently provides meaningful joint visualisations and achieves significantly higher label transfer accuracy than existing methods using a complex cell atlas data and a biologically varying multi-modal data. This suggests scJoint is effective in overcoming the heterogeneity in different modalities towards a more comprehensive understanding of cellular phenotypes.

Topology of synaptic connectivity constrains neuronal stimulus representation, predicting two complementary coding strategies

In motor-related brain regions, movement intention has been successfully decoded from in-vivo spike train by isolating a lower-dimension manifold that the high-dimensional spiking activity is constrained to. The mechanism enforcing this constraint remains unclear, although it has been hypothesized to be implemented by the connectivity of the sampled neurons. We test this idea and explore the interactions between local synaptic connectivity and its ability to encode information in a lower dimensional manifold through simulations of a detailed microcircuit model with realistic sources of noise. We confirm that even in isolation such a model can encode the identity of different stimuli in a lower-dimensional space. We then demonstrate that the reliability of the encoding depends on the connectivity between the sampled neurons by specifically sampling populations whose connectivity maximizes certain topological metrics. Finally, we developed an alternative method for determining stimulus identity from the activity of neurons by combining their spike trains with their recurrent connectivity. We found that this method performs better for sampled groups of neurons that perform worse under the classical approach, predicting the possibility of two separate encoding strategies in a single microcircuit.