Explainable Neural Network Ensembles

Neural networks are known for providing impressive classification performance, and the ensemble learning technique is further acting as a catalyst to enhance this performance by integrating multiple networks. But like neural networks, neural network ensembles are also considered as a black-box because they cannot explain their decision making process. So, despite having high classification performance, neural networks and their ensembles are not suited for some applications which require explainable decisions. However, the rule extraction technique can overcome this drawback by representing the knowledge learned by a neural network in the guise of interpretable decision rules. A rule extraction algorithm provides neural networks with the power to justify their classification responses through explainable classification rules. Several rule extraction algorithms exist to extract classification rules from neural networks, but only a few of them generates rules using neural network ensembles. So this paper proposes an algorithm named Rule Extraction using Ensemble of Neural Network Ensembles (RE-E-NNES) to demonstrate the high performance of neural network ensembles through rule extraction. RE-E-NNES extracts classification rules by ensembling several neural network ensembles. Results show the efficacy of the proposed RE-E-NNES algorithm compared to different existing rule extraction algorithms.

Using Neural Network Ensembles to Separate Biogeochemical and Physical Components in Earth System Models

Earth system models (ESMs) are useful tools for predicting and understanding past and future aspects of the climate system. However, the biological and physical parameters used in ESMs can have wide variations in their estimates. Even small changes in these parameters can yield unexpected results without a clear explanation of how a particular outcome was reached. The standard method for estimating ESM sensitivity is to compare spatiotemporal distributions of variables from different runs of a single ESM. However, a potential pitfall of this method is that ESM output could match observational patterns because of compensating errors. For example, if a model predicts overly weak upwelling and low nutrient concentrations, it may compensate for this by allowing phytoplankton to have a high sensitivity to nutrients. Recently, it has been demonstrated that neural network ensembles (NNEs) are capable of extracting relationships between predictor and target variables within ocean biogeochemical models. Being able to view the relationships between variables, along with spatiotemporal distributions, allows for a more mechanistically based examination of ESM outputs. Here, we investigated whether we could apply NNEs to help us determine why different ESMs produce different results. We tested this using three cases. The first and second case use different runs of the same ESM, except the physical circulations differ between them in the first case while the biological equations differ between them in the second. Our results indicate that the NNEs were capable of extracting the relationships between variables, allowing us to distinguish between differences due to changes in circulation (which do not change relationships) from changes in biogeochemical formulation (which do change relationships).

Accurate prediction of clinical stroke scales and improved biomarkers of motor impairment from robotic measurements

Objective One of the greatest challenges in clinical trial design is dealing with the subjectivity and variability introduced by human raters when measuring clinical end-points. We hypothesized that robotic measures that capture the kinematics of human movements collected longitudinally in patients after stroke would bear a significant relationship to the ordinal clinical scales and potentially lead to the development of more sensitive motor biomarkers that could improve the efficiency and cost of clinical trials. Materials and methods We used clinical scales and a robotic assay to measure arm movement in 208 patients 7, 14, 21, 30 and 90 days after acute ischemic stroke at two separate clinical sites. The robots are low impedance and low friction interactive devices that precisely measure speed, position and force, so that even a hemiparetic patient can generate a complete measurement profile. These profiles were used to develop predictive models of the clinical assessments employing a combination of artificial ant colonies and neural network ensembles. Results The resulting models replicated commonly used clinical scales to a cross-validated R2 of 0.73, 0.75, 0.63 and 0.60 for the Fugl-Meyer, Motor Power, NIH stroke and modified Rankin scales, respectively. Moreover, when suitably scaled and combined, the robotic measures demonstrated a significant increase in effect size from day 7 to 90 over historical data (1.47 versus 0.67). Discussion and conclusion These results suggest that it is possible to derive surrogate biomarkers that can significantly reduce the sample size required to power future stroke clinical trials.

Dense cellular segmentation for EM using 2D–3D neural network ensembles

Biologists who use electron microscopy (EM) images to build nanoscale 3D models of whole cells and their organelles have historically been limited to small numbers of cells and cellular features due to constraints in imaging and analysis. This has been a major factor limiting insight into the complex variability of cellular environments. Modern EM can produce gigavoxel image volumes containing large numbers of cells, but accurate manual segmentation of image features is slow and limits the creation of cell models. Segmentation algorithms based on convolutional neural networks can process large volumes quickly, but achieving EM task accuracy goals often challenges current techniques. Here, we define dense cellular segmentation as a multiclass semantic segmentation task for modeling cells and large numbers of their organelles, and give an example in human blood platelets. We present an algorithm using novel hybrid 2D–3D segmentation networks to produce dense cellular segmentations with accuracy levels that outperform baseline methods and approach those of human annotators. To our knowledge, this work represents the first published approach to automating the creation of cell models with this level of structural detail.