Avoiding pathologies in very deep networks

Choosing appropriate architectures and regularization strategies of deep networks is crucial to good predictive performance. To shed light on this problem, we analyze the analogous problem of constructing useful priors on compositions of functions. Specifically, we study the deep Gaussian process, a type of infinitely-wide, deep neural network. We show that in standard architectures, the representational capacity of the network tends to capture fewer degrees of freedom as the number of layers increases, retaining only a single degree of freedom in the limit. We propose an alternate network architecture which does not suffer from this pathology. We also examine deep covariance functions, obtained by composing infinitely many feature transforms. Lastly, we characterize the class of models obtained by performing dropout on Gaussian processes.

An important pitfall of pseudo-static finite element analysis

Finite Element (FE) pseudo-static analysis can provide a good compromise between simplified methods of dynamic analysis and time domain analysis. The pseudo-static FE approach can accurately model the in situ, stresses prior to seismic loading (when it follows a static analysis simulating the construction sequence) is relatively simple and not as computationally expensive as the time domain approach. However this method should be used with caution as the results can be sensitive to the choice of the mesh dimensions. In this paper two simple examples of pseudo-static finite element analysis are examined parametrically, a homogeneous slope and a cantilever retaining wall, exploring the sensitivity of the pseudo-static analysis results on the adopted mesh size. The mesh dependence was found to be more pronounced for problems with high critical seismic coefficients values (e.g. gentle slopes or small walls), as in these cases a generalised layer failure mechanism is developed simultaneously with the slope or wall mechanism. In general the mesh width was found not to affect notably the predicted value of critical seismic coefficient but to have a major impact on the predicted movements. © 2012 Elsevier Ltd.

Amorphous silicon thin film transistor biosensing system

Electronic systems are a very good platform for sensing biological signals for fast point-of-care diagnostics or threat detection. One of the solutions is the lab-on-a-chip integrated circuit (IC), which is low cost and high reliability, offering the possibility for label-free detection. In recent years, similar integrated biosensors based on the conventional complementary metal oxide semiconductor (CMOS) technology have been reported. However, post-fabrication processes are essential for all classes of CMOS biochips, requiring biocompatible electrode deposition and circuit encapsulation. In this work, we present an amorphous silicon (a-Si) thin film transistor (TFT) array based sensing approach, which greatly simplifies the fabrication procedures and even decreases the cost of the biosensor. The device contains several identical sensor pixels with amplifiers to boost the sensitivity. Ring oscillator and logic circuits are also integrated to achieve different measurement methodologies, including electro-analytical methods such as amperometric and cyclic voltammetric modes. The system also supports different operational modes. For example, depending on the required detection arrangement, a sample droplet could be placed on the sensing pads or the device could be immersed into the sample solution for real time in-situ measurement. The entire system is designed and fabricated using a low temperature TFT process that is compatible to plastic substrates. No additional processing is required prior to biological measurement. A Cr/Au double layer is used for the biological-electronic interface. The success of the TFT-based system used in this work will open new avenues for flexible label-free or low-cost disposable biosensors. © 2013 Materials Research Society.