Over the past decade, Deep Convolutional Neural Networks (DCNNs) have shown remarkable performance in most computer vision tasks. These tasks traditionally use a fixed dataset, and the model, once trained, is deployed as is. Adding new information to such a model presents a challenge due to complex training issues, such as "catastrophic forgetting", and sensitivity to hyperparameter tuning. However, in this modern world, data is constantly evolving, and our deep learning models are required to adapt to these changes. In this paper, we propose an adaptive hierarchical network structure composed of DCNNs that can grow and learn as new data becomes available. The network grows in a tree-like fashion to accommodate new classes of data, while preserving the ability to distinguish the previously trained classes. The network organizes the incrementally available data into feature-driven superclasses and improves upon existing hierarchical CNN models by adding the capability of self-growth. The proposed hierarchical model, when compared against fine-tuning a deep network, achieves significant reduction of training effort, while maintaining competitive accuracy on CIFAR-10 and CIFAR-100.

The advances in the field of machine learning using neuromorphic systems have paved the pathway for extensive research on possibilities of hardware implementations of neural networks. Various memristive technologies such as oxide-based devices, spintronics and phase change materials have been explored to implement the core functional units of neuromorphic systems, namely the synaptic network, and the neuronal functionality, in a fast and energy efficient manner. However, various non-idealities in the crossbar implementations of the synaptic arrays can significantly degrade performance of neural networks and hence, impose restrictions on feasible crossbar sizes. In this work, we build mathematical models of various non-idealities that occur in crossbar implementations such as source resistance, neuron resistance and chip-to-chip device variations and analyze their impact on the classification accuracy of a fully connected network (FCN) and convolutional neural network (CNN) trained with standard training algorithm. We show that a network trained under ideal conditions can suffer accuracy degradation as large as 59.84% for FCNs and 62.4% for CNNs when implemented on non-ideal crossbars for relevant non-ideality ranges. This severely constrains the sizes for crossbars. As a solution, we propose a technology aware training algorithm which incorporates the mathematical models of the non-idealities in the standard training algorithm. We demonstrate that our proposed methodology achieves significant recovery of testing accuracy within 1.9% of the ideal accuracy for FCNs and 1.5% for CNNs. We further show that our proposed training algorithm can potentially allow the use of significantly larger crossbar arrays of sizes 784×500 for FCNs and 4096×512 for CNNs with a minor or no trade-off in accuracy.

Deep neural networks are biologically-inspired class of algorithms that have recently demonstrated state-of-the-art accuracy in large scale classification and recognition tasks. Hardware acceleration of deep networks is of paramount importance to ensure their ubiquitous presence in future computing platforms. Indeed, a major landmark that enables efficient hardware accelerators for deep networks is the recent advances from the machine learning community that have demonstrated the viability of aggressively scaled deep binary networks. In this paper, we demonstrate how deep binary networks can be accelerated in modified von-Neumann machines by enabling binary convolutions within the SRAM array. In general, binary convolutions consist of bit-wise XNOR followed by a populationcount (popcount). We present two proposals − one based on charge sharing approach to perform vector XNORs and approximate popcount and another based on bit-wsie XNORs followed by a digital bit-tree adder for accurate popcount. We highlight the various trade-offs in terms of circuit complexity, speed-up and classification accuracy for both the approaches. Few key techniques presented as a part of the manuscript is use of low-precision, low overhead ADC, to achieve a fairly accurate popcount for the charge-sharing scheme and proposal for sectioning of the SRAM array by adding switches onto the read-bitlines, thereby achieving improved parallelism. Our results on a benchmark image classification dataset CIFAR-10 on a binarized neural network architecture show energy improvements of 6.1× and 2.3× for the two proposals, compared to conventional SRAM banks. In terms of latency, improvements of 15.8× and 8.1× were achieved for the two respective proposals.

No abstract

No abstract

No abstract

Spiking neural networks (SNNs) offer a promising alternative to current artificial neural networks to enable low-power event-driven neuromorphic hardware. Spike-based neuromorphic applications require processing and extracting meaningful information from spatio-temporal data, represented as series of spike trains over time. In this paper, we propose a method to synthesize images from multiple modalities in a spike-based environment. We use spiking auto-encoders to convert image and audio inputs into compact spatio-temporal representations that is then decoded for image synthesis. For this, we use a direct training algorithm that computes loss on the membrane potential of the output layer and back-propagates it by using a sigmoid approximation of the neuron's activation function to enable differentiability. The spiking autoencoders are benchmarked on MNIST and Fashion-MNIST and achieve very low reconstruction loss, comparable to ANNs. Then, spiking autoencoders are trained to learn meaningful spatio-temporal representations of the data, across the two modalities—audio and visual. We synthesize images from audio in a spike-based environment by first generating, and then utilizing such shared multi-modal spatio-temporal representations. Our audio to image synthesis model is tested on the task of converting TI-46 digits audio samples to MNIST images. We are able to synthesize images with high fidelity and the model achieves competitive performance against ANNs.

No abstract