When building a unified vision system or gradually adding new capabilities to a system, the usual assumption is that training data for all tasks is always available. However, as the number of tasks grows, storing and retraining on such data becomes infeasible. A new problem arises where we add new capabilities to a Convolutional Neural Network (CNN), but the training data for its existing capabilities are unavailable. We propose our Learning without Forgetting method, which uses only new task data to train the network while preserving the original capabilities. Our method performs favorably compared to commonly used feature extraction and fine-tuning adaption techniques and performs similarly to multitask learning that uses original task data we assume unavailable. A more surprising observation is that Learning without Forgetting may be able to replace fine-tuning with similar old and new task datasets for improved new task performance.

Learning without Forgetting

Guided by brain-like ‘spiking’ computational frameworks, neuromorphic computing—brain-inspired computing for machine intelligence—promises to realize artificial intelligence while reducing the energy requirements of computing platforms. This interdisciplinary field began with the implementation of silicon circuits for biological neural routines, but has evolved to encompass the hardware implementation of algorithms with spike-based encoding and event-driven representations. Here we provide an overview of the developments in neuromorphic computing for both algorithms and hardware and highlight the fundamentals of learning and hardware frameworks. We discuss the main challenges and the future prospects of neuromorphic computing, with emphasis on algorithm–hardware codesign. The authors review the advantages and future prospects of neuromorphic computing, a multidisciplinary engineering concept for energy-efficient artificial intelligence with brain-inspired functionality.

Towards spike-based machine intelligence with neuromorphic computing.

Deep learning in spiking neural networks

Spiking neural networks (SNNs) can potentially offer an efficient way of doing inference because the neurons in the networks are sparsely activated and computations are event-driven. Previous work showed that simple continuous-valued deep Convolutional Neural Networks (CNNs) can be converted into accurate spiking equivalents. These networks did not include certain common operations such as max-pooling, softmax, batch-normalization and Inception-modules. This paper presents spiking equivalents of these operations therefore allowing conversion of nearly arbitrary CNN architectures. We show conversion of popular CNN architectures, including VGG-16 and Inception-v3, into SNNs that produce the best results reported to date on MNIST, CIFAR-10 and the challenging ImageNet dataset. SNNs can trade off classification error rate against the number of available operations whereas deep continuous-valued neural networks require a fixed number of operations to achieve their classification error rate. From the examples of LeNet for MNIST and BinaryNet for CIFAR-10, we show that with an increase in error rate of a few percentage points, the SNNs can achieve more than 2x reductions in operations compared to the original CNNs. This highlights the potential of SNNs in particular when deployed on power-efficient neuromorphic spiking neuron chips, for use in embedded applications.

/pdf/conversion-of-continuous-valued-deep-networks-to-efficient-1ltkloo6ev.pdf

Conversion of Continuous-Valued Deep Networks to Efficient Event-Driven Networks for Image Classification.

Event cameras are bio-inspired sensors that differ from conventional frame cameras: Instead of capturing images at a fixed rate, they asynchronously measure per-pixel brightness changes, and output a stream of events that encode the time, location and sign of the brightness changes. Event cameras offer attractive properties compared to traditional cameras: high temporal resolution (in the order of is), very high dynamic range (140dB vs. 60dB), low power consumption, and high pixel bandwidth (on the order of kHz) resulting in reduced motion blur. Hence, event cameras have a large potential for robotics and computer vision in challenging scenarios for traditional cameras, such as low-latency, high speed, and high dynamic range. However, novel methods are required to process the unconventional output of these sensors in order to unlock their potential. This paper provides a comprehensive overview of the emerging field of event-based vision, with a focus on the applications and the algorithms developed to unlock the outstanding properties of event cameras. We present event cameras from their working principle, the actual sensors that are available and the tasks that they have been used for, from low-level vision (feature detection and tracking, optic flow, etc.) to high-level vision (reconstruction, segmentation, recognition). We also discuss the techniques developed to process events, including learning-based techniques, as well as specialized processors for these novel sensors, such as spiking neural networks. Additionally, we highlight the challenges that remain to be tackled and the opportunities that lie ahead in the search for a more efficient, bio-inspired way for machines to perceive and interact with the world.

Event-based Vision: A Survey

Benchmarks have played a vital role in the advancement of visual object recognition and other fields of computer vision (LeCun et al., 1998; Deng et al., 2009;). The challenges posed by these standard datasets have helped identify and overcome the shortcomings of existing approaches, and have led to great advances of the state of the art. Even the recent massive increase of interest in deep learning methods can be attributed to their success in difficult benchmarks such as ImageNet (Krizhevsky et al., 2012; LeCun et al., 2015). Neuromorphic vision uses silicon retina sensors such as the dynamic vision sensor (DVS; Lichtsteiner et al., 2008). These sensors and their DAVIS (Dynamic and Active-pixel Vision Sensor) and ATIS (Asynchronous Time-based Image Sensor) derivatives (Brandli et al., 2014; Posch et al., 2014) are inspired by biological vision by generating streams of asynchronous events indicating local log-intensity brightness changes. They thereby greatly reduce the amount of data to be processed, and their dynamic nature makes them a good fit for domains such as optical flow, object tracking, action recognition, or dynamic scene understanding. Compared to classical computer vision, neuromorphic vision is a younger and much smaller field of research, and lacks benchmarks, which impedes the progress of the field. To address this we introduce the largest event-based vision benchmark dataset published to date, hoping to satisfy a growing demand and stimulate challenges for the community. In particular, the availability of such benchmarks should help the development of algorithms processing event-based vision input, allowing a direct fair comparison of different approaches. We have explicitly chosen mostly dynamic vision tasks such as action recognition or tracking, which could benefit from the strengths of neuromorphic vision sensors, although algorithms that exploit these features are largely missing.

A major reason for the lack of benchmarks is that currently neuromorphic vision sensors are only available as RD see Tan et al. (2015) for an informative review. Unlabeled DVS data was made available around 2007 in the jAER project1 and was used for development of spike timing-based unsupervised feature learning e.g., in Bichler et al. (2012). The first labeled and published event-based neuromorphic vision sensor benchmarks were created from the MNIST digit recognition dataset by jiggling the image on the screen (see Serrano-Gotarredona and Linares-Barranco, 2015 for an informative history) and later to reduce frame artifacts by jiggling the camera view with a pan-tilt unit (Orchard et al., 2015). These datasets automated the scene movement necessary to generate DVS output from the static images, and will be an important step forward for evaluating neuromorphic object recognition systems such as spiking deep networks (Perez-Carrasco et al., 2013; O'Connor et al., 2013; Cao et al., 2014; Diehl et al., 2015), which so far have been tested mostly on static image datasets converted into Poisson spike trains. But static image recognition is not the ideal use case for event-based vision sensors that are designed for dynamic scenes. Recently several additional DVS datasets were made available in the Frontiers research topic “Benchmarks and Challenges for Neuromorphic Engineering”2; in particular for navigation using multiple sensor modalities (Barranco et al., 2016) and for developing and benchmarking DVS and DAVIS optical flow methods (Rueckauer and Delbruck, 2016).

This data report summarizes a new benchmark dataset in which we converted established visual video benchmarks for object tracking, action recognition and object recognition into spiking neuromorphic datasets, recorded with the DVS output (Lichtsteiner et al., 2008) of a DAVIS camera (Berner et al., 2013; Brandli et al., 2014). This report presents our approach for sensor calibration and capture of frame-based videos into neuromorphic vision datasets with minimal human intervention. We converted four widely used dynamic datasets: the VOT Challenge 2015 Dataset (Kristan et al., 2016), TrackingDataset3, the UCF-50 Action Recognition Dataset (Reddy and Shah, 2012), and the Caltech-256 Object Category Dataset (Griffin et al., 2006). We conclude with statistics and summaries of the datasets.

/pdf/dvs-benchmark-datasets-for-object-tracking-action-4wmtf20uq4.pdf

DVS Benchmark Datasets for Object Tracking, Action Recognition, and Object Recognition.

To help meet the increasing need for dynamic vision sensor (DVS) event camera data, this paper proposes the v2e toolbox that generates realistic synthetic DVS events from intensity frames. It also clarifies incorrect claims about DVS motion blur and latency characteristics in recent literature. Unlike other toolboxes, v2e includes pixel-level Gaussian event threshold mismatch, finite intensity-dependent bandwidth, and intensity-dependent noise. Realistic DVS events are useful in training networks for uncontrolled lighting conditions. The use of v2e synthetic events is demonstrated in two experiments. The first experiment is object recognition with N-Caltech 101 dataset. Results show that pretraining on various v2e lighting conditions improves generalization when transferred on real DVS data for a ResNet model. The second experiment shows that for night driving, a car detector trained with v2e events shows an average accuracy improvement of 40% compared to the YOLOv3 trained on intensity frames.

/pdf/v2e-from-video-frames-to-realistic-dvs-events-245qvsxisf.pdf

v2e: From Video Frames to Realistic DVS Events

Deep convolutional neural networks (CNNs) have shown great potential for numerous real-world machine learning applications, but performing inference in large CNNs in real-time remains a challenge. We have previously demonstrated that traditional CNNs can be converted into deep spiking neural networks (SNNs), which exhibit similar accuracy while reducing both latency and computational load as a consequence of their data-driven, event-based style of computing. Here we provide a novel theory that explains why this conversion is successful, and derive from it several new tools to convert a larger and more powerful class of deep networks into SNNs. We identify the main sources of approximation errors in previous conversion methods, and propose simple mechanisms to fix these issues. Furthermore, we develop spiking implementations of common CNN operations such as max-pooling, softmax, and batch-normalization, which allow almost loss-less conversion of arbitrary CNN architectures into the spiking domain. Empirical evaluation of different network architectures on the MNIST and CIFAR10 benchmarks leads to the best SNN results reported to date.

Theory and Tools for the Conversion of Analog to Spiking Convolutional Neural Networks

Plain recurrent networks greatly suffer from the vanishing gradient problem while Gated Neural Networks (GNNs) such as Long-short Term Memory (LSTM) and Gated Recurrent Unit (GRU) deliver promising results in many sequence learning tasks through sophisticated network designs. This paper shows how we can address this problem in a plain recurrent network by analyzing the gating mechanisms in GNNs. We propose a novel network called the Recurrent Identity Network (RIN) which allows a plain recurrent network to overcome the vanishing gradient problem while training very deep models without the use of gates. We compare this model with IRNNs and LSTMs on multiple sequence modeling benchmarks. The RINs demonstrate competitive performance and converge faster in all tasks. Notably, small RIN models produce 12%--67% higher accuracy on the Sequential and Permuted MNIST datasets and reach state-of-the-art performance on the bAbI question answering dataset.

/pdf/overcoming-the-vanishing-gradient-problem-in-plain-recurrent-uo70kxftzk.pdf

Yuhuang Hu

Papers

Conversion of Continuous-Valued Deep Networks to Efficient Event-Driven Networks for Image Classification.

DVS Benchmark Datasets for Object Tracking, Action Recognition, and Object Recognition.

v2e: From Video Frames to Realistic DVS Events

Theory and Tools for the Conversion of Analog to Spiking Convolutional Neural Networks

Overcoming the vanishing gradient problem in plain recurrent networks