Shih-Chii Liu

Qinyu Chen, Zuowen Wang, Shih-Chii Liu et al.

This paper presents a sparse Change-Based Convolutional Long Short-Term Memory (CB-ConvLSTM) model for event-based eye tracking, key for next-generation wearable healthcare technology such as AR/VR headsets. We leverage the benefits of retina-inspired event cameras, namely their low-latency response and sparse output event stream, over traditional frame-based cameras. Our CB-ConvLSTM architecture efficiently extracts spatio-temporal features for pupil tracking from the event stream, outperforming conventional CNN structures. Utilizing a delta-encoded recurrent path enhancing activation sparsity, CB-ConvLSTM reduces arithmetic operations by approximately 4.7$\times$ without losing accuracy when tested on a \texttt{v2e}-generated event dataset of labeled pupils. This increase in efficiency makes it ideal for real-time eye tracking in resource-constrained devices. The project code and dataset are openly available at \url{https://github.com/qinche106/cb-convlstm-eyetracking}.

Jason Yik, Korneel Van den Berghe, Douwe den Blanken et al. · eth-zurich

Neuromorphic computing shows promise for advancing computing efficiency and capabilities of AI applications using brain-inspired principles. However, the neuromorphic research field currently lacks standardized benchmarks, making it difficult to accurately measure technological advancements, compare performance with conventional methods, and identify promising future research directions. Prior neuromorphic computing benchmark efforts have not seen widespread adoption due to a lack of inclusive, actionable, and iterative benchmark design and guidelines. To address these shortcomings, we present NeuroBench: a benchmark framework for neuromorphic computing algorithms and systems. NeuroBench is a collaboratively-designed effort from an open community of researchers across industry and academia, aiming to provide a representative structure for standardizing the evaluation of neuromorphic approaches. The NeuroBench framework introduces a common set of tools and systematic methodology for inclusive benchmark measurement, delivering an objective reference framework for quantifying neuromorphic approaches in both hardware-independent (algorithm track) and hardware-dependent (system track) settings. In this article, we outline tasks and guidelines for benchmarks across multiple application domains, and present initial performance baselines across neuromorphic and conventional approaches for both benchmark tracks. NeuroBench is intended to continually expand its benchmarks and features to foster and track the progress made by the research community.

Simon Narduzzi, Siavash A. Bigdeli, Shih-Chii Liu et al.

Reducing energy consumption is a critical point for neural network models running on edge devices. In this regard, reducing the number of multiply-accumulate (MAC) operations of Deep Neural Networks (DNNs) running on edge hardware accelerators will reduce the energy consumption during inference. Spiking Neural Networks (SNNs) are an example of bio-inspired techniques that can further save energy by using binary activations, and avoid consuming energy when not spiking. The networks can be configured for equivalent accuracy on a task through DNN-to-SNN conversion frameworks but their conversion is based on rate coding therefore the synaptic operations can be high. In this work, we look into different techniques to enforce sparsity on the neural network activation maps and compare the effect of different training regularizers on the efficiency of the optimized DNNs and SNNs.

Longbiao Cheng, Ashutosh Pandey, Buye Xu et al.

This paper introduces a new Dynamic Gated Recurrent Neural Network (DG-RNN) for compute-efficient speech enhancement models running on resource-constrained hardware platforms. It leverages the slow evolution characteristic of RNN hidden states over steps, and updates only a selected set of neurons at each step by adding a newly proposed select gate to the RNN model. This select gate allows the computation cost of the conventional RNN to be reduced during network inference. As a realization of the DG-RNN, we further propose the Dynamic Gated Recurrent Unit (D-GRU) which does not require additional parameters. Test results obtained from several state-of-the-art compute-efficient RNN-based speech enhancement architectures using the DNS challenge dataset, show that the D-GRU based model variants maintain similar speech intelligibility and quality metrics comparable to the baseline GRU based models even with an average 50% reduction in GRU computes.

Yuhuang Hu, Shih-Chii Liu

Deep Neural Networks, particularly Convolutional Neural Networks (ConvNets), have achieved incredible success in many vision tasks, but they usually require millions of parameters for good accuracy performance. With increasing applications that use ConvNets, updating hundreds of networks for multiple tasks on an embedded device can be costly in terms of memory, bandwidth, and energy. Approaches to reduce this cost include model compression and parameter-efficient models that adapt a subset of network layers for each new task. This work proposes a novel parameter-efficient kernel modulation (KM) method that adapts all parameters of a base network instead of a subset of layers. KM uses lightweight task-specialized kernel modulators that require only an additional 1.4% of the base network parameters. With multiple tasks, only the task-specialized KM weights are communicated and stored on the end-user device. We applied this method in training ConvNets for Transfer Learning and Meta-Learning scenarios. Our results show that KM delivers up to 9% higher accuracy than other parameter-efficient methods on the Transfer Learning benchmark.

Pehuén Moure, Niclas Pokel, Bilal Bounajma et al.

Automatic speech recognition (ASR) systems remain brittle on dysarthric and other atypical speech. Recent audio-language models raise the possibility of improving performance by conditioning on additional clinical context at inference time, but it is unclear whether these models can make use of such information. We introduce a benchmark built on the Speech Accessibility Project (SAP) dataset that tests whether diagnosis labels, clinician-derived speech ratings, and progressively richer clinical descriptions improve transcription accuracy for dysarthric speech. Across matched comparisons on nine models, we find that current models do not meaningfully use this context: diagnosis-informed and clinically detailed prompts yield negligible improvements and often degrade word error rate. We complement the prompting analysis with context-dependent fine-tuning, showing that LoRA adaptation with a mixture of clinical prompt formats achieves a WER of 0.066, a 52% relative reduction over the frozen baseline, while preserving performance when context is unavailable. Subgroup analyses reveal significant gains for Down syndrome and mild-severity speakers. These results clarify where current models fall short and provide a testbed for measuring progress toward more inclusive ASR.

Zuowen Wang, Chang Gao, Zongwei Wu et al.

This survey reviews the AIS 2024 Event-Based Eye Tracking (EET) Challenge. The task of the challenge focuses on processing eye movement recorded with event cameras and predicting the pupil center of the eye. The challenge emphasizes efficient eye tracking with event cameras to achieve good task accuracy and efficiency trade-off. During the challenge period, 38 participants registered for the Kaggle competition, and 8 teams submitted a challenge factsheet. The novel and diverse methods from the submitted factsheets are reviewed and analyzed in this survey to advance future event-based eye tracking research.

Gonçalo Guiomar, Elia Torre, Pehuen Moure et al.

Do language models make decisions under uncertainty like humans do, and what role does chain-of-thought (CoT) reasoning play in the underlying decision process? We introduce an active probabilistic reasoning task that cleanly separates sampling (actively acquiring evidence) from inference (integrating evidence toward a decision). Benchmarking humans and a broad set of contemporary large language models against near-optimal reference policies reveals a consistent pattern: extended reasoning is the key determinant of strong performance, driving large gains in inference and producing belief trajectories that become strikingly human-like, while yielding only modest improvements in active sampling. To explain these differences, we fit a mechanistic model that captures systematic deviations from optimal behavior via four interpretable latent variables: memory, strategy, choice bias, and occlusion awareness. This model places humans and models in a shared low-dimensional cognitive space, reproduces behavioral signatures across agents, and shows how chain-of-thought shifts language models toward human-like regimes of evidence accumulation and belief-to-choice mapping, tightening alignment in inference while leaving a persistent gap in information acquisition.

Joachim Ott, Shih-Chii Liu

This work proposes a model for continual learning on tasks involving temporal sequences, specifically, human motions. It improves on a recently proposed brain-inspired replay model (BI-R) by building a biologically-inspired conditional temporal variational autoencoder (BI-CTVAE), which instantiates a latent mixture-of-Gaussians for class representation. We investigate a novel continual-learning-to-generate (CL2Gen) scenario where the model generates motion sequences of different classes. The generative accuracy of the model is tested over a set of tasks. The final classification accuracy of BI-CTVAE on a human motion dataset after sequentially learning all action classes is 78%, which is 63% higher than using no-replay, and only 5.4% lower than a state-of-the-art offline trained GRU model.

Longbiao Cheng, Ashutosh Pandey, Buye Xu et al.

Deep learning-based speech enhancement (SE) methods often face significant computational challenges when needing to meet low-latency requirements because of the increased number of frames to be processed. This paper introduces the SlowFast framework which aims to reduce computation costs specifically when low-latency enhancement is needed. The framework consists of a slow branch that analyzes the acoustic environment at a low frame rate, and a fast branch that performs SE in the time domain at the needed higher frame rate to match the required latency. Specifically, the fast branch employs a state space model where its state transition process is dynamically modulated by the slow branch. Experiments on a SE task with a 2 ms algorithmic latency requirement using the Voice Bank + Demand dataset show that our approach reduces computation cost by 70% compared to a baseline single-branch network with equivalent parameters, without compromising enhancement performance. Furthermore, by leveraging the SlowFast framework, we implemented a network that achieves an algorithmic latency of just 62.5 μs (one sample point at 16 kHz sample rate) with a computation cost of 100 M MACs/s, while scoring a PESQ-NB of 3.12 and SISNR of 16.62.

Yuanxi Wang, Zuowen Wang, Shih-Chii Liu

This paper presents an efficient deep learning solution for decoding motor movements from neural recordings in non-human primates. An Autoencoder Gated Recurrent Unit (AEGRU) model was adopted as the model architecture for this task. The autoencoder is only used during the training stage to achieve better generalization. Together with the preprocessing techniques, our model achieved 0.71 $R^2$ score, surpassing the baseline models in Neurobench and is ranked first for $R^2$ in the IEEE BioCAS 2024 Grand Challenge on Neural Decoding. Model pruning is also applied leading to a reduction of 41.4% of the multiply-accumulate (MAC) operations with little change in the $R^2$ score compared to the unpruned model.

Xi Chen, Chang Gao, Zuowen Wang et al.

Recurrent Neural Networks (RNNs) are useful in temporal sequence tasks. However, training RNNs involves dense matrix multiplications which require hardware that can support a large number of arithmetic operations and memory accesses. Implementing online training of RNNs on the edge calls for optimized algorithms for an efficient deployment on hardware. Inspired by the spiking neuron model, the Delta RNN exploits temporal sparsity during inference by skipping over the update of hidden states from those inactivated neurons whose change of activation across two timesteps is below a defined threshold. This work describes a training algorithm for Delta RNNs that exploits temporal sparsity in the backward propagation phase to reduce computational requirements for training on the edge. Due to the symmetric computation graphs of forward and backward propagation during training, the gradient computation of inactivated neurons can be skipped. Results show a reduction of $\sim$80% in matrix operations for training a 56k parameter Delta LSTM on the Fluent Speech Commands dataset with negligible accuracy loss. Logic simulations of a hardware accelerator designed for the training algorithm show 2-10X speedup in matrix computations for an activation sparsity range of 50%-90%. Additionally, we show that the proposed Delta RNN training will be useful for online incremental learning on edge devices with limited computing resources.

Longbiao Cheng, Shih-Chii Liu

Recent studies have shown that post-deployment adaptation can improve the robustness of speech enhancement models in unseen noise conditions. However, existing methods often incur prohibitive computational and memory costs, limiting their suitability for on-device deployment. In this work, we investigate model adaptation in realistic settings with dynamic acoustic scene changes and propose a lightweight framework that augments a frozen backbone with low-rank adapters updated via self-supervised training. Experiments on sequential scene evaluations spanning 111 environments across 37 noise types and three signal-to-noise ratio ranges, including the challenging [-8, 0] dB range, show that our method updates fewer than 1% of the base model's parameters while achieving an average 1.51 dB SI-SDR improvement within only 20 updates per scene. Compared to state-of-the-art approaches, our framework achieves competitive or superior perceptual quality with smoother and more stable convergence, demonstrating its practicality for lightweight on-device adaptation of speech enhancement models under real-world acoustic conditions.

Niclas Pokel, Pehuén Moure, Roman Boehringer et al.

Speech impairments resulting from congenital disorders, such as cerebral palsy, down syndrome, or apert syndrome, as well as acquired brain injuries due to stroke, traumatic accidents, or tumors, present major challenges to automatic speech recognition (ASR) systems. Despite recent advancements, state-of-the-art ASR models like Whisper still struggle with non-normative speech due to limited training data availability and high acoustic variability. Moreover, collecting and annotating non-normative speech is burdensome: speaking is effortful for many affected individuals, while laborious annotation often requires caregivers familiar with the speaker. This work introduces a novel ASR personalization method based on Bayesian Low-rank Adaptation for data-efficient fine-tuning. We validate our method on the English UA-Speech dataset and a newly collected German speech dataset, BF-Sprache, from a child with structural speech impairment. The dataset and approach are designed to reflect the challenges of low-resource settings that include individuals with speech impairments. Our method significantly improves ASR accuracy for impaired speech while maintaining data and annotation efficiency, offering a practical path toward inclusive ASR.

Joachim Ott, Zuowen Wang, Shih-Chii Liu

Event cameras are advantageous for tasks that require vision sensors with low-latency and sparse output responses. However, the development of deep network algorithms using event cameras has been slow because of the lack of large labelled event camera datasets for network training. This paper reports a method for creating new labelled event datasets by using a text-to-X model, where X is one or multiple output modalities, in the case of this work, events. Our proposed text-to-events model produces synthetic event frames directly from text prompts. It uses an autoencoder which is trained to produce sparse event frames representing event camera outputs. By combining the pretrained autoencoder with a diffusion model architecture, the new text-to-events model is able to generate smooth synthetic event streams of moving objects. The autoencoder was first trained on an event camera dataset of diverse scenes. In the combined training with the diffusion model, the DVS gesture dataset was used. We demonstrate that the model can generate realistic event sequences of human gestures prompted by different text statements. The classification accuracy of the generated sequences, using a classifier trained on the real dataset, ranges between 42% to 92%, depending on the gesture group. The results demonstrate the capability of this method in synthesizing event datasets.

Qinyu Chen, Kwantae Kim, Chang Gao et al.

This paper introduces DeltaKWS, to the best of our knowledge, the first $Δ$RNN-enabled fine-grained temporal sparsity-aware KWS IC for voice-controlled devices. The 65 nm prototype chip features a number of techniques to enhance performance, area, and power efficiencies, specifically: 1) a bio-inspired delta-gated recurrent neural network ($Δ$RNN) classifier leveraging temporal similarities between neighboring feature vectors extracted from input frames and network hidden states, eliminating unnecessary operations and memory accesses; 2) an IIR BPF-based FEx that leverages mixed-precision quantization, low-cost computing structure and channel selection; 3) a 24 kB 0.6 V near-$V_\text{TH}$ weight SRAM that achieves 6.6X lower read power than the foundry-provided SRAM. From chip measurement results, we show that the DeltaKWS achieves an 11/12-class GSCD accuracy of 90.5%/89.5% respectively and energy consumption of 36 nJ/decision in 65 nm CMOS process. At 87% temporal sparsity, computing latency and energy/inference are reduced by 2.4X/3.4X, respectively. The IIR BPF-based FEx, $Δ$RNN accelerator, and 24 kB near-$V_\text{TH}$ SRAM blocks occupy 0.084 mm$^{2}$, 0.319 mm$^{2}$, and 0.381 mm$^{2}$ respectively (0.78 mm$^{2}$ in total).

Qinyu Chen, Congyi Sun, Chang Gao et al.

Epilepsy is a common disease of the nervous system. Timely prediction of seizures and intervention treatment can significantly reduce the accidental injury of patients and protect the life and health of patients. This paper presents a neuromorphic Spiking Convolutional Transformer, named Spiking Conformer, to detect and predict epileptic seizure segments from scalped long-term electroencephalogram (EEG) recordings. We report evaluation results from the Spiking Conformer model using the Boston Children's Hospital-MIT (CHB-MIT) EEG dataset. By leveraging spike-based addition operations, the Spiking Conformer significantly reduces the classification computational cost compared to the non-spiking model. Additionally, we introduce an approximate spiking neuron layer to further reduce spike-triggered neuron updates by nearly 38% without sacrificing accuracy. Using raw EEG data as input, the proposed Spiking Conformer achieved an average sensitivity rate of 94.9% and a specificity rate of 99.3% for the seizure detection task, and 96.8%, 89.5% for the seizure prediction task, and needs >10x fewer operations compared to the non-spiking equivalent model.

Ilya Kiselev, Chang Gao, Shih-Chii Liu

Including local automatic gain control (AGC) circuitry into a silicon cochlea design has been challenging because of transistor mismatch and model complexity. To address this, we present an alternative system-level algorithm that implements channel-specific AGC in a silicon spiking cochlea by measuring the output spike activity of individual channels. The bandpass filter gain of a channel is adapted dynamically to the input amplitude so that the average output spike rate stays within a defined range. Because this AGC mechanism only needs counting and adding operations, it can be implemented at low hardware cost in a future design. We evaluate the impact of the local AGC algorithm on a classification task where the input signal varies over 32 dB input range. Two classifier types receiving cochlea spike features were tested on a speech versus noise classification task. The logistic regression classifier achieves an average of 6% improvement and 40.8% relative improvement in accuracy when the AGC is enabled. The deep neural network classifier shows a similar improvement for the AGC case and achieves a higher mean accuracy of 96% compared to the best accuracy of 91% from the logistic regression classifier.

Zuowen Wang, Yuhuang Hu, Shih-Chii Liu

Event cameras report local changes of brightness through an asynchronous stream of output events. Events are spatially sparse at pixel locations with little brightness variation. We propose using a visual transformer (ViT) architecture to leverage its ability to process a variable-length input. The input to the ViT consists of events that are accumulated into time bins and spatially separated into non-overlapping sub-regions called patches. Patches are selected when the number of nonzero pixel locations within a sub-region is above a threshold. We show that by fine-tuning a ViT model on the selected active patches, we can reduce the average number of patches fed into the backbone during the inference by at least 50% with only a minor drop (0.34%) of the classification accuracy on the N-Caltech101 dataset. This reduction translates into a decrease of 51% in Multiply-Accumulate (MAC) operations and an increase of 46% in the inference speed using a server CPU.

Shu Wang, Yuhuang Hu, Shih-Chii Liu

Spiking silicon cochlea sensors encode sound as an asynchronous stream of spikes from different frequency channels. The lack of labeled training datasets for spiking cochleas makes it difficult to train deep neural networks on the outputs of these sensors. This work proposes a self-supervised method called Temporal Network Grafting Algorithm (T-NGA), which grafts a recurrent network pretrained on spectrogram features so that the network works with the cochlea event features. T-NGA training requires only temporally aligned audio spectrograms and event features. Our experiments show that the accuracy of the grafted network was similar to the accuracy of a supervised network trained from scratch on a speech recognition task using events from a software spiking cochlea model. Despite the circuit non-idealities of the spiking silicon cochlea, the grafted network accuracy on the silicon cochlea spike recordings was only about 5% lower than the supervised network accuracy using the N-TIDIGITS18 dataset. T-NGA can train networks to process spiking audio sensor events in the absence of large labeled spike datasets.

Chang Gao, Tobi Delbruck, Shih-Chii Liu

Long Short-Term Memory (LSTM) recurrent networks are frequently used for tasks involving time-sequential data such as speech recognition. Unlike previous LSTM accelerators that either exploit spatial weight sparsity or temporal activation sparsity, this paper proposes a new accelerator called "Spartus" that exploits spatio-temporal sparsity to achieve ultra-low latency inference. Spatial sparsity is induced using a new Column-Balanced Targeted Dropout (CBTD) structured pruning method, producing structured sparse weight matrices for a balanced workload. The pruned networks running on Spartus hardware achieve weight sparsity levels of up to 96% and 94% with negligible accuracy loss on the TIMIT and the Librispeech datasets. To induce temporal sparsity in LSTM, we extend the previous DeltaGRU method to the DeltaLSTM method. Combining spatio-temporal sparsity with CBTD and DeltaLSTM saves on weight memory access and associated arithmetic operations. The Spartus architecture is scalable and supports real-time online speech recognition when implemented on small and large FPGAs. Spartus per-sample latency for a single DeltaLSTM layer of 1024 neurons averages 1 us. Exploiting spatio-temporal sparsity on our test LSTM network using the TIMIT dataset leads to 46X speedup of Spartus over its theoretical hardware performance to achieve 9.4 TOp/s effective batch-1 throughput and 1.1 TOp/s/W power efficiency.

Shih-Chii Liu, John Paul Strachan, Arindam Basu

Operations typically used in machine learning al-gorithms (e.g. adds and soft max) can be implemented bycompact analog circuits. Analog Application-Specific Integrated Circuit (ASIC) designs that implement these algorithms using techniques such as charge sharing circuits and subthreshold transistors, achieve very high power efficiencies. With the recent advances in deep learning algorithms, focus has shifted to hardware digital accelerator designs that implement the prevalent matrix-vector multiplication operations. Power in these designs is usually dominated by the memory access power of off-chip DRAM needed for storing the network weights and activations. Emerging dense non-volatile memory technologies can help to provide on-chip memory and analog circuits can be well suited to implement the needed multiplication-vector operations coupled with in-computing memory approaches. This paper presents abrief review of analog designs that implement various machine learning algorithms. It then presents an outlook for the use ofanalog circuits in low-power deep network accelerators suitable for edge or tiny machine learning applications.

Yuhuang Hu, Shih-Chii Liu, Tobi Delbruck

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.

Yuhuang Hu, Jonathan Binas, Daniel Neil et al.

Neuromorphic event cameras are useful for dynamic vision problems under difficult lighting conditions. To enable studies of using event cameras in automobile driving applications, this paper reports a new end-to-end driving dataset called DDD20. The dataset was captured with a DAVIS camera that concurrently streams both dynamic vision sensor (DVS) brightness change events and active pixel sensor (APS) intensity frames. DDD20 is the longest event camera end-to-end driving dataset to date with 51h of DAVIS event+frame camera and vehicle human control data collected from 4000km of highway and urban driving under a variety of lighting conditions. Using DDD20, we report the first study of fusing brightness change events and intensity frame data using a deep learning approach to predict the instantaneous human steering wheel angle. Over all day and night conditions, the explained variance for human steering prediction from a Resnet-32 is significantly better from the fused DVS+APS frames (0.88) than using either DVS (0.67) or APS (0.77) data alone.

Tobi Delbruck, Shih-Chii Liu

The energy consumed by running large deep neural networks (DNNs) on hardware accelerators is dominated by the need for lots of fast memory to store both states and weights. This large required memory is currently only economically viable through DRAM. Although DRAM is high-throughput and low-cost memory (costing 20X less than SRAM), its long random access latency is bad for the unpredictable access patterns in spiking neural networks (SNNs). In addition, accessing data from DRAM costs orders of magnitude more energy than doing arithmetic with that data. SNNs are energy-efficient if local memory is available and few spikes are generated. This paper reports on our developments over the last 5 years of convolutional and recurrent deep neural network hardware accelerators that exploit either spatial or temporal sparsity similar to SNNs but achieve SOA throughput, power efficiency and latency even with the use of DRAM for the required storage of the weights and states of large DNNs.

Yuhuang Hu, Tobi Delbruck, Shih-Chii Liu

Novel vision sensors such as thermal, hyperspectral, polarization, and event cameras provide information that is not available from conventional intensity cameras. An obstacle to using these sensors with current powerful deep neural networks is the lack of large labeled training datasets. This paper proposes a Network Grafting Algorithm (NGA), where a new front end network driven by unconventional visual inputs replaces the front end network of a pretrained deep network that processes intensity frames. The self-supervised training uses only synchronously-recorded intensity frames and novel sensor data to maximize feature similarity between the pretrained network and the grafted network. We show that the enhanced grafted network reaches competitive average precision (AP50) scores to the pretrained network on an object detection task using thermal and event camera datasets, with no increase in inference costs. Particularly, the grafted network driven by thermal frames showed a relative improvement of 49.11% over the use of intensity frames. The grafted front end has only 5--8% of the total parameters and can be trained in a few hours on a single GPU equivalent to 5% of the time that would be needed to train the entire object detector from labeled data. NGA allows new vision sensors to capitalize on previously pretrained powerful deep models, saving on training cost and widening a range of applications for novel sensors.

Chang Gao, Rachel Gehlhar, Aaron D. Ames et al.

Lower leg prostheses could improve the life quality of amputees by increasing comfort and reducing energy to locomote, but currently control methods are limited in modulating behaviors based upon the human's experience. This paper describes the first steps toward learning complex controllers for dynamical robotic assistive devices. We provide the first example of behavioral cloning to control a powered transfemoral prostheses using a Gated Recurrent Unit (GRU) based recurrent neural network (RNN) running on a custom hardware accelerator that exploits temporal sparsity. The RNN is trained on data collected from the original prosthesis controller. The RNN inference is realized by a novel EdgeDRNN accelerator in real-time. Experimental results show that the RNN can replace the nominal PD controller to realize end-to-end control of the AMPRO3 prosthetic leg walking on flat ground and unforeseen slopes with comparable tracking accuracy. EdgeDRNN computes the RNN about 240 times faster than real time, opening the possibility of running larger networks for more complex tasks in the future. Implementing an RNN on this real-time dynamical system with impacts sets the ground work to incorporate other learned elements of the human-prosthesis system into prosthesis control.

Yi Luo, Enea Ceolini, Cong Han et al.

Beamforming has been extensively investigated for multi-channel audio processing tasks. Recently, learning-based beamforming methods, sometimes called \textit{neural beamformers}, have achieved significant improvements in both signal quality (e.g. signal-to-noise ratio (SNR)) and speech recognition (e.g. word error rate (WER)). Such systems are generally non-causal and require a large context for robust estimation of inter-channel features, which is impractical in applications requiring low-latency responses. In this paper, we propose filter-and-sum network (FaSNet), a time-domain, filter-based beamforming approach suitable for low-latency scenarios. FaSNet has a two-stage system design that first learns frame-level time-domain adaptive beamforming filters for a selected reference channel, and then calculate the filters for all remaining channels. The filtered outputs at all channels are summed to generate the final output. Experiments show that despite its small model size, FaSNet is able to outperform several traditional oracle beamformers with respect to scale-invariant signal-to-noise ratio (SI-SNR) in reverberant speech enhancement and separation tasks. Moreover, when trained with a frequency-domain objective function on the CHiME-3 dataset, FaSNet achieves 14.3\% relative word error rate reduction (RWERR) compared with the baseline model. These results show the efficacy of FaSNet particularly in reverberant and noisy signal conditions.

Bodo Rückauer, Nicolas Känzig, Shih-Chii Liu et al.

Mobile and embedded applications require neural networks-based pattern recognition systems to perform well under a tight computational budget. In contrast to commonly used synchronous, frame-based vision systems and CNNs, asynchronous, spiking neural networks driven by event-based visual input respond with low latency to sparse, salient features in the input, leading to high efficiency at run-time. The discrete nature of the event-based data streams makes direct training of asynchronous neural networks challenging. This paper studies asynchronous spiking neural networks, obtained by conversion from a conventional CNN trained on frame-based data. As an example, we consider a CNN trained to steer a robot to follow a moving target. We identify possible pitfalls of the conversion and demonstrate how the proposed solutions bring the classification accuracy of the asynchronous network to only 3\% below the performance of the original synchronous CNN, while requiring 12x fewer computations. While being applied to a simple task, this work is an important step towards low-power, fast, and embedded neural networks-based vision solutions for robotic applications.

Matthew Thornton, Jithendar Anumula, Shih-Chii Liu

Recurrent neural networks can be difficult to train on long sequence data due to the well-known vanishing gradient problem. Some architectures incorporate methods to reduce RNN state updates, therefore allowing the network to preserve memory over long temporal intervals. To address these problems of convergence, this paper proposes a timing-gated LSTM RNN model, called the Gaussian-gated LSTM (g-LSTM). The time gate controls when a neuron can be updated during training, enabling longer memory persistence and better error-gradient flow. This model captures long-temporal dependencies better than an LSTM and the time gate parameters can be learned even from non-optimal initialization values. Because the time gate limits the updates of the neuron state, the number of computes needed for the network update is also reduced. By adding a computational budget term to the training loss, we can obtain a network which further reduces the number of computes by at least 10x. Finally, by employing a temporal curriculum learning schedule for the g-LSTM, we can reduce the convergence time of the equivalent LSTM network on long sequences.

Yuhuang Hu, Adrian Huber, Jithendar Anumula et al.

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.

Jonathan Binas, Daniel Neil, Shih-Chii Liu et al.

Event cameras, such as dynamic vision sensors (DVS), and dynamic and active-pixel vision sensors (DAVIS) can supplement other autonomous driving sensors by providing a concurrent stream of standard active pixel sensor (APS) images and DVS temporal contrast events. The APS stream is a sequence of standard grayscale global-shutter image sensor frames. The DVS events represent brightness changes occurring at a particular moment, with a jitter of about a millisecond under most lighting conditions. They have a dynamic range of >120 dB and effective frame rates >1 kHz at data rates comparable to 30 fps (frames/second) image sensors. To overcome some of the limitations of current image acquisition technology, we investigate in this work the use of the combined DVS and APS streams in end-to-end driving applications. The dataset DDD17 accompanying this paper is the first open dataset of annotated DAVIS driving recordings. DDD17 has over 12 h of a 346x260 pixel DAVIS sensor recording highway and city driving in daytime, evening, night, dry and wet weather conditions, along with vehicle speed, GPS position, driver steering, throttle, and brake captured from the car's on-board diagnostics interface. As an example application, we performed a preliminary end-to-end learning study of using a convolutional neural network that is trained to predict the instantaneous steering angle from DVS and APS visual data.

Stefan Braun, Daniel Neil, Enea Ceolini et al.

Recent work on encoder-decoder models for sequence-to-sequence mapping has shown that integrating both temporal and spatial attention mechanisms into neural networks increases the performance of the system substantially. In this work, we report on the application of an attentional signal not on temporal and spatial regions of the input, but instead as a method of switching among inputs themselves. We evaluate the particular role of attentional switching in the presence of dynamic noise in the sensors, and demonstrate how the attentional signal responds dynamically to changing noise levels in the environment to achieve increased performance on both audio and visual tasks in three commonly-used datasets: TIDIGITS, Wall Street Journal, and GRID. Moreover, the proposed sensor transformation network architecture naturally introduces a number of advantages that merit exploration, including ease of adding new sensors to existing architectures, attentional interpretability, and increased robustness in a variety of noisy environments not seen during training. Finally, we demonstrate that the sensor selection attention mechanism of a model trained only on the small TIDIGITS dataset can be transferred directly to a pre-existing larger network trained on the Wall Street Journal dataset, maintaining functionality of switching between sensors to yield a dramatic reduction of error in the presence of noise.

Alessandro Aimar, Hesham Mostafa, Enrico Calabrese et al.

Convolutional neural networks (CNNs) have become the dominant neural network architecture for solving many state-of-the-art (SOA) visual processing tasks. Even though Graphical Processing Units (GPUs) are most often used in training and deploying CNNs, their power efficiency is less than 10 GOp/s/W for single-frame runtime inference. We propose a flexible and efficient CNN accelerator architecture called NullHop that implements SOA CNNs useful for low-power and low-latency application scenarios. NullHop exploits the sparsity of neuron activations in CNNs to accelerate the computation and reduce memory requirements. The flexible architecture allows high utilization of available computing resources across kernel sizes ranging from 1x1 to 7x7. NullHop can process up to 128 input and 128 output feature maps per layer in a single pass. We implemented the proposed architecture on a Xilinx Zynq FPGA platform and present results showing how our implementation reduces external memory transfers and compute time in five different CNNs ranging from small ones up to the widely known large VGG16 and VGG19 CNNs. Post-synthesis simulations using Mentor Modelsim in a 28nm process with a clock frequency of 500 MHz show that the VGG19 network achieves over 450 GOp/s. By exploiting sparsity, NullHop achieves an efficiency of 368%, maintains over 98% utilization of the MAC units, and achieves a power efficiency of over 3TOp/s/W in a core area of 6.3mm$^2$. As further proof of NullHop's usability, we interfaced its FPGA implementation with a neuromorphic event camera for real time interactive demonstrations.

Daniel Neil, Jun Haeng Lee, Tobi Delbruck et al.

Many neural networks exhibit stability in their activation patterns over time in response to inputs from sensors operating under real-world conditions. By capitalizing on this property of natural signals, we propose a Recurrent Neural Network (RNN) architecture called a delta network in which each neuron transmits its value only when the change in its activation exceeds a threshold. The execution of RNNs as delta networks is attractive because their states must be stored and fetched at every timestep, unlike in convolutional neural networks (CNNs). We show that a naive run-time delta network implementation offers modest improvements on the number of memory accesses and computes, but optimized training techniques confer higher accuracy at higher speedup. With these optimizations, we demonstrate a 9X reduction in cost with negligible loss of accuracy for the TIDIGITS audio digit recognition benchmark. Similarly, on the large Wall Street Journal speech recognition benchmark even existing networks can be greatly accelerated as delta networks, and a 5.7x improvement with negligible loss of accuracy can be obtained through training. Finally, on an end-to-end CNN trained for steering angle prediction in a driving dataset, the RNN cost can be reduced by a substantial 100X.

Joachim Ott, Zhouhan Lin, Ying Zhang et al.

Recurrent Neural Networks (RNNs) produce state-of-art performance on many machine learning tasks but their demand on resources in terms of memory and computational power are often high. Therefore, there is a great interest in optimizing the computations performed with these models especially when considering development of specialized low-power hardware for deep networks. One way of reducing the computational needs is to limit the numerical precision of the network weights and biases, and this will be addressed for the case of RNNs. We present results from the use of different stochastic and deterministic reduced precision training methods applied to two major RNN types, which are then tested on three datasets. The results show that the stochastic and deterministic ternarization, pow2- ternarization, and exponential quantization methods gave rise to low-precision RNNs that produce similar and even higher accuracy on certain datasets, therefore providing a path towards training more efficient implementations of RNNs in specialized hardware.

Daniel Neil, Michael Pfeiffer, Shih-Chii Liu

Recurrent Neural Networks (RNNs) have become the state-of-the-art choice for extracting patterns from temporal sequences. However, current RNN models are ill-suited to process irregularly sampled data triggered by events generated in continuous time by sensors or other neurons. Such data can occur, for example, when the input comes from novel event-driven artificial sensors that generate sparse, asynchronous streams of events or from multiple conventional sensors with different update intervals. In this work, we introduce the Phased LSTM model, which extends the LSTM unit by adding a new time gate. This gate is controlled by a parametrized oscillation with a frequency range that produces updates of the memory cell only during a small percentage of the cycle. Even with the sparse updates imposed by the oscillation, the Phased LSTM network achieves faster convergence than regular LSTMs on tasks which require learning of long sequences. The model naturally integrates inputs from sensors of arbitrary sampling rates, thereby opening new areas of investigation for processing asynchronous sensory events that carry timing information. It also greatly improves the performance of LSTMs in standard RNN applications, and does so with an order-of-magnitude fewer computes at runtime.

Joachim Ott, Zhouhan Lin, Ying Zhang et al.

Recurrent Neural Networks (RNNs) produce state-of-art performance on many machine learning tasks but their demand on resources in terms of memory and computational power are often high. Therefore, there is a great interest in optimizing the computations performed with these models especially when considering development of specialized low-power hardware for deep networks. One way of reducing the computational needs is to limit the numerical precision of the network weights and biases. This has led to different proposed rounding methods which have been applied so far to only Convolutional Neural Networks and Fully-Connected Networks. This paper addresses the question of how to best reduce weight precision during training in the case of RNNs. We present results from the use of different stochastic and deterministic reduced precision training methods applied to three major RNN types which are then tested on several datasets. The results show that the weight binarization methods do not work with the RNNs. However, the stochastic and deterministic ternarization, and pow2-ternarization methods gave rise to low-precision RNNs that produce similar and even higher accuracy on certain datasets therefore providing a path towards training more efficient implementations of RNNs in specialized hardware.

Jonathan Binas, Daniel Neil, Giacomo Indiveri et al.

The operations used for neural network computation map favorably onto simple analog circuits, which outshine their digital counterparts in terms of compactness and efficiency. Nevertheless, such implementations have been largely supplanted by digital designs, partly because of device mismatch effects due to material and fabrication imperfections. We propose a framework that exploits the power of deep learning to compensate for this mismatch by incorporating the measured device variations as constraints in the neural network training process. This eliminates the need for mismatch minimization strategies and allows circuit complexity and power-consumption to be reduced to a minimum. Our results, based on large-scale simulations as well as a prototype VLSI chip implementation indicate a processing efficiency comparable to current state-of-art digital implementations. This method is suitable for future technology based on nanodevices with large variability, such as memristive arrays.

Stefan Braun, Daniel Neil, Shih-Chii Liu

The performance of automatic speech recognition systems under noisy environments still leaves room for improvement. Speech enhancement or feature enhancement techniques for increasing noise robustness of these systems usually add components to the recognition system that need careful optimization. In this work, we propose the use of a relatively simple curriculum training strategy called accordion annealing (ACCAN). It uses a multi-stage training schedule where samples at signal-to-noise ratio (SNR) values as low as 0dB are first added and samples at increasing higher SNR values are gradually added up to an SNR value of 50dB. We also use a method called per-epoch noise mixing (PEM) that generates noisy training samples online during training and thus enables dynamically changing the SNR of our training data. Both the ACCAN and the PEM methods are evaluated on a end-to-end speech recognition pipeline on the Wall Street Journal corpus. ACCAN decreases the average word error rate (WER) on the 20dB to -10dB SNR range by up to 31.4% when compared to a conventional multi-condition training method.

Giacomo Indiveri, Shih-Chii Liu

A striking difference between brain-inspired neuromorphic processors and current von Neumann processors architectures is the way in which memory and processing is organized. As Information and Communication Technologies continue to address the need for increased computational power through the increase of cores within a digital processor, neuromorphic engineers and scientists can complement this need by building processor architectures where memory is distributed with the processing. In this paper we present a survey of brain-inspired processor architectures that support models of cortical networks and deep neural networks. These architectures range from serial clocked implementations of multi-neuron systems to massively parallel asynchronous ones and from purely digital systems to mixed analog/digital systems which implement more biological-like models of neurons and synapses together with a suite of adaptation and learning mechanisms analogous to the ones found in biological nervous systems. We describe the advantages of the different approaches being pursued and present the challenges that need to be addressed for building artificial neural processing systems that can display the richness of behaviors seen in biological systems.

Shaista Hussain, Shih-Chii Liu, Arindam Basu

This paper presents a spike-based model which employs neurons with functionally distinct dendritic compartments for classifying high dimensional binary patterns. The synaptic inputs arriving on each dendritic subunit are nonlinearly processed before being linearly integrated at the soma, giving the neuron a capacity to perform a large number of input-output mappings. The model utilizes sparse synaptic connectivity; where each synapse takes a binary value. The optimal connection pattern of a neuron is learned by using a simple hardware-friendly, margin enhancing learning algorithm inspired by the mechanism of structural plasticity in biological neurons. The learning algorithm groups correlated synaptic inputs on the same dendritic branch. Since the learning results in modified connection patterns, it can be incorporated into current event-based neuromorphic systems with little overhead. This work also presents a branch-specific spike-based version of this structural plasticity rule. The proposed model is evaluated on benchmark binary classification problems and its performance is compared against that achieved using Support Vector Machine (SVM) and Extreme Learning Machine (ELM) techniques. Our proposed method attains comparable performance while utilizing 10 to 50% less computational resources than the other reported techniques.