Pushing the frontiers of time-series information processing in ever-growing edge devices with stringent resources has been impeded by the system's ability to process information and learn locally on the device. Local processing and learning typically demand intensive computations and massive storage as the process involves retrieving information and tuning hundreds of parameters back in time. In this work, we developed a memristor-based echo state network accelerator that features efficient temporal data processing and in-situ online learning. The proposed design is benchmarked using various datasets involving real-world tasks, such as forecasting the load energy consumption and weather conditions. The experimental results illustrate that the hardware model experiences a marginal degradation (~4.8%) in performance as compared to the software model. This is mainly attributed to the limited precision and dynamic range of network parameters when emulated using memristor devices. The proposed system is evaluated for lifespan, robustness, and energy-delay product. It is observed that the system demonstrates a reasonable robustness for device failure below 10%, which may occur due to stuck-at faults. Furthermore, 246X reduction in energy consumption is achieved when compared to a custom CMOS digital design implemented at the same technology node.

We introduce a new challenge to test the STEM skills of neural models. The problems in the real world often require solutions, combining knowledge from STEM (science, technology, engineering, and math). Unlike existing datasets, our dataset requires the understanding of multimodal vision-language information of STEM. Our dataset features one of the largest and most comprehensive datasets for the challenge. It includes 448 skills and 1,073,146 questions spanning all STEM subjects. Compared to existing datasets that often focus on examining expert-level ability, our dataset includes fundamental skills and questions designed based on the K-12 curriculum. We also add state-of-the-art foundation models such as CLIP and GPT-3.5-Turbo to our benchmark. Results show that the recent model advances only help master a very limited number of lower grade-level skills (2.5% in the third grade) in our dataset. In fact, these models are still well below (averaging 54.7%) the performance of elementary students, not to mention near expert-level performance. To understand and increase the performance on our dataset, we teach the models on a training split of our dataset. Even though we observe improved performance, the model performance remains relatively low compared to average elementary students. To solve STEM problems, we will need novel algorithmic innovations from the community.

Scaling laws for large language models (LLMs) have provided useful guidance on how to train ever larger models for predictable performance gains. Time series forecasting shares a similar sequential structure to language, and is amenable to large-scale transformer architectures. Here we show that foundational decoder-only time series transformer models exhibit analogous scaling-behavior to LLMs, while architectural details (aspect ratio and number of heads) have a minimal effect over broad ranges. We assemble a large corpus of heterogenous time series data on which to train, and establish, for the first time, power-law scaling relations with respect to parameter count, dataset size, and training compute, spanning five orders of magnitude.

Large language models (LLMs) are trained on a vast amount of human-written data, but data providers often remain uncredited. In response to this issue, data valuation (or data attribution), which quantifies the contribution or value of each data to the model output, has been discussed as a potential solution. Nevertheless, applying existing data valuation methods to recent LLMs and their vast training datasets has been largely limited by prohibitive compute and memory costs. In this work, we focus on influence functions, a popular gradient-based data valuation method, and significantly improve its scalability with an efficient gradient projection strategy called LoGra that leverages the gradient structure in backpropagation. We then provide a theoretical motivation of gradient projection approaches to influence functions to promote trust in the data valuation process. Lastly, we lower the barrier to implementing data valuation systems by introducing LogIX, a software package that can transform existing training code into data valuation code with minimal effort. In our data valuation experiments, LoGra achieves competitive accuracy against more expensive baselines while showing up to 6,500x improvement in throughput and 5x reduction in GPU memory usage when applied to Llama3-8B-Instruct and the 1B-token dataset.

Multivariate time-series forecasting (MTSF) stands as a compelling field within the machine learning community. Diverse neural network based methodologies deployed in MTSF applications have demonstrated commendable efficacy. Despite the advancements in model performance, comprehending the rationale behind the model's behavior remains an enigma. Our proposed model, the Neural ForeCasting Layer (NFCL), employs a straightforward amalgamation of neural networks. This uncomplicated integration ensures that each neural network contributes inputs and predictions independently, devoid of interference from other inputs. Consequently, our model facilitates a transparent explication of forecast results. This paper introduces NFCL along with its diverse extensions. Empirical findings underscore NFCL's superior performance compared to nine benchmark models across 15 available open datasets. Notably, NFCL not only surpasses competitors but also provides elucidation for its predictions. In addition, Rigorous experimentation involving diverse model structures bolsters the justification of NFCL's unique configuration.

Local climate zones (LCZs) established a standard classification system to categorize the landscape universe for improved urban climate studies. Existing LCZ mapping is guided by human interaction with geographic information systems (GIS) or modelled from remote sensing (RS) data. GIS-based methods do not scale to large areas. However, RS-based methods leverage machine learning techniques to automatize LCZ classification from RS. Yet, RS-based methods require huge amounts of manual labels for training. We propose a novel LCZ mapping framework, termed AutoLCZ, to extract the LCZ classification features from high-resolution RS modalities. We study the definition of numerical rules designed to mimic the LCZ definitions. Those rules model geometric and surface cover properties from LiDAR data. Correspondingly, we enable LCZ classification from RS data in a GIS-based scheme. The proposed AutoLCZ method has potential to reduce the human labor to acquire accurate metadata. At the same time, AutoLCZ sheds light on the physical interpretability of RS-based methods. In a proof-of-concept for New York City (NYC) we leverage airborne LiDAR surveys to model 4 LCZ features to distinguish 10 LCZ types. The results indicate the potential of AutoLCZ as promising avenue for large-scale LCZ mapping from RS data.

State-space models (SSMs) that utilize linear, time-invariant (LTI) systems are known for their effectiveness in learning long sequences. However, these models typically face several challenges: (i) they require specifically designed initializations of the system matrices to achieve state-of-the-art performance, (ii) they require training of state matrices on a logarithmic scale with very small learning rates to prevent instabilities, and (iii) they require the model to have exponentially decaying memory in order to ensure an asymptotically stable LTI system. To address these issues, we view SSMs through the lens of Hankel operator theory, which provides us with a unified theory for the initialization and training of SSMs. Building on this theory, we develop a new parameterization scheme, called HOPE, for LTI systems that utilizes Markov parameters within Hankel operators. This approach allows for random initializations of the LTI systems and helps to improve training stability, while also provides the SSMs with non-decaying memory capabilities. Our model efficiently implements these innovations by nonuniformly sampling the transfer functions of LTI systems, and it requires fewer parameters compared to canonical SSMs. When benchmarked against HiPPO-initialized models such as S4 and S4D, an SSM parameterized by Hankel operators demonstrates improved performance on Long-Range Arena (LRA) tasks. Moreover, we use a sequential CIFAR-10 task with padded noise to empirically corroborate our SSM's long memory capacity.

Traditionally, the task of sampling from un-normalized densities is largely delegated to MCMC methods. However, modern machine learning developments in optimal transport and generative modeling have greatly expanded the toolbox we have available for performing such tasks, which can further benefit from powerful advancements in deep learning. By reducing the problem to performing empirical risk minimization with neural networks, they hold promise especially for sampling from high-dimensional and multimodal distributions compared to MCMC-based alternatives. In this work, we propose novel training objectives that are amenable to tractable approximations for sampling (without access to data from the target, as typically considered in the MCMC setup), and demonstrate their numerical advantages compared to several related methods that are also built around forward-backward SDEs through time-reversals. Before elaborating on these connections and synthesis in Section 2, we briefly summarize our contributions below. We present a transport/control based sampler using Schrödinger bridge (SB) idea (generalizable to extended state space and nonlinear prior), which compared to MCMC-based approaches has two benefits: (1) alleviate metastability without being confined to make local moves in the state space; (2) based on interpolating path, it comes with the ability to provide low-variance, unbiased normalizing constant estimates using the trajectory information (c.f.

The Twelfth International Conference on Learning Representations (ICLR2024) took place from 7-11 May in Vienna. The program included workshops, contributed talks, affinity group events, and socials. There were also seven invited talks that covered a broad range of topics. In this post, we give a summary of the talk by Priya Donti. Priya's research focuses on machine learning for forecasting, optimization, and control in power grids. She is an Assistant Professor and the Silverman (1968) Family Career Development Professor at MIT.

The predictive brain hypothesis suggests that perception can be interpreted as the process of minimizing the error between predicted perception tokens generated by an internal world model and actual sensory input tokens. When implementing working examples of this hypothesis in the context of in-air sonar, significant difficulties arise due to the sparse nature of the reflection model that governs ultrasonic sensing. Despite these challenges, creating consistent world models using sonar data is crucial for implementing predictive processing of ultrasound data in robotics. In an effort to enable robust robot behavior using ultrasound as the sole exteroceptive sensor modality, this paper introduces EchoPT, a pretrained transformer architecture designed to predict 2D sonar images from previous sensory data and robot ego-motion information. We detail the transformer architecture that drives EchoPT and compare the performance of our model to several state-of-the-art techniques. In addition to presenting and evaluating our EchoPT model, we demonstrate the effectiveness of this predictive perception approach in two robotic tasks.