In AI development, the dominant paradigm is that the more training data, the better. OpenAI's GPT-2 model had a data set consisting of 40 gigabytes of text. GPT-3, which ChatGPT is based on, was trained on 570 GB of data. OpenAI has not shared how big the data set for its latest model, GPT-4, is. But that hunger for larger models is now coming back to bite the company. In the past few weeks, several Western data protection authorities have started investigations into how OpenAI collects and processes the data powering ChatGPT.

This paper introduces a new learning-based approach for approximately solving the Kidney-Exchange Problem (KEP), an NP-hard problem on graphs. The problem consists of, given a pool of kidney donors and patients waiting for kidney donations, optimally selecting a set of donations to optimize the quantity and quality of transplants performed while respecting a set of constraints about the arrangement of these donations. The proposed technique consists of two main steps: the first is a Graph Neural Network (GNN) trained without supervision; the second is a deterministic non-learned search heuristic that uses the output of the GNN to find paths and cycles. To allow for comparisons, we also implemented and tested an exact solution method using integer programming, two greedy search heuristics without the machine learning module, and the GNN alone without a heuristic. We analyze and compare the methods and conclude that the learning-based two-stage approach is the best solution quality, outputting approximate solutions on average 1.1 times more valuable than the ones from the deterministic heuristic alone.

A scalable and computationally efficient framework is designed to fingerprint real-world Bluetooth devices. We propose an embedding-assisted attentional framework (Mbed-ATN) suitable for fingerprinting actual Bluetooth devices. Its generalization capability is analyzed in different settings and the effect of sample length and anti-aliasing decimation is demonstrated. The embedding module serves as a dimensionality reduction unit that maps the high dimensional 3D input tensor to a 1D feature vector for further processing by the ATN module. Furthermore, unlike the prior research in this field, we closely evaluate the complexity of the model and test its fingerprinting capability with real-world Bluetooth dataset collected under a different time frame and experimental setting while being trained on another. Our study reveals a 9.17x and 65.2x lesser memory usage at a sample length of 100 kS when compared to the benchmark - GRU and Oracle models respectively. Further, the proposed Mbed-ATN showcases 16.9x fewer FLOPs and 7.5x lesser trainable parameters when compared to Oracle. Finally, we show that when subject to anti-aliasing decimation and at greater input sample lengths of 1 MS, the proposed Mbed-ATN framework results in a 5.32x higher TPR, 37.9% fewer false alarms, and 6.74x higher accuracy under the challenging real-world setting.

This paper presents BERT-CTC, a novel formulation of end-to-end speech recognition that adapts BERT for connectionist temporal classification (CTC). Our formulation relaxes the conditional independence assumptions used in conventional CTC and incorporates linguistic knowledge through the explicit output dependency obtained by BERT contextual embedding. BERT-CTC attends to the full contexts of the input and hypothesized output sequences via the self-attention mechanism. This mechanism encourages a model to learn inner/inter-dependencies between the audio and token representations while maintaining CTC's training efficiency. During inference, BERT-CTC combines a mask-predict algorithm with CTC decoding, which iteratively refines an output sequence. The experimental results reveal that BERT-CTC improves over conventional approaches across variations in speaking styles and languages. Finally, we show that the semantic representations in BERT-CTC are beneficial towards downstream spoken language understanding tasks.

Mining data streams is one of the main studies in machine learning area due to its application in many knowledge areas. One of the major challenges on mining data streams is concept drift, which requires the learner to discard the current concept and adapt to a new one. Ensemble-based drift detection algorithms have been used successfully to the classification task but usually maintain a fixed size ensemble of learners running the risk of needlessly spending processing time and memory. In this paper we present improvements to the Scale-free Network Regressor (SFNR), a dynamic ensemble-based method for regression that employs social networks theory. In order to detect concept drifts SFNR uses the Adaptive Window (ADWIN) algorithm. Results show improvements in accuracy, especially in concept drift situations and better performance compared to other state-of-the-art algorithms in both real and synthetic data.

Online reinforcement learning and other adaptive sampling algorithms are increasingly used in digital intervention experiments to optimize treatment delivery for users over time. In this work, we focus on longitudinal user data collected by a large class of adaptive sampling algorithms that are designed to optimize treatment decisions online using accruing data from multiple users. Combining or "pooling" data across users allows adaptive sampling algorithms to potentially learn faster. However, by pooling, these algorithms induce dependence between the sampled user data trajectories; we show that this can cause standard variance estimators for i.i.d. data to underestimate the true variance of common estimators on this data type. We develop novel methods to perform a variety of statistical analyses on such adaptively sampled data via Z-estimation. Specifically, we introduce the \textit{adaptive} sandwich variance estimator, a corrected sandwich estimator that leads to consistent variance estimates under adaptive sampling. Additionally, to prove our results we develop novel theoretical tools for empirical processes on non-i.i.d., adaptively sampled longitudinal data which may be of independent interest. This work is motivated by our efforts in designing experiments in which online reinforcement learning algorithms optimize treatment decisions, yet statistical inference is essential for conducting analyses after experiments conclude.

This work contributes to developing an agent based on deep reinforcement learning capable of acting in a beyond visual range (BVR) air combat simulation environment. The paper presents an overview of building an agent representing a high-performance fighter aircraft that can learn and improve its role in BVR combat over time based on rewards calculated using operational metrics. Also, through self-play experiments, it expects to generate new air combat tactics never seen before. Finally, we hope to examine a real pilot's ability, using virtual simulation, to interact in the same environment with the trained agent and compare their performances. This research will contribute to the air combat training context by developing agents that can interact with real pilots to improve their performances in air defense missions.

This technical report presents the application of a recurrent memory to extend the context length of BERT, one of the most effective Transformer-based models in natural language processing. By leveraging the Recurrent Memory Transformer architecture, we have successfully increased the model's effective context length to an unprecedented two million tokens, while maintaining high memory retrieval accuracy. Our method allows for the storage and processing of both local and global information and enables information flow between segments of the input sequence through the use of recurrence. Our experiments demonstrate the effectiveness of our approach, which holds significant potential to enhance long-term dependency handling in natural language understanding and generation tasks as well as enable large-scale context processing for memory-intensive applications.

This paper considers the problem of learning safe policies in the context of reinforcement learning (RL). In particular, we consider the notion of probabilistic safety. This is, we aim to design policies that maintain the state of the system in a safe set with high probability. This notion differs from cumulative constraints often considered in the literature. The challenge of working with probabilistic safety is the lack of expressions for their gradients. Indeed, policy optimization algorithms rely on gradients of the objective function and the constraints. To the best of our knowledge, this work is the first one providing such explicit gradient expressions for probabilistic constraints. It is worth noting that the gradient of this family of constraints can be applied to various policy-based algorithms. We demonstrate empirically that it is possible to handle probabilistic constraints in a continuous navigation problem.

Endovascular guidewire manipulation is essential for minimally-invasive clinical applications (Percutaneous Coronary Intervention (PCI), Mechanical thrombectomy techniques for acute ischemic stroke (AIS), or Transjugular intrahepatic portosystemic shunt (TIPS)). All procedures commonly require 3D vessel geometries from 3D CTA (Computed Tomography Angiography) images. During these procedures, the clinician generally places a guiding catheter in the ostium of the relevant vessel and then manipulates a wire through the catheter and across the blockage. The clinician only uses X-ray fluoroscopy intermittently to visualize and guide the catheter, guidewire, and other devices. However, clinicians still passively control guidewires/catheters by relying on limited indirect observation (i.e., 2D partial view of devices, and intermittent updates due to radiation limit) from X-ray fluoroscopy. Modeling and controlling the guidewire manipulation in coronary vessels remains challenging because of the complicated interaction between guidewire motions with different physical properties (i.e., loads, coating) and vessel geometries with lumen conditions resulting in a highly non-linear system. This paper introduces a scalable learning pipeline to train AI-based agent models toward automated endovascular predictive device controls. First, we create a scalable environment by pre-processing 3D CTA images, providing patient-specific 3D vessel geometry and the centerline of the coronary. Next, we apply a large quantity of randomly generated motion sequences from the proximal end to generate wire states associated with each environment using a physics-based device simulator. Then, we reformulate the control problem to a sequence-to-sequence learning problem, in which we use a Transformer-based model, trained to handle non-linear sequential forward/inverse transition functions.