In the summer of 2012, Google made a big media splash when it showed that its researchers "trained a network of 1,000 computers wired up like a brain to recognize cats." While AI, neural networks and its most recent rebranding, "deep learning," were already established fields with decades of research and countless real-world applications behind them, the world at large (and all its cats) took notice. Deep learning, a branch of AI that closely mimics how neurons wire and fire, was becoming more powerful: The massive amounts of digital data and compute power needed for training these systems were now available to companies like Google. Since 2012, applications of AI have expanded to both the consumer and enterprise realms. For instance, AI can be applied to make smart phone pictures more beautiful, delete spam messages, recognize faces, translate languages, make video games more appealing and optimize sales engagements, among many others.

I recently read the Fast AI deep learning book and wanted to summarise some of the many advanced takeaways & tricks I got from it. I'm going to leave out the basic things because there's enough posts about them, i'm just focusing on what I found new or special in the book. I've also put the insights into a deck on save all to help you remember them over the long-term. I would massively recommend using a spaced repetition app (video explanation) like anki or save all for the things you learn otherwise you'll just forget so much of what is important. Please let me know if you found this helpful and if there are any other training tricks you use that we should also know about?

Robotics being used in healthcare can be defined as the application and use of robotic technology to diagnose, treat, correct, restore, or modify a body function, body part, or disease. However, there are many other ways we can see robotics in healthcare used for daily operations. In the healthcare sector, robotics can play a role in sanitization, triage, and even performing actual surgeries; but there is a lot of space in between in which robots in this industry do not meet the standard definition listed above. More often than not, science fiction brings about real innovation and creation. Robots were once imagined as something so far outside of human control that they could never truly exist.

University of Waterloo researchers are using deep learning and computer vision to develop autonomous exoskeleton legs to help users walk, climb stairs, and avoid obstacles. The ExoNet project, described in an early-access paper on "Frontiers in Robotics and AI", fits users with wearable cameras. AI software processes the camera's video stream, and is being trained to recognize surrounding features such as stairs and doorways, and then determine the best movements to take. "Our control approach wouldn't necessarily require human thought," said Brokoslaw Laschowski, Ph.D. candidate in systems design engineering and lead author on the ExoNet project. "Similar to autonomous cars that drive themselves, we're designing autonomous exoskeletons that walk for themselves."

Lee Sedol, a world champion in the Chinese strategy board game Go, faced a new kind of adversary at a 2016 match in Seoul. Developers at DeepMind, an artificial intelligence startup acquired by Google, had fed 30 million Go moves into a deep neural network. Their creation, dubbed AlphaGo, then figured out which moves worked by playing millions of games against itself, learning at a faster rate than any human ever could. The match, which AlphaGo won 4 to 1, "was the moment when the new movement in artificial intelligence exploded into the public consciousness," technology journalist Cade Metz writes in his engaging new book, "Genius Makers: The Mavericks Who Brought AI to Google, Facebook, and the World." Metz, who covers AI for The New York Times and previously wrote for Wired magazine, is well positioned to chart the decades-long effort to build artificially intelligent machines.

We review a recently introduced Structured Deep Kernel Network (SDKN) approach that is capable of dealing with high-dimensional and huge datasets - and enjoys typical standard machine learning approximation properties. We extend the SDKN to combine it with standard machine learning modules and compare it with Neural Networks on the scientific challenge of data-driven prediction of closure terms of turbulent flows. We show experimentally that the SDKNs are capable of dealing with large datasets and achieve near-perfect accuracy on the given application.

Though pre-trained language models such as Bert and XLNet, have rapidly advanced the state-of-the-art on many NLP tasks, they implicit semantics only relying on surface information between words in corpus. Intuitively, background knowledge influences the efficacy of understanding. Inspired by this common sense, we focus on improving model pretraining by leveraging explicit knowledge. Different from recent research that optimize pretraining model by knowledge masking strategies, we propose a simple but general method to combine explicit knowledge with pretraining. To be specific, we first match knowledge facts from knowledge graph (KG) and then add a knowledge injunction layer to transformer directly without changing its architecture. The present study seeks to find the direct impact of explicit knowledge on transformer per-training. We conduct experiments on various datasets for different downstream tasks. The experimental results show that solely by adding external knowledge to transformer can improve the learning performance on many NLP tasks.

Predicting agents' future trajectories plays a crucial role in modern AI systems, yet it is challenging due to intricate interactions exhibited in multi-agent systems, especially when it comes to collision avoidance. To address this challenge, we propose to learn congestion patterns as contextual cues explicitly and devise a novel "Sense--Learn--Reason--Predict" framework by exploiting advantages of three different doctrines of thought, which yields the following desirable benefits: (i) Representing congestion as contextual cues via latent factors subsumes the concept of social force commonly used in physics-based approaches and implicitly encodes the distance as a cost, similar to the way a planning-based method models the environment. (ii) By decomposing the learning phases into two stages, a "student" can learn contextual cues from a "teacher" while generating collision-free trajectories. To make the framework computationally tractable, we formulate it as an optimization problem and derive an upper bound by leveraging the variational parametrization. In experiments, we demonstrate that the proposed model is able to generate collision-free trajectory predictions in a synthetic dataset designed for collision avoidance evaluation and remains competitive on the commonly used NGSIM US-101 highway dataset.

Deep Neural Networks (DNNs) are popularly used for implementing autonomy related tasks in automotive Cyber-Physical Systems (CPSs). However, these networks have been shown to make erroneous predictions to anomalous inputs, which manifests either due to Out-of-Distribution (OOD) data or adversarial attacks. To detect these anomalies, a separate DNN called assurance monitor is often trained and used in parallel to the controller DNN, increasing the resource burden and latency. We hypothesize that a single network that can perform controller predictions and anomaly detection is necessary to reduce the resource requirements. Deep-Radial Basis Function (RBF) networks provide a rejection class alongside the class predictions, which can be utilized for detecting anomalies at runtime. However, the use of RBF activation functions limits the applicability of these networks to only classification tasks. In this paper, we show how the deep-RBF network can be used for detecting anomalies in CPS regression tasks such as continuous steering predictions. Further, we design deep-RBF networks using popular DNNs such as NVIDIA DAVE-II, and ResNet20, and then use the resulting rejection class for detecting adversarial attacks such as a physical attack and data poison attack. Finally, we evaluate these attacks and the trained deep-RBF networks using a hardware CPS testbed called DeepNNCar and a real-world German Traffic Sign Benchmark (GTSB) dataset. Our results show that the deep-RBF networks can robustly detect these attacks in a short time without additional resource requirements.

The wide adoption of Electronic Health Records (EHR) has resulted in large amounts of clinical data becoming available, which promises to support service delivery and advance clinical and informatics research. Deep learning techniques have demonstrated performance in predictive analytic tasks using EHRs yet they typically lack model result transparency or explainability functionalities and require cumbersome pre-processing tasks. Moreover, EHRs contain heterogeneous and multi-modal data points such as text, numbers and time series which further hinder visualisation and interpretability. This paper proposes a deep learning framework to: 1) encode patient pathways from EHRs into images, 2) highlight important events within pathway images, and 3) enable more complex predictions with additional intelligibility. The proposed method relies on a deep attention mechanism for visualisation of the predictions and allows predicting multiple sequential outcomes.