For software development, there are many methodologies, patterns, and techniques to build, deploy, and run applications. DevOps is the state of the art methodology that describes a software engineering culture with a holistic view of software development and operation. For data science, there is a lot of information on how machine and deep learning models can be built. The operational aspects seem to still be evolving. I'm currently trying to understand better how to deploy models in the cloud and how to use them efficiently in applications.

We could use the same approach taken by many research works that generate images and paint the square with a stack of deconvolution (transposed convolution) layers. To test this idea, we created a dataset consisting of randomly placed 9 9 squares on a 64 64 canvas, as shown in Figure 1b. To evaluate how well models generalize, we define two train/test splits: a uniform split, where all possible center locations are randomly divided 80 percent/20 percent into train vs. test sets, and a quadrant split, where the canvas is divided into four quadrants: squares centered in the first three quadrants are put in the train set and squares in the last quadrant in the test set.

An earlier version of this post was published on Off the Convex Path. It is reposted here with the author's permission. In the last few years, deep learning practitioners have proposed a litany of different sequence models. Although recurrent neural networks were once the tool of choice, now models like the autoregressive Wavenet or the Transformer are replacing RNNs on a diverse set of tasks. In this post, we explore the trade-offs between recurrent and feed-forward models.

An earlier version of this post was published on Off the Convex Path. It is reposted here with the author's permission. In the last few years, deep learning practitioners have proposed a litany of different sequence models. Although recurrent neural networks were once the tool of choice, now models like the autoregressive Wavenet or the Transformer are replacing RNNs on a diverse set of tasks. In this post, we explore the trade-offs between recurrent and feed-forward models.

In the 15 years since I first introduced the term "artificial general intelligence" (AGI), the AI field has advanced tremendously. We now have self-driving cars, automated face recognition and image captioning, machine translation and expert AI game-players, and so much more. However, these achievements remain essentially in the domain of "narrow AI"--AI that carries out tasks based on specifically-supplied data or rules, or carefully-created training situations. AIs that can generalize to unanticipated domains and confront the world as autonomous agents are still part of the road ahead. The question remains: what do we need to do to get from today's narrow AI tools, which have become mainstream in business and society, to the AGI envisioned by futurists and science fiction authors?

This course is all about the application of deep learning and neural networks to reinforcement learning. If you've taken my first reinforcement learning class, then you know that reinforcement learning is on the bleeding edge of what we can do with AI. Specifically, the combination of deep learning with reinforcement learning has led to AlphaGo beating a world champion in the strategy game Go, it has led to self-driving cars, and it has led to machines that can play video games at a superhuman level. Reinforcement learning has been around since the 70s but none of this has been possible until now. The world is changing at a very fast pace. The state of California is changing their regulations so that self-driving car companies can test their cars without a human in the car to supervise.

U.K.-based Deepmind has created artificial intelligence software that can read lips. An artificial intelligence (AI) program from DeepMind can read lips better than professional lip readers after reviewing thousands of hours of YouTube videos along with transcripts via machine learning. The researchers tested the program on 37 minutes of video it had not previously viewed, and it misidentified only 41% of the words. In comparison, the best previous computer method, which focuses on individual letters instead of phonemes, had a 77% word error rate, while professional lip readers had a 93% error rate in the same test, which lacked context or body language. Columbia University's Hassan Akbari says the AI, if incorporated into a phone, would enable hearing-impaired users to have a "translator" with them wherever they go.

Arguably the biggest challenge in applying neural networks is tuning the hyperparameters, in particular the learning rate. The sensitivity to the learning rate is due to the reliance on backpropagation to train the network. In this paper we present the first application of Implicit Stochastic Gradient Descent (ISGD) to train neural networks, a method known in convex optimization to be unconditionally stable and robust to the learning rate. Our key contribution is a novel layer-wise approximation of ISGD which makes its updates tractable for neural networks. Experiments show that our method is more robust to high learning rates and generally outperforms standard backpropagation on a variety of tasks.

In automatic speech recognition (ASR) what a user says depends on the particular context she is in. Typically, this context is represented as a set of word n-grams. In this work, we present a novel, all-neural, end-to-end (E2E) ASR sys- tem that utilizes such context. Our approach, which we re- fer to as Contextual Listen, Attend and Spell (CLAS) jointly- optimizes the ASR components along with embeddings of the context n-grams. During inference, the CLAS system can be presented with context phrases which might contain out-of- vocabulary (OOV) terms not seen during training. We com- pare our proposed system to a more traditional contextualiza- tion approach, which performs shallow-fusion between inde- pendently trained LAS and contextual n-gram models during beam search. Across a number of tasks, we find that the pro- posed CLAS system outperforms the baseline method by as much as 68% relative WER, indicating the advantage of joint optimization over individually trained components. Index Terms: speech recognition, sequence-to-sequence models, listen attend and spell, LAS, attention, embedded speech recognition.

Obtaining magnetic resonance images (MRI) with high resolution and generating quantitative image-based biomarkers for assessing tissue biochemistry is crucial in clinical and research applications. How- ever, acquiring quantitative biomarkers requires high signal-to-noise ratio (SNR), which is at odds with high-resolution in MRI, especially in a single rapid sequence. In this paper, we demonstrate how super-resolution can be utilized to maintain adequate SNR for accurate quantification of the T2 relaxation time biomarker, while simultaneously generating high- resolution images. We compare the efficacy of resolution enhancement using metrics such as peak SNR and structural similarity. We assess accuracy of cartilage T2 relaxation times by comparing against a standard reference method. Our evaluation suggests that SR can successfully maintain high-resolution and generate accurate biomarkers for accelerating MRI scans and enhancing the value of clinical and research MRI.