We present $\Gamma$-nets, a method for generalizing value function estimation over timescale. By using the timescale as one of the estimator's inputs we can estimate value for arbitrary timescales. As a result, the prediction target for any timescale is available and we are free to train on multiple timescales at each timestep. Here we empirically evaluate $\Gamma$-nets in the policy evaluation setting. We first demonstrate the approach on a square wave and then on a robot arm using linear function approximation. Next, we consider the deep reinforcement learning setting using several Atari video games. Our results show that $\Gamma$-nets can be effective for predicting arbitrary timescales, with only a small cost in accuracy as compared to learning estimators for fixed timescales. $\Gamma$-nets provide a method for compactly making predictions at many timescales without requiring a priori knowledge of the task, making it a valuable contribution to ongoing work on model-based planning, representation learning, and lifelong learning algorithms.

Wow, what a week it was. The EMNLP conference gave us many treats to chew on such as the growing popularity of cross-lingual learning and the continued adoption of knowledge graphs in language models. Because of all this action, this week's Cypher will be a bit longer than usual. What were some of the top keywords in EMNLP papers? Stephen Mayhew et al were live tweeting during the conference (thank you) and sharing all the action.

Back In 2000, Goldman Sachs employed 600 people to execute stock trades for the investment bank's major clients. By 2017, that workforce had reportedly dwindled to just two traders -- the others had been replaced by automated trading systems that can handle millions of transactions per minute. The rise of artificial intelligence threatens many more college-educated workers as the technology becomes more sophisticated and is more widely adopted by a range of industries, according to a new Brookings Institution analysis. In fact, the researchers say, AI is five times as likely to displace college grads than those without a degree. The research "suggests that better-educated, better-paid workers (along with manufacturing and production workers) will be the most affected by the new AI technologies, with some exceptions," write Mark Muro, Jacob Whiton and Robert Maxim of Brookings.

I started developing AI algorithms for handwriting recognition at my part-time student job while doing my Undergraduate degree in Computer Science. Since then, over the last 20 years or so, I have strived to combine my work in the industry with academic research. I did my Graduate degree in Computer Vision and completed my Ph.D. in Machine Learning while having quite an intensive career in the industry in parallel with my studies. In the industry, I've worked on all kinds of data and applications, including medical imaging, educational multimedia, mobile advertising, financial time series, video, text and speech processing for public safety, and other projects. When I began working with the product and business aspects of R&D, I felt that I needed to strengthen the relevant skills, so I went back to school and got an additional Master's degree in Technology Management.

Mukesh Ambani's Jio Institute will start its first academic session in 2021. It will focus on undergraduate courses in emerging technologies like artificial intelligence, data science, machine learning, and digital media and marketing. The digital courses will be led by Shailesh Kumar, who is the chief data scientist for Jio. The Human Resource & Development (HRD) Ministry had granted the'eminence' status to the Jio Institute even before it went operational. Mukesh Ambani's Jio Institute which will start its academic session in 2021 -- will offer courses in emerging technologies like artificial intelligence, data science, machine learning, digital media and marketing.

In machine learning, most tasks can be easily categorized into one of two different classes: supervised learning problems or unsupervised learning problems. In supervised learning, data has labels or classes appended to it, while in the case of unsupervised learning the data is unlabeled. Let's take a close look at why this distinction is important and look at some of the algorithms associated with each type of learning. Most machine learning tasks are in the domain of supervised learning. In supervised learning algorithms, the individual instances/data points in the dataset have a class or label assigned to them.

A modern data analytics framework should allow for rapid scaling and processing at scale. It also should be able to intelligently take only valuable data sources and present data in a consumable fashion. Cuelogic helps to rapidly prototype models that can meet data strategies.

We live in a data society. Journalists are becoming data analysts and data curators, and computation is an essential tool for reporting. Data and computation reshape the way a reporter sees the world and composes a story. They also control the operation of the information ecosystem she sends her journalism into, influencing where it finds audiences and generates discussion. So every reporting beat is now a data beat, and computation is an essential tool for investigation. But digitization is affected by inequities, leaving gaps that often reflect the very disparities reporters seek to illustrate. Computation is creating new systems of power and inequality in the world. We rely on journalists, the "explainers of last resort"[1], to hold these new constellations of power to account. We report on computation, not just with computation. While a term with considerable history and mystery, artificial intelligence (AI) represents the most recent bundling of data and computation to optimize business decisions, automate tasks, and, from the point of view of a reporter, learn about the world. The relationship between a journalist and AI is not unlike the process of developing sources or cultivating fixers. As with human sources, artificial intelligences may be knowledgeable, but they are not free of subjectivetivity in their design -- they also need to be contextualized and qualified. Ethical questions of introducing AI in journalism abound. But since AI has once again captured the public imagination, it is hard to have a clear-eyed discussion about the issues involved with journalism's call to both report on and with these new computational tools. And so our article will alternate a discussion of issues facing the profession today with a "slant narrative" -- indicated because these sections are in italics. The slant narrative starts with the 1964 World's Fair and a partnership between IBM and The New York Times, winds through commentary by Joseph Weizenbaum, a famed figure in AI research in the 1960s, and ends in 1983 with the shuttering of one of the most ambitious information delivery systems of the time. The simplicity of the role of computation in the slant narrative will help us better understand our contemporary situation with AI. But we begin our article with context for the use of data and computation in journalism -- a short, and certainly incomplete, history before we settle into the rhythm of alternating narratives. Reporters depend on data, and through computation they make sense of that data. This reliance is not new. Joseph Pulitzer listed a series of topics that should be taught to aspiring journalists in his 1904 article "The College of Journalism."

To acquire a new skill, humans learn better and faster if a tutor, based on their current knowledge level, informs them of how much attention they should pay to particular content or practice problems. Similarly, a machine learning model could potentially be trained better with a scorer that "adapts" to its current learning state and estimates the importance of each training data instance. Training such an adaptive scorer efficiently is a challenging problem; in order to precisely quantify the effect of a data instance at a given time during the training, it is typically necessary to first complete the entire training process. To efficiently optimize data usage, we propose a reinforcement learning approach called Differentiable Data Selection (DDS). In DDS, we formulate a scorer network as a learnable function of the training data, which can be efficiently updated along with the main model being trained. Specifically, DDS updates the scorer with an intuitive reward signal: it should up-weigh the data that has a similar gradient with a dev set upon which we would finally like to perform well. Without significant computing overhead, DDS delivers strong and consistent improvements over several strong baselines on two very different tasks of machine translation and image classification.

This loss discrepancy across groups is especially problematic in critical applications that impact people's lives (Berk, 2012; Chouldechova, 2017). Despite the vast literature on removing loss discrepancy (Hardt et al., 2016; Khani et al., 2019; Agarwal et al., 2018; Zafar et al., 2017), the direct removal of loss discrepancy might introduce other problems such as intragroup loss discrepancy (Lipton et al., 2018) and adverse long-term impacts (Liu et al., 2018). Therefore, it is important to understand the source of loss discrepancy. Why do such loss discrepancies exist? The literature generally studies sources of loss discrepancy due to an "information deficiency" of one group--that is, one group has, for example, more noise (Corbett-Davies et al., 2017), lessPreliminary work, under review.