Imitating the Agile Locomotion Skills of Four-legged Animals

Imitating the agile locomotion skills of animals has been a longstanding challenge in robotics. Manually-designed controllers have been able to reproduce many complex behaviors, but building such controllers is time-consuming and difficult. According to Xue Bin Peng (Google Research and University of California, Berkeley) and his co-authors, reinforcement learning provides an interesting alternative for automating the manual effort involved in the development of controllers. In their work, they present “an imitation learning system that enables legged robots to learn agile locomotion skills by imitating real-world animals” (Xue Bin Peng et al. 2020). They show “that by leveraging reference motion data, a single learning-based approach is able to automatically synthesize controllers for a diverse repertoire behaviors for legged robots” (Xue Bin Peng et al. 2020). By incorporating sample efficient domain adaptation techniques into the training process, their system “is able to learn adaptive policies in simulation that can then be quickly adapted for real-world deployment” (Xue Bin Peng et al. 2020). For demonstration purposes, the scientists trained “a quadruped robot to perform a variety of agile behaviors ranging from different locomotion gaits to dynamic hops and turns” (Xue Bin Peng et al. 2020).

Towards a Human-like Chatbot

Google is currently working on Meena, a particular chatbot, which should be able to have arbitrary conversations and be used in many contexts. In their paper “Towards a Human-like Open-Domain Chatbot“, the developers present the 2.6 billion parameters end-to-end trained neural conversational model. They show that Meena “can conduct conversations that are more sensible and specific than existing state-of-the-art chatbots”. “Such improvements are reflected through a new human evaluation metric that we propose for open-domain chatbots, called Sensibleness and Specificity Average (SSA), which captures basic, but important attributes for human conversation. Remarkably, we demonstrate that perplexity, an automatic metric that is readily available to any neural conversational models, highly correlates with SSA.” (Google AI Blog) The company draws a comparison with OpenAI GPT-2, a model used in “Talk to Transformer” and Harmony, among others, which uses 1.5 billion parameters and is based on the text content of 8 million web pages.

Robots that Learn as They Go

“Alphabet X, the company’s early research and development division, has unveiled the Everyday Robot project, whose aim is to develop a ‘general-purpose learning robot.’ The idea is to equip robots with cameras and complex machine-learning software, letting them observe the world around them and learn from it without needing to be taught every potential situation they may encounter.” (MIT Technology Review, 23 November 2019) This was reported by MIT Technology Review on 23 November 2019 in the article “Alphabet X’s ‘Everyday Robot’ project is making machines that learn as they go”. The approach of Alphabet X seems to be well though-out and target-oriented. In a way, it is oriented towards human learning. One could also teach robots human language in this way. With the help of microphones, cameras and machine learning, they would gradually understand us better and better. For example, they observe how we point to and comment on a person. Or they perceive that we point to an object and say a certain term – and after some time they conclude that this is the name of the object. However, such frameworks pose ethical and legal challenges. You can’t just designate cities as such test areas. The result would be comprehensive surveillance in public spaces. Specially established test areas, on the other hand, would probably not have the same benefits as “natural environments”. Many questions still need to be answered.

Talk to Transformer

Artificial intelligence is spreading into more and more application areas. American scientists have now developed a system that can supplement texts: “Talk to Transformer”. The user enters a few sentences – and the AI system adds further passages. “The system is based on a method called DeepQA, which is based on the observation of patterns in the data. This method has its limitations, however, and the system is only effective for data on the order of 2 million words, according to a recent news article. For instance, researchers say that the system cannot cope with the large amounts of data from an academic paper. Researchers have also been unable to use this method to augment texts from academic sources. As a result, DeepQA will have limited application, according to the researchers. The scientists also note that there are more applications available in the field of text augmentation, such as automatic transcription, the ability to translate text from one language to another and to translate text into other languages.” The sentences in quotation marks are not from the author of this blog. They were written by the AI system itself. You can try it via talktotransformer.com.

Learning How to Behave

In October 2019 Springer VS published the “Handbuch Maschinenethik” (“Handbook Machine Ethics”) with German and English contributions. Editor is Oliver Bendel (Zurich, Switzerland). One of the articles was written by Bertram F. Malle (Brown University, Rhode Island) and Matthias Scheutz (Tufts University, Massachusetts). From the abstract: “We describe a theoretical framework and recent research on one key aspect of robot ethics: the development and implementation of a robot’s moral competence. As autonomous machines take on increasingly social roles in human communities, these machines need to have some level of moral competence to ensure safety, acceptance, and justified trust. We review the extensive and complex elements of human moral competence and ask how analogous competences could be implemented in a robot. We propose that moral competence consists of five elements, two constituents (moral norms and moral vocabulary) and three activities (moral judgment, moral action, and moral communication). A robot’s computational representations of social and moral norms is a prerequisite for all three moral activities. However, merely programming in advance the vast network of human norms is impossible, so new computational learning algorithms are needed that allow robots to acquire and update the context-specific and graded norms relevant to their domain of deployment. Moral vocabulary is needed primarily for moral communication, which expresses moral judgments of others’ violations and explains one’s own moral violations – to justify them, apologize, or declare intentions to do better. Current robots have at best rudimentary moral competence, but with improved learning and reasoning they may begin to show the kinds of capacities that humans will expect of future social robots.” (Abstract “Handbuch Maschinenethik”). The book is available via www.springer.com.

Honey, I shrunk the AI

Some months ago, researchers at the University of Massachusetts showed the climate toll of machine learning, especially deep learning. Training Google’s BERT, with its 340 million data parameters, emitted nearly as much carbon as a round-trip flight between the East and West coasts. According to Technology Review, the trend could also accelerate the concentration of AI research into the hands of a few big tech companies. “Under-resourced labs in academia or countries with fewer resources simply don’t have the means to use or develop such computationally expensive models.” (Technology Review, 4 October 2019) In response, some researchers are focused on shrinking the size of existing models without losing their capabilities. The magazine wrote enthusiastically: “Honey, I shrunk the AI” (Technology Review, 4 October 2019) There are advantages not only with regard to the environment and to the access to state-of-the-art AI. According to Technology Review, tiny models will help bring the latest AI advancements to consumer devices. “They avoid the need to send consumer data to the cloud, which improves both speed and privacy. For natural-language models specifically, more powerful text prediction and language generation could improve myriad applications like autocomplete on your phone and voice assistants like Alexa and Google Assistant.” (Technology Review, 4 October 2019)

Towards Full Body Fakes

“Within the field of deepfakes, or ‘synthetic media’ as researchers call it, much of the attention has been focused on fake faces potentially wreaking havoc on political reality, as well as other deep learning algorithms that can, for instance, mimic a person’s writing style and voice. But yet another branch of synthetic media technology is fast evolving: full body deepfakes.” (Fast Company, 21 September 2019) Last year, researchers from the University of California Berkeley demonstrated in an impressive way how deep learning can be used to transfer dance moves from a professional onto the bodies of amateurs. Also in 2018, a team from the University of Heidelberg published a paper on teaching machines to realistically render human movements. And in spring of this year, a Japanese company developed an AI that can generate whole body models of nonexistent persons. “While it’s clear that full body deepfakes have interesting commercial applications, like deepfake dancing apps or in fields like athletics and biomedical research, malicious use cases are an increasing concern amid today’s polarized political climate riven by disinformation and fake news.” (Fast Company, 21 September 2019) Was anyone really in this area, did he or she really take part in a demonstration and throw stones? In the future you won’t know for sure.

The New Dangers of Face Recognition

The dangers of face recognition are discussed more and more. A new initiative is aimed at banning the use of technology to monitor the American population. The AI Now Institute already warned of the risks in 2018, as did Oliver Bendel. The ethicist had a particular use in mind. In the 21st century, face recognition is increasingly attempted to connect to the pseudoscience of physiognomy, which has its origins in ancient times. From the appearance of persons, a conclusion is drawn to their inner self, and attempts are made to identify character traits, personality traits and temperament, or political and sexual orientation. Biometrics plays a role in this concept. It was founded in the eighteenth century, when physiognomy under the lead of Johann Caspar Lavater had its dubious climax. In his paper “The Uncanny Return of Physiognomy”, Oliver Bendel elaborates the basic principles of this topic; selected projects from research and practice are presented and, from an ethical perspective, the possibilities of face recognition are subjected to fundamental critique in this context, including the above examples. The philosopher presented his paper on 27 March 2018 at Stanford University (“AI and Society: Ethics, Safety and Trustworthiness in Intelligent Agents”, AAAI 2018 Spring Symposium Series). The PDF is available here.

Fighting Deepfakes with Deepfakes

A deepfake (or deep fake) is a picture or video created with the help of artificial intelligence that looks authentic but is not. Also the methods and techniques in this context are labeled with the term. Machine learning and especially deep learning are used. With deepfakes one wants to create objects of art and visual objects or means for discreditation, manipulation and propaganda. Politics and pornography are therefore closely interwoven with the phenomenon. According to Futurism, Facebook is teaming up with a consortium of Microsoft researchers and several prominent universities for a “Deepfake Detection Challenge”. “The idea is to build a data set, with the help of human user input, that’ll help neural networks detect what is and isn’t a deepfake. The end result, if all goes well, will be a system that can reliably fake videos online. Similar data sets already exist for object or speech recognition, but there isn’t one specifically made for detecting deepfakes yet.” (Futurism, 5 September 2019) The winning team will get a prize – presumably a higher sum of money. Facebook is investing a total of 10 million dollars in the competition.

An AI System for Multiple-choice Tests

According to the New York Times, the Allen Institute for Artificial Intelligence unveiled a new system that correctly answered more than 90 percent of the questions on an eighth-grade science test and more than 80 percent on a 12th-grade exam. Is it really a breakthrough for AI technology, as the title of the article claims? This is quite controversial among experts. The newspaper is optimistic: “The system, called Aristo, is an indication that in just the past several months researchers have made significant progress in developing A.I. that can understand languages and mimic the logic and decision-making of humans.” (NYT, 4 September 2019) Aristo was built for multiple-choice tests. “It took standard exams written for students in New York, though the Allen Institute removed all questions that included pictures and diagrams.” (NYT, 4 September 2019) Some questions could be answered by simple information retrieval. There are numerous systems that access Google and Wikipedia, including artifacts of machine ethics like LIEBOT and BESTBOT. But for the answers to other questions logical thinking was required. Perhaps Aristo is helping to abolish multiple-choice tests – not so much because it can solve them, but because they are often not effective.