This position paper argues that, in order to understand AI, we cannot rely on our existing vocabulary of human words. Instead, we should strive to develop neologisms: new words that represent precise human concepts that we want to teach machines, or machine concepts that we need to learn. We start from the premise that humans and machines have differing concepts. This means interpretability can be framed as a communication problem: humans must be able to reference and control machine concepts, and communicate human concepts to machines. Creating a shared human-machine language through developing neologisms, we believe, could solve this communication problem. Successful neologisms achieve a useful amount of abstraction: not too detailed, so they're reusable in many contexts, and not too high-level, so they convey precise information. As a proof of concept, we demonstrate how a "length neologism" enables controlling LLM response length, while a "diversity neologism" allows sampling more variable responses. Taken together, we argue that we cannot understand AI using our existing vocabulary, and expanding it through neologisms creates opportunities for both controlling and understanding machines better.

What do artificial neural networks (ANNs) learn? The machine learning (ML) community shares the narrative that ANNs must develop abstract human concepts to perform complex tasks. Some go even further and believe that these concepts are stored in individual units of the network. Based on current research, I systematically investigate the assumptions underlying this narrative. I conclude that ANNs are indeed capable of performing complex prediction tasks, and that they may learn human and non-human concepts to do so. However, evidence indicates that ANNs do not represent these concepts in individual units.

Recent years have witnessed important advancements in our understanding of the psychological underpinnings of subjective properties of visual information, such as aesthetics, memorability, or induced emotions. Concurrently, computational models of objective visual properties such as semantic labelling and geometric relationships have made significant breakthroughs using the latest achievements in machine learning and large-scale data collection. There has also been limited but important work exploiting these breakthroughs to improve computational modelling of subjective visual properties. The time is ripe to explore how advances in both of these fields of study can be mutually enriching and lead to further progress. This book combines perspectives from psychology and machine learning to showcase a new, unified understanding of how images and videos influence high-level visual perception - particularly interestingness, affective values and emotions, aesthetic values, memorability, novelty, complexity, visual composition and stylistic attributes, and creativity.

DeepMind and Google Brain researchers and former World Chess Champion Vladimir Kramnik explore how human knowledge is acquired and how chess concepts are represented in the AlphaZero neural network via concept probing, behavioral analysis, and an examination of its activations. The world has quietly crowned a new chess champion. While it has now been over two decades since a human has been honored with that title, the latest victor represents a breakthrough in another significant way: It's an algorithm that can be generalized to other learning tasks. AlphaZero, the new reigning champion, acquired all its chess know-how in a mere four hours. AlphaZero is almost as different from its fellow AI chess competitors as Deep Blue was from Gary Kasparov, back when the latter first faced off against a supercomputer in 1996.

What is learned by sophisticated neural network agents such as AlphaZero? This question is of both scientific and practical interest. If the representations of strong neural networks bear no resemblance to human concepts, our ability to understand faithful explanations of their decisions will be restricted, ultimately limiting what we can achieve with neural network interpretability. In this work we provide evidence that human knowledge is acquired by the AlphaZero neural network as it trains on the game of chess. By probing for a broad range of human chess concepts we show when and where these concepts are represented in the AlphaZero network. We also provide a behavioural analysis focusing on opening play, including qualitative analysis from chess Grandmaster Vladimir Kramnik. Finally, we carry out a preliminary investigation looking at the low-level details of AlphaZero's representations, and make the resulting behavioural and representational analyses available online.

Deep neural networks are known to learn opaque, uninterpretable representations that lie beyond the grasp of human understanding. As such, from both scientific and practical viewpoints, it is intriguing to explore what is actually being learned and how in the case of superhuman self-taught neural network agents such as AlphaZero. In the new paper Acquisition of Chess Knowledge in AlphaZero, DeepMind and Google Brain researchers and former World Chess Champion Vladimir Kramnik explore how and to what extent human knowledge is acquired by AlphaZero and how chess concepts are represented in its network. They do this via comprehensive concept probing, behavioural analysis, and examination of AlphaZero's activations. The researchers premise their study with the idea that if the representations of strong neural networks like AlphaZero bear no resemblance to human concepts, our ability to understand faithful explanations of their decisions will be restricted, ultimately limiting what we can achieve with neural network interpretability. This enables them to build up a picture of what is learned, when it was learned during training, and where in the network it is computed.

If you're handed a note that asks you to draw a picture of a bird with a yellow body, black wings and a short beak, chances are you'll start with a rough outline of a bird, then glance back at the note, see the yellow part and reach for a yellow pen to fill in the body, read the note again and reach for a black pen to draw the wings and, after a final check, shorten the beak and define it with a reflective glint. Then, for good measure, you might sketch a tree branch where the bird rests. Now, there's a bot that can do that, too. The new artificial intelligence technology under development in Microsoft's research labs is programmed to pay close attention to individual words when generating images from caption-like text descriptions. This deliberate focus produced a nearly three-fold boost in image quality compared to the previous state-of-the-art technique for text-to-image generation, according to results on an industry standard test reported in a research paper posted on arXiv.org.

Sophisticated autonomous AI may need to base its behavior on fuzzy concepts such as well-being or rights. These concepts cannot be given an explicit formal definition, but obtaining desired behavior still requires a way to instill the concepts in an AI system. To solve the problem, we review evidence suggesting that the human brain generates its concepts using a relatively limited set of rules and mechanisms. This suggests that it might be feasible to build AI systems that use similar criteria for generating their own concepts, and could thus learn similar concepts as humans do. Major challenges to this approach include the embodied nature of human thought, evolutionary vestiges in cognition, the social nature of concepts, and the need to compare conceptual representations between humans and AI systems.