'Vital reading. This is the book on artificial intelligence we need right now.' Mike Krieger, cofounder of Instagram Artificial intelligence is rapidly dominating every aspect of our modern lives influencing the news we consume, whether we get a mortgage, and even which friends wish us happy birthday. But as algorithms make ever more decisions on our behalf, how do we ensure they do what we want? And fairly? This conundrum - dubbed 'The Alignment Problem' by experts - is the subject of this timely and important book. From the AI program which cheats at computer games to the sexist algorithm behind Google Translate, bestselling author Brian Christian explains how, as AI develops, we rapidly approach a collision between artificial intelligence and ethics. If we stand by, we face a future with unregulated algorithms that propagate our biases - and worse - violate our most sacred values. Urgent and fascinating, this is an accessible primer to the most important issue facing AI researchers today.
Neuronale Netze sind Schlüsselelemente des Deep Learning und der Künstlichen Intelligenz, die heute zu Erstaunlichem in der Lage sind. Dennoch verstehen nur wenige, wie Neuronale Netze tatsächlich funktionieren. Dieses Buch nimmt Sie mit auf eine unterhaltsame Reise, die mit ganz einfachen Ideen beginnt und Ihnen Schritt für Schritt zeigt, wie Neuronale Netze arbeiten. Dafür brauchen Sie keine tieferen Mathematik-Kenntnisse, denn alle mathematischen Konzepte werden behutsam und mit vielen Illustrationen erläutert. Dann geht es in die Praxis: Sie programmieren Ihr eigenes Neuronales Netz mit Python und bringen ihm bei, handgeschriebene Zahlen zu erkennen, bis es eine Performance wie ein professionell entwickeltes Netz erreicht. Zum Schluss lassen Sie das Netz noch auf einem Raspberry Pi Zero laufen. - Tariq Rashid hat eine besondere Fähigkeit, schwierige Konzepte verständlich zu erklären, dadurch werden Neuronale Netze für jeden Interessierten zugänglich und praktisch nachvollziehbar.
This dissertation addresses the problem of searching a target within a region by sequential queries with noisy responses. A Bayesian decision maker is responsible to collect observation samples so as to enhance his knowledge about the true location in a speedy manner. When the response is noiseless, the classical binary search solves such a problem optimally. Noisy binary search, on the other hand, has also been formulated and studied extensively in theory over the past 60 years since Horstein (1963). However, the algorithms developed in noisy binary search problem find limited practical applications in real-world engineer problem. Motivated by bridging theory and practice, we formulate the noisy binary search problem by identifying practical scenarios and constraints that naturally rises with practical applications such as spectrum sensing in cognitive communication, AoA estimation by adaptive beamforming in large antenna array system, visual image inspection, bit-wise data transmission, heavy hitter detection in network system, etc. The first part of the dissertation (Chapter 2) focuses on theoretical understanding and developing noisy binary search algorithms under those practical constraints. Three algorithms sortPM, dyaPM, hiePM are proposed. Using the extrinsic Jensen Divergence from information theory, we provide upper bound for the expected search time of each of the algorithms. By comparing with an information theoretic lower bound, we demonstrate the asymptotic optimality and suboptimality of the proposed algorithms (asymptotic in the resolution of the target location). The second part of the dissertation applies the proposed hiePM to practical problems. In particular, Chapter 3 demonstrates the application of hiePM on the data transmission problem with noiseless feedback. The dyadic hierarchical query area of hiePM relates directly to the bit representation of the data stream. This simplifies significantly the corresponding adaptive encoding scheme and allows a bit-wise encoding. Chapter 4 considers the initial beam alignment problem in 5G mmWave communication using beamforming. With a single-path channel model, the problem is reduced to actively searching the Angle-of-Arrival (AoA) of the signal sent from the user to the Base Station (BS). hiePM is applied to adaptively and sequentially choose the beamforming from the hierarchical beamforming codebook. The proposed algorithm is compared to prior works of initial beam alignment that employs linear beam search, repeat binary search, or random beam search, respectively, and gives the state-of-art performance in terms of both AoA estimation error at the end of the initial alignment, and the spectral efficiency during the communication phase.
This thesis provides a novel conceptual contribution to artificial intelligence (AI) safety by finding a tractable method for solving the AI value alignment problem: the creation of more complete audience models using narrative information extraction techniques from the field of computational narratology. With a thorough analysis of results from the field of computational narratology, I show that research into narrative for autonomous agents can contribute to solving the AI value alignment problem. In short, we can create artificial intelligence systems that automatically act in the best interest of humanity by teaching them to read and understand stories.The novelty of this thesis lies in the combination of two disparate academic fields: AI safety and computational narratology. Reviewing the current work and ongoing issues in both fields, I show that methods used in computational narratology to model stories can be used to solve the value alignment problem from the field of AI safety. In Chapter 2, I show why value alignment is the best solution to the problem of controlling intelligent agents. In Chapter 2, I discuss how stories encode tacit human values, and how the creation of a better audience model will contribute to solving the value alignment problem. In Chapter 3, I present two case studies providing evidence that value alignment from narrative information extraction is not only viable, but effective. Finally, I conclude by acknowledging the shortcomings of the field and pressing areas of future work.
Aggregate data objects (such as arrays) are distributed across the processor memories when compiling a data-parallel language for a distributed-memory machine. The mapping determines the amount of communication needed to bring operands of parallel operations into alignment with each other. A common approach is to break the mapping into two stages: an alignment that maps all the objects to an abstract template, followed by a distribution that maps the template to the processors. This paper describes algorithms for solving the various facets of the alignment problem: axis and stride alignment, static and mobile offset alignment, and replication labeling. We show that optimal axis and stride alignment is NP-complete for general program graphs, and give a heuristic method that can explore the space of possible solutions in a number of ways. We show that some of these strategies can give better solutions than a simple greedy approach proposed earlier. We also show how local graph contractions can reduce the size of the problem significantly without changing the best solution. This allows more complex and effective heuristics to be used. We show how to model the static offset alignment problem using linear programming, and we show that loop-dependent mobile offset alignment is sometimes necessary for optimum performance. We describe an algorithm with for determining mobile alignments for objects within do loops. We also identify situations in which replicated alignment is either required by the program itself or can be used to improve performance. We describe an algorithm based on network flow that replicates objects so as to minimize the total amount of broadcast communication in replication. Chatterjee, Siddhartha and Gilbert, John R. and Oliker, Leonid and Schreiber, Robert and Sheffler, Thomas J. Ames Research Center NAS2-13721...