Machine Learning is the study of methods for programming computers to learn. Computers are applied to a wide range of tasks, and for most of these it is relatively easy for programmers to design and implement the necessary software. However, there are many tasks for which this is difficult or impossible. These can be divided into four general categories. First, there are problems for which there exist no human experts. For example, in modern automated manufacturing facilities, there is a need to predict machine failures before they occur by analyzing sensor readings. Because the machines are new, there are no human experts who can be interviewed by a programmer to provide the knowledge necessary to build a computer system. A machine learning system can study recorded data and subsequent machine failures and learn prediction rules. Second, there are problems where human experts exist, but where they are unable to explain their expertise. This is the case in many perceptual tasks, such as speech recognition, hand-writing recognition, and natural language understanding. Virtually all humans exhibit expert-level abilities on these tasks, but none of them can describe the detailed steps that they follow as they perform them. Fortunately, humans can provide machines with examples of the inputs and correct outputs for these tasks, so machine learning algorithms can learn to map the inputs to the outputs. Third, there are problems where phenomena are changing rapidly. In finance, for example, people would like to predict the future behavior of the stock market, of consumer purchases, or of exchange rates. These behaviors change frequently, so that even if a programmer could construct a good predictive computer program, it would need to be rewritten frequently. A learning program can relieve the programmer of this burden by constantly modifying and tuning a set of learned prediction rules. Fourth, there are applications that need to be customized for each computer user separately. Consider, for example, a program to filter unwanted electronic mail messages. Different users will need different filters. It is unreasonable to expect each user to program his or her own rules, and it is infeasible to provide every user with a software engineer to keep the rules up-to-date. A machine learning system can learn which mail messages the user rejects and maintain the filtering rules automatically. Machine learning addresses many of the same research questions as the fields of statistics, data mining, and psychology, but with differences of emphasis. Statistics focuses on understanding the phenomena that have generated the data, often with the goal of testing different hypotheses about those phenomena. Data mining seeks to find patterns in the data that are understandable by people. Psychological studies of human learning aspire to understand the mechanisms underlying the various learning behaviors exhibited by people (concept learning, skill acquisition, strategy change, etc.).

Machine learning

Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. To address this limitation, we introduce "deep compression", a three stage pipeline: pruning, trained quantization and Huffman coding, that work together to reduce the storage requirement of neural networks by 35x to 49x without affecting their accuracy. Our method first prunes the network by learning only the important connections. Next, we quantize the weights to enforce weight sharing, finally, we apply Huffman coding. After the first two steps we retrain the network to fine tune the remaining connections and the quantized centroids. Pruning, reduces the number of connections by 9x to 13x; Quantization then reduces the number of bits that represent each connection from 32 to 5. On the ImageNet dataset, our method reduced the storage required by AlexNet by 35x, from 240MB to 6.9MB, without loss of accuracy. Our method reduced the size of VGG-16 by 49x from 552MB to 11.3MB, again with no loss of accuracy. This allows fitting the model into on-chip SRAM cache rather than off-chip DRAM memory. Our compression method also facilitates the use of complex neural networks in mobile applications where application size and download bandwidth are constrained. Benchmarked on CPU, GPU and mobile GPU, compressed network has 3x to 4x layerwise speedup and 3x to 7x better energy efficiency.

https://forum.stanford.edu/events/posterslides/DeepCompressionCompressingDeepNeuralNetworkswithPruning,TrainedQuantizationandHuffmanCoding.pdf

Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding

Planning algorithms are impacting technical disciplines and industries around the world, including robotics, computer-aided design, manufacturing, computer graphics, aerospace applications, drug design, and protein folding. This coherent and comprehensive book unifies material from several sources, including robotics, control theory, artificial intelligence, and algorithms. The treatment is centered on robot motion planning but integrates material on planning in discrete spaces. A major part of the book is devoted to planning under uncertainty, including decision theory, Markov decision processes, and information spaces, which are the “configuration spaces” of all sensor-based planning problems. The last part of the book delves into planning under differential constraints that arise when automating the motions of virtually any mechanical system. Developed from courses taught by the author, the book is intended for students, engineers, and researchers in robotics, artificial intelligence, and control theory as well as computer graphics, algorithms, and computational biology.

Planning Algorithms: Introductory Material

Model Checking

A classic introduction to artificial intelligence intended to bridge the gap between theory and practice, "Principles of Artificial Intelligence" describes fundamental AI ideas that underlie applications such as natural language processing, automatic programming, robotics, machine vision, automatic theorem proving, and intelligent data retrieval. Rather than focusing on the subject matter of the applications, the book is organized around general computational concepts involving the kinds of data structures used, the types of operations performed on the data structures, and the properties of the control strategies used. "Principles of Artificial Intelligence"evolved from the author's courses and seminars at Stanford University and University of Massachusetts, Amherst, and is suitable for text use in a senior or graduate AI course, or for individual study.

Principles of Artificial Intelligence

We resolve an old problem, namely to design a ‘natural’ machine model for accepting the complements of recursively enumerable languages.

Turing machines with one-sided advice and the acceptance of the co-RE languages

AI research is continually challenged to explain cognitive processes as being computational. Whereas existing notions of computing seem to have their limits for it, we contend that the recent, epistemic approach to computations may hold the key to understanding cognition from this perspective. In this approach, computations are seen as processes generating knowledge over a suitable knowledge domain, within the framework of a suitable knowledge theory. This, machine-independent, understanding of computation allows us to explain a variety of higher cognitive functions such as accountability, self-awareness, introspection, free will, creativity, anticipation and curiosity in computational terms. It also opens the way to understanding the self-improving mechanisms behind the development of intelligence. The argumentation does not depend on any technological analogies.

Epistemic Computation and Artificial Intelligence

First we review amortized fully-dynamic polylogarithmic algorithms for connectivity, minimum spanning trees (MST), 2-edgeand biconnectivity. Second we discuss how they yield improved static algorithms: connectivity for constructing a tree from homeomorphic subtrees, 2-edge connectivity for finding unique matchings in graphs, and MST for packing spanning trees in graphs. The application of MST for spanning tree packing is new and when boot-strapped, it yields a fully-dynamic polylogarithmic algorithm for approximating general edge connectivity within a factor √ 2 + o(1). Finally, on the more practical side, we will discuss how output sensitive algorithms for dynamic shortest paths have been applied successfully to speed up local search algorithms for improving routing on the internet, roughly doubling the capacity. 1 Dynamic Graph Algorithms In this talk, we will discuss some simple dynamic graph algorithms and their applications within static graph problems. As a new result, we will derive a fully dynamic polylogarithmic algorithm approximating the edge connectivity λ within a factor √ 2 + o(1), that is, the algorithm will output a value between λ/ √ 2 + o(1) and λ ×2 + o(1). The talk is not intended as a general survey of dynamic graph algorithms and their applications. Rather its goal is just to present a few nice illustrations of the potent relationship between dynamic graph algorithms and their applications in static graph problems, showing contexts in which dynamic graph algorithms play a role similar to that played by priority queues for greedy algorithms. In a fully dynamic graph problem, we are considering a graph G over a fixed vertex set V , |V | = n. The graph G may be updated by insertions and deletions of edges. Unless otherwise stated, we assume that we start with an empty edge set. We will review the fully dynamic graph algorithms of Holm et al. [11] for connectivity, minimum spanning trees (MST), 2-edge, and biconnectivity in undirected graphs. For the connectivity type problems, the updates may be interspersed by queries on (2-edge-/bi-) connectivity of the graph or between specified vertices. For MST, the fully dynamic algorithm should update the MST in connection with each update to the graph: an inserted edge might have to go into the MST, and if an MST edge is deleted, we should replace with the lightest edge possible. M.M. Halldórsson (Ed.): SWAT 2000, LNCS 1851, pp. 1–9, 2000. c © Springer-Verlag Berlin Heidelberg 2000 2 M. Thorup and D.R. Karger Both updates and queries are presented on-line, meaning that we have to respond to an update or query without knowing anything about the future. The time bounds for these operations are polylogarithmic but amortized meaning that we only bound the average operation time over any sequence of operations, starting with no edges. In our later applications for static graph problems, we only care about the total amount of time spent over all dynamic graph operations, and hence the amortized time bounds suffice. The above mentioned results are all for undirected graphs. For directed graphs there are very few results. In a recent break-through, King [16] showed how to maintain the full transitive closure of a graph in Õ(n2) amortized time per update. Further, she showed how to maintain all pairs shortest paths in O(n2.5 √ log C) time per update if C is the maximum weight in the graph. However, if one is is just interested in maintaining whether t can be reached from s for two fixed vertices s and t, nobody knows how to do this in o(m) time. On the more practical side, Ramalingan and Reps [24] have suggested a lazy implementation of Dijkstra’s [4] single source shortest paths algorithm for a dynamic directed graph. If X is the number of vertices that change distance from the source s in connection with an arc insertion or deletion, they can update a shortest path tree from s in Õ( ∑ v∈X degree(v)) time. Although this does not in general improve over the Õ(m) time it takes to compute a single source shortest path tree from scratch, there has been experimental evidence suggesting that this kind of laziness is worthwhile in connection with internet like topologies [7].

Algorithm Theory - SWAT 2000

In this paper, we consider a Video-on-Demand storage server which operates in a heterogeneous environment and faces (a) fluctuations in workload, (b) network congestion, and (c) failure of server and/or network components. network). We present scheduling of data retrieval techniques which are capable of dynamically resolving congestion which occurs on the storage subsystem due to expected and unexpected changes in I/O bandwidth demand for a scalable video server in a multi-disk environment. techniques.

Advances in Multimedia Information Systems

In the late nineteen sixties it was observed that the r.e. languages form an infinite proper hierarchy RE1⊂RE2⊂⋯ based on the size of the Turing machines that accept them. We examine the fundamental position of the finite languages and their complements in the hierarchy. We show that for every finite language L one has that L, L¯∈REn for some n≤p⋅(m−⌊ log⁡2p ⌋+1)+1 where m is the length of the longest word in L, c is the cardinality of L, and p=min⁡{c,2m−1}. If L∈REn, then L¯∈REs for some s=O(n+m). We also prove that for every n, there is a finite language Ln with m=O(nlog⁡2n) such that Ln∉REn but Ln, L¯n∈REs for some s=O(nlog⁡2n). Several further results are shown that how the hierarchy can be separated by increasing chains of finite languages. The proofs make use of several auxiliary results for Turing machines with advice.

/pdf/separating-the-classes-of-recursively-enumerable-languages-1j17mtat7h.pdf

Jan van Leeuwen

Papers

Turing machines with one-sided advice and the acceptance of the co-RE languages

Epistemic Computation and Artificial Intelligence

Algorithm Theory - SWAT 2000

Advances in Multimedia Information Systems

Separating the Classes of Recursively Enumerable Languages Based on Machine Size