Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers—8× deeper than VGG nets [40] but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation.

/pdf/deep-residual-learning-for-image-recognition-b75jg61qrq.pdf

Deep Residual Learning for Image Recognition

Learning to store information over extended time intervals by recurrent backpropagation takes a very long time, mostly because of insufficient, decaying error backflow. We briefly review Hochreiter's (1991) analysis of this problem, then address it by introducing a novel, efficient, gradient based method called long short-term memory (LSTM). Truncating the gradient where this does not do harm, LSTM can learn to bridge minimal time lags in excess of 1000 discrete-time steps by enforcing constant error flow through constant error carousels within special units. Multiplicative gate units learn to open and close access to the constant error flow. LSTM is local in space and time; its computational complexity per time step and weight is O. 1. Our experiments with artificial data involve local, distributed, real-valued, and noisy pattern representations. In comparisons with real-time recurrent learning, back propagation through time, recurrent cascade correlation, Elman nets, and neural sequence chunking, LSTM leads to many more successful runs, and learns much faster. LSTM also solves complex, artificial long-time-lag tasks that have never been solved by previous recurrent network algorithms.

Long short-term memory

In this work we investigate the effect of the convolutional network depth on its accuracy in the large-scale image recognition setting. Our main contribution is a thorough evaluation of networks of increasing depth using an architecture with very small (3x3) convolution filters, which shows that a significant improvement on the prior-art configurations can be achieved by pushing the depth to 16-19 weight layers. These findings were the basis of our ImageNet Challenge 2014 submission, where our team secured the first and the second places in the localisation and classification tracks respectively. We also show that our representations generalise well to other datasets, where they achieve state-of-the-art results. We have made our two best-performing ConvNet models publicly available to facilitate further research on the use of deep visual representations in computer vision.

Very Deep Convolutional Networks for Large-Scale Image Recognition

Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers---8x deeper than VGG nets but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. 
The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation.

Deep learning is a form of machine learning that enables computers to learn from experience and understand the world in terms of a hierarchy of concepts. Because the computer gathers knowledge from experience, there is no need for a human computer operator to formally specify all the knowledge that the computer needs. The hierarchy of concepts allows the computer to learn complicated concepts by building them out of simpler ones; a graph of these hierarchies would be many layers deep. This book introduces a broad range of topics in deep learning. The text offers mathematical and conceptual background, covering relevant concepts in linear algebra, probability theory and information theory, numerical computation, and machine learning. It describes deep learning techniques used by practitioners in industry, including deep feedforward networks, regularization, optimization algorithms, convolutional networks, sequence modeling, and practical methodology; and it surveys such applications as natural language processing, speech recognition, computer vision, online recommendation systems, bioinformatics, and videogames. Finally, the book offers research perspectives, covering such theoretical topics as linear factor models, autoencoders, representation learning, structured probabilistic models, Monte Carlo methods, the partition function, approximate inference, and deep generative models. Deep Learning can be used by undergraduate or graduate students planning careers in either industry or research, and by software engineers who want to begin using deep learning in their products or platforms. A website offers supplementary material for both readers and instructors.

Deep Learning

Nonword pronunciation is a form of generalization behavior that has been at the center of debates about models of word recognition, the role of rules in explaining behavior, and the adequacy of the parallel distributed processing approach. An experiment yielded data concerning the pronunciation of a large corpus of nonwords. The data were then used to assess 2 models of naming: a model developed by D. C. Plaut and J. L. McClelland (1993), which is similar to the one described by M. S. Seidenberg and J. L. McClelland (1989) but uses improved orthographic and phonological representations, and the grapheme-phoneme correspondence rules of M. Coltheart, B. Curtis, P. Atkins, and M. Haller's (1993) dual-route model. Both models generate plausible nonword pronunciations and match subjects' responses accurately. The dual-route model does so by using rules that generate correct output for most words but mispronounce a significant number of exceptions. The parallel distributed processing model does so by finding a set of weights that allow it to generate correct output for both "rule-governed" items and exceptions. Some ways in which the two approaches differ and other issues facing them are also discussed. The task of reading words and nonwords aloud has played a central role in the development of models of word recognition. Reading makes use of knowledge concerning the correspondences between the orthographic and phonological forms of words. This information is used in recognizing words and pronouncing them aloud (see Seidenberg, in press, for a review). Current models differ in their assumptions about how this knowledge is acquired, represented, and used. Dual-route models assume that there are separate lexical and sublexical procedures for generating pronunciations (for an overview, see Patterson & Coltheart, 1987; for critiques, see Humphreys & Evett, 1985, and Van Orden, Pennington, & Stone, 1990). The specific version of the dual-route model developed by Coltheart and his colleagues (Coltheart, 1978,1987; Coltheart, Curtis, Atkins, & Haller, 1993) assumes that in alphabetic writing systems, knowledge of the correspondences between orthography and phonology is represented in terms of rules translating graphemes into phonemes. The rules are used in naming words whose pronunciations they correctly specify (sometimes

Nonword pronunciation and models of word recognition.

Semantic cognition requires conceptual representations shaped by verbal and nonverbal experience and executive control processes that regulate activation of knowledge to meet current situational demands. A complete model must also account for the representation of concrete and abstract words, of taxonomic and associative relationships, and for the role of context in shaping meaning. We present the first major attempt to assimilate all of these elements within a unified, implemented computational framework. Our model combines a hub-and-spoke architecture with a buffer that allows its state to be influenced by prior context. This hybrid structure integrates the view, from cognitive neuroscience, that concepts are grounded in sensory-motor representation with the view, from computational linguistics, that knowledge is shaped by patterns of lexical co-occurrence. The model successfully codes knowledge for abstract and concrete words, associative and taxonomic relationships, and the multiple meanings of homonyms, within a single representational space. Knowledge of abstract words is acquired through (a) their patterns of co-occurrence with other words and (b) acquired embodiment, whereby they become indirectly associated with the perceptual features of co-occurring concrete words. The model accounts for executive influences on semantics by including a controlled retrieval mechanism that provides top-down input to amplify weak semantic relationships. The representational and control elements of the model can be damaged independently, and the consequences of such damage closely replicate effects seen in neuropsychological patients with loss of semantic representation versus control processes. Thus, the model provides a wide-ranging and neurally plausible account of normal and impaired semantic cognition. (PsycINFO Database Record

/pdf/concepts-control-and-context-a-connectionist-account-of-ls8cct62fd.pdf

Concepts, control and context: A connectionist account of normal and disordered semantic cognition

An experiment is reported confirming the existence of the “word-letter phenomenon” (WLP): At tachistoscopic exposure durations, each letter of a four-letter word is perceived more accurately than a single letter. Data obtained rule out several artifactual interpretations, including the possibility that perception of letters in a word is facilitated merely by the presence of adjacent contours. The WLP is shown to depend critically on what type of display is used as a preand postexposure field. While a masking field of high-contrast random contours produced a large and reliable WLP, a plain white field eliminated the phenomenon entirely. This pattern of results suggests several ways in which perception at the word level may differ from perception at the letter level.

https://link.springer.com/content/pdf/10.3758/BF03212406.pdf

Visual factors in word perception

Although Broadbent concedes that we are probably correct in supposing that memory representations are distributed, he argues that psychological evidence is irrelevant to our argument because our point is relevant only at what Marr (1982) has called the implementation^ level of description and that psychological theory is only properly concerned with what Marr calls the computational level. We believe that Broadbent is wrong on both counts. First, our model is stated at a third level between the other two, Marr's representational and algorithmic level. Second, we believe that psychology is properly concerned with all three of these levels and that the information processing approach to psychology has been primarily concerned with the same level that we are, namely, the algorithmic level. Thus, our model is a competitor of the logogen model and other models of human information processing. We discuss these and other aspects of the question of levels, concluding that distributed models may ultimately provide more compelling accounts of a number of aspects of cognitive processes than other, competing algorithmic accounts.

James L. McClelland

Papers

Nonword pronunciation and models of word recognition.

Concepts, control and context: A connectionist account of normal and disordered semantic cognition

Visual factors in word perception

Levels indeed! A response to Broadbent

Constructive memory and memory distortions: A parallel-distributed processing approach.