I address whether neural networks perform computations in the sense of computability theory and computer science. I explicate and defend
the following theses. (1) Many neural networks compute—they perform computations. (2) Some neural networks compute in a classical way.
Ordinary digital computers, which are very large networks of logic gates, belong in this class of neural networks. (3) Other neural networks
compute in a non-classical way. (4) Yet other neural networks (...) do not perform computations. Brains may well fall into this last class. (shrink)

This work describes a mathematical concept of modelling field theory and its applications to a variety of problems, while offering a view of the relationships among mathematics, computational concepts in neural networks, semiotics, and concepts of mind in psychology and philosophy.

Artificial neural networks (ANNs) are new mathematical techniques which can be used for modelling real neural networks, but also for data categorisation and inference tasks in any empirical science. This means that they have a twofold interest for the philosopher. First, ANN theory could help us to understand the nature of mental phenomena such as perceiving, thinking, remembering, inferring, knowing, wanting and acting. Second, because ANNs are such powerful instruments for data classification and inference, their use (...) also leads us into the problems of induction and probability. Ever since David Hume expressed his famous doubts about induction, the principles of scientific inference have been a central concern for philosophers. (shrink)

In this paper an Artificial Neural Network (ANN) model was used to help cars dealers recognize the many characteristics of cars, including manufacturers, their location and classification of cars according to several categories including: Make, Model, Type, Origin, DriveTrain, MSRP, Invoice, EngineSize, Cylinders, Horsepower, MPG_Highway, Weight, Wheelbase, Length. ANN was used in prediction of the number of miles per gallon when the car is driven in the city(MPG_City). The results showed that ANN model was able to predict MPG_City with (...) 97.50 % accuracy. The factor of DriveTrain has the most influence on MPG_City evaluation. Similar studies can be carried out for the evaluation of other characteristics of cars. (shrink)

Neural networks have long been used to study linguistic phenomena spanning the domains of phonology, morphology, syntax, and semantics. Of these domains, semantics is somewhat unique in that there is little clarity concerning what a model needs to be able to do in order to provide an account of how the meanings of complex linguistic expressions, such as sentences, are understood. We argue that one thing such models need to be able to do is generate predictions about which (...) further sentences are likely to follow from a given sentence; these define the sentence's “inferential role.” We then show that it is possible to train a tree-structured neural network model to generate very simple examples of such inferential roles using the recently released Stanford Natural Language Inference (SNLI) dataset. On an empirical front, we evaluate the performance of this model by reporting entailment prediction accuracies on a set of test sentences not present in the training data. We also report the results of a simple study that compares human plausibility ratings for both human-generated and model-generated entailments for a random selection of sentences in this test set. On a more theoretical front, we argue in favor of a revision to some common assumptions about semantics: understanding a linguistic expression is not only a matter of mapping it onto a representation that somehow constitutes its meaning; rather, understanding a linguistic expression is mainly a matter of being able to draw certain inferences. Inference should accordingly be at the core of any model of semantic cognition. (shrink)

Deep neural networks (DNNs) have had extraordinary successes in classifying photographic images of objects and are often described as the best models of biological vision. This conclusion is largely based on three sets of findings: (1) DNNs are more accurate than any other model in classifying images taken from various datasets, (2) DNNs do the best job in predicting the pattern of human errors in classifying objects taken from various behavioral datasets, and (3) DNNs do the best job (...) in predicting brain signals in response to images taken from various brain datasets (e.g., single cell responses or fMRI data). However, these behavioral and brain datasets do not test hypotheses regarding what features are contributing to good predictions and we show that the predictions may be mediated by DNNs that share little overlap with biological vision. More problematically, we show that DNNs account for almost no results from psychological research. This contradicts the common claim that DNNs are good, let alone the best, models of human object recognition. We argue that theorists interested in developing biologically plausible models of human vision need to direct their attention to explaining psychological findings. More generally, theorists need to build models that explain the results of experiments that manipulate independent variables designed to test hypotheses rather than compete on making the best predictions. We conclude by briefly summarizing various promising modeling approaches that focus on psychological data. (shrink)

Human decision making shows systematic simplifications and deviations from the tenets of rationality (‘heuristics’) that may lead to suboptimal decisional outcomes (‘cognitive biases’). There are currently three prevailing theoretical perspectives on the origin of heuristics and cognitive biases: a cognitive-psychological, an ecological and an evolutionary perspective. However, these perspectives are mainly descriptive and none of them provides an overall explanatory framework for the underlying mechanisms of cognitive biases. To enhance our understanding of cognitive heuristics and biases we propose a (...) class='Hi'>neural network framework for cognitive biases, which explains why our brain systematically tends to default to heuristic (‘Type 1’) decision making. We argue that many cognitive biases arise from intrinsic brain mechanisms that are fundamental for the working of biological neural networks. In order to substantiate our viewpoint, we discern and explain four basic neural network principles: (1) Association, (2) Compatibility (3) Retainment, and (4) Focus. These principles are inherent to (all) neural networks which were originally optimized to perform concrete biological, perceptual, and motor functions. They form the basis for our inclinations to associate and combine (unrelated) information, to prioritize information that is compatible with our present state (such as knowledge, opinions and expectations), to retain given information that sometimes could better be ignored, and to focus on dominant information while ignoring relevant information that is not directly activated. The supposed mechanisms are complementary and not mutually exclusive. For different cognitive biases they may all contribute in varying degrees to distortion of information. The present viewpoint not only complements the earlier three viewpoints, but also provides a unifying and binding framework for many cognitive bias phenomena. (shrink)

A central methodological innovation in cognitive science has been the development of connectionist or neural network models of psychological phenomena. These models may also comprise a theoretically integrative and methodologically rigorous approach to psychiatric phenomena. In this paper we employ connectionist theory to conceptualize obsessive-compulsive disorder . We discuss salient phenomenological and neurobiological findings of the illness, and then reformulate these using neural network models. Several features and mechanisms of OCD may be explicated in terms of disordered (...) class='Hi'>networks. Neural network modeling appears to constitute a novel and potentially fertile approach to psychiatric disorders such as OCD. (shrink)

Deep neural networks (DNNs) have found many useful applications in recent years. Of particular interest have been those instances where their successes imitate human cognition and many consider artificial intelligences to offer a lens for understanding human intelligence. Here, we criticize the underlying conflation between the predictive and explanatory power of DNNs by examining the goals of modeling.

The literature on common pool resource (CPR) governance lists numerous factors that influence whether a given CPR system achieves ecological long-term sustainability. Up to now there is no comprehensive model to integrate these factors or to explain success within or across cases and sectors. Difficulties include the absence of large-N-studies (Poteete 2008), the incomparability of single case studies, and the interdependence of factors (Agrawal and Chhatre 2006). We propose (1) a synthesis of 24 success factors based on the current SES (...) framework and a literature review; (2) the application of neural networks on a database of CPR management case studies in an attempt to test the viability of this synthesis. This method allows us to obtain an implicit quantitative and rather precise model of the interdependencies in CPR systems. Given such a model, every success factor in each case can be manipulated separately, yielding different predictions for success. This could become be a fast and inexpensive way to analyze, predict and optimize performance for communities world-wide facing CPR challenges. Existing theoretical frameworks could be improved as well. (shrink)

An adaptive neural network control problem is addressed for a class of nonlinear hydraulic servo-systems with time-varying state constraints. In view of the low precision problem of the traditional hydraulic servo-system which is caused by the tracking errors surpassing appropriate bound, the previous works have shown that the constraint for the system is a good way to solve the low precision problem. Meanwhile, compared with constant constraints, the time-varying state constraints are more general in the actual systems. Therefore, when (...) the states of the system are forced to obey bounded time-varying constraint conditions, the high precision tracking performance of the system can be easily realized. In order to achieve this goal, the time-varying barrier Lyapunov function is used to prevent the states from violating time-varying constraints. By the backstepping design, the adaptive controller will be obtained. A radial basis function neural network is used to estimate the uncertainties. Based on analyzing the stability of the hydraulic servo-system, we show that the error signals are bounded in the compacts sets; the time-varying state constrains are never violated and all singles of the hydraulic servo-system are bounded. The simulation and experimental results show that the tracking accuracy of system is improved and the controller has fast tracking ability and strong robustness. (shrink)

If artificial neural networks are ever to form the foundation for higher level cognitive behaviors in machines or to realize their full potential as explanatory devices for human cognition, they must show signs of autonomy, multifunction operation, and intersystem integration that are absent in most existing models. This model begins to address these issues by integrating predictive learning, sequence interleaving, and sequence creation components to simulate a spectrum of higher-order cognitive behaviors which have eluded the grasp of simpler (...) systems. Its capabilities are described based on simulations calling for increasing levels of functionality and are used to show how the model can progress from fundamental sequence learning and recall tasks to sophisticated behaviors such as an ability to solve simple mathematical expressions and a creative capacity for the formation and application of inductive rules. (shrink)

In this paper an Artificial Neural Network (ANN) model was used to help cars dealers recognize the many characteristics of cars, including manufacturers, their location and classification of cars according to several categories including: Buying, Maint, Doors, Persons, Lug_boot, Safety, and Overall. ANN was used in forecasting car acceptability. The results showed that ANN model was able to predict the car acceptability with 99.12 %. The factor of Safety has the most influence on car acceptability evaluation. Comparative study method (...) is suitable for the evaluation of car acceptability forecasting, can also be extended to all other areas. (shrink)

The paper aims to discuss the results of testing a neural network which classifies the vowels of the vocalic system with the [ATR] (Advanced Tongue Root) contrast based on the data of Akebu (Kwa family). The acoustic nature of the [ATR] feature is yet understudied. The only reliable acoustic correlate of [ATR] is the magnitude of the first formant (F1) which can be also modulated by tongue height, resulting in significant overlap between high [-ATR] vowels and mid [+ATR] vowels. (...) Other acoustic metrics which had been associated with the [ATR], such as F1 bandwidth (B1), relative intensity of F1 to F2 (A1-A2), etc., are typically inconsistent across vowel types and speakers. The values of four metrics – F1, F2, A1-A2, B1 – were used for training and testing the neural network. We tested four versions of the model differing in the presence of the fifth variable encoding the speaker and the number of hidden layers. The models which included the variable encoding the speaker achieved slightly higher accuracy, whereas the precision and recall metrics of the three-layer model were generally higher than those with two hidden layers. (shrink)

Diabetes is one of the most common diseases worldwide where a cure is not found for it yet. Annually it cost a lot of money to care for people with diabetes. Thus the most important issue is the prediction to be very accurate and to use a reliable method for that. One of these methods is using artificial intelligence systems and in particular is the use of Artificial Neural Networks (ANN). So in this paper, we used artificial (...) class='Hi'>neural networks to predict whether a person is diabetic or not. The criterion was to minimize the error function in neural network training using a neural network model. After training the ANN model, the average error function of the neural network was equal to 0.01 and the accuracy of the prediction of whether a person is diabetics or not was 87.3%. (shrink)

Computer simulations show that an unstructured neural-network model [Shultz, T. R., & Bale, A. C. (2001). Infancy, 2, 501–536] covers the essential features␣of infant learning of simple grammars in an artificial language [Marcus, G. F., Vijayan, S., Bandi Rao, S., & Vishton, P. M. (1999). Science, 283, 77–80], and generalizes to examples both outside and inside of the range of training sentences. Knowledge-representation analyses confirm that these networks discover that duplicate words in the sentences are nearly identical and (...) that they use this near-identity relation to distinguish sentences that are consistent or inconsistent with a familiar grammar. Recent simulations that were claimed to show that this model did not really learn these grammars [Vilcu, M., & Hadley, R. F. (2005). Minds and Machines, 15, 359–382] confounded syntactic types with speech sounds and did not perform standard statistical tests of results. (shrink)

As a type of evidence glass can be very useful contact trace material in a wide range of offences including burglaries and robberies, hit-and-run accidents, murders, assaults, ram-raids, criminal damage and thefts of and from motor vehicles. All of that offer the potential for glass fragments to be transferred from anything made of glass which breaks, to whoever or whatever was responsible. Variation in manufacture of glass allows considerable discrimination even with tiny fragments. In this study, we worked glass classification (...) and testing of artificial neural network model created by the JustNN. The aim of the study is help investigator in identifying the type of glass found in arena of the crime. The Neural Network model was trained and validated using the type of glass dataset. The accuracy of model in predicting the type of glass reached 96.7%. Thus neural network is suitable for predicating type of glasses. (shrink)

Background: Although convolutional neural networks (CNNs) achieve high diagnostic accuracy for detecting Alzheimer’s disease (AD) dementia based on magnetic resonance imaging (MRI) scans, they are not yet applied in clinical routine. One important reason for this is a lack of model comprehensibility. Recently developed visualization methods for deriving CNN relevance maps may help to fill this gap as they allow the visualization of key input image features that drive the decision of the model. We investigated whether models with (...) higher accuracy also rely more on discriminative brain regions predefined by prior knowledge. Methods: We trained a CNN for the detection of AD in N = 663 T1-weighted MRI scans of patients with dementia and amnestic mild cognitive impairment (MCI) and verified the accuracy of the models via cross-validation and in three independent samples including in total N = 1655 cases. We evaluated the association of relevance scores and hippocampus volume to validate the clinical utility of this approach. To improve model comprehensibility, we implemented an interactive visualization of 3D CNN relevance maps, thereby allowing intuitive model inspection. Results: Across the three independent datasets, group separation showed high accuracy for AD dementia versus controls (AUC ≥ 0.91) and moderate accuracy for amnestic MCI versus controls (AUC ≈ 0.74). Relevance maps indicated that hippocampal atrophy was considered the most informative factor for AD detection, with additional contributions from atrophy in other cortical and subcortical regions. Relevance scores within the hippocampus were highly correlated with hippocampal volumes (Pearson’s r ≈ −0.86, p < 0.001). Conclusion: The relevance maps highlighted atrophy in regions that we had hypothesized a priori. This strengthens the comprehensibility of the CNN models, which were trained in a purely data-driven manner based on the scans and diagnosis labels. The high hippocampus relevance scores as well as the high performance achieved in independent samples support the validity of the CNN models in the detection of AD-related MRI abnormalities. The presented data-driven and hypothesis-free CNN modeling approach might provide a useful tool to automatically derive discriminative features for complex diagnostic tasks where clear clinical criteria are still missing, for instance for the differential diagnosis between various types of dementia. (shrink)

The ontology of social objects and facts remains a field of continued controversy. This situation complicates the life of social scientists who seek to make predictive models of social phenomena. For the purposes of modelling a social phenomenon, we would like to avoid having to make any controversial ontological commitments. The overwhelming majority of models in the social sciences, including statistical models, are built upon ontological assumptions that can be questioned. Recently, however, artificial neural networks have made their (...) way into the social sciences, raising the question whether they can avoid controversial ontological assumptions. ANNs are largely distinguished from other statistical and machine learning techniques by being a representation-learning technique. That is, researchers can let the neural networks select which features of the data to use for internal representation instead of imposing their preconceptions. On this basis, I argue that neural networks can avoid ontological assumptions to a greater degree than common statistical models in the social sciences. I then go on, however, to establish that ANNs are not ontologically innocent either. The use of ANNs in the social sciences introduces ontological assumptions typically in at least two ways, via the input and via the architecture. (shrink)

Engineers use neural networks to control systems too complex for conventional engineering solutions. To examine the behavior of individual hidden units would defeat the purpose of this approach because it would be largely uninterpretable. Yet neurophysiologists spend their careers doing just that! Hidden units contain bits and scraps of signals that yield only arcane hints about network function and no information about how its individual units process signals. Most literature on single-unit recordings attests to this grim fact. On (...) the other hand, knowing a system's function and describing it with elegant mathematics tell one very little about what to expect of interneuronal behavior. Examples of simple networks based on neurophysiology are taken from the oculomotor literature to suggest how single-unit interpretability might decrease with increasing task complexity. It is argued that trying to explain how any real neural network works on a cell-by-cell, reductionist basis is futile and we may have to be content with trying to understand the brain at higher levels of organization. (shrink)

How might artificial neural networks (ANNs) inform cognitive science? Often cognitive scientists use ANNs but do not examine their internal structures. In this paper, we use ANNs to explore how cognition might represent musical properties. We train ANNs to classify musical chords, and we interpret network structure to determine what representations ANNs discover and use. We find connection weights between input units and hidden units can be described using Fourier phase spaces, a representation studied in musical set theory. (...) We find the total signal coming through these weighted connection weights is a measure of the similarity between two Fourier structures: the structure of the hidden unit's weights and the structure of the stimulus. This is surprising because neither of these Fourier structures is computed by the hidden unit. We then show how output units use such similarity measures to classify chords. However, we also find different types of units—units that use different activation functions—use this similarity measure very differently. This result, combined with other findings, indicates that while our networks are related to the Fourier analysis of musical sets, they do not perform Fourier analyses of the kind usually described in musical set theory. Our results show Fourier representations of music are not limited to musical set theory. Our results also suggest how cognitive psychologists might explore Fourier representations in musical cognition. Critically, such theoretical and empirical implications require researchers to understand how network structure converts stimuli into responses. (shrink)

This article deals with detection and accommodation of sensor faults in longitudinal dynamics of an F8 aircraft model. Both the detection of the fault and reconfiguration of the failed sensor are done with the help of neural network-based models. Detection of a sensor fault is done with the help of knowledge-based neural network fault detection. Apart from KBNNFD, another neural network model is developed in this article for the reconfiguration of the failed sensor. A model-based approach of (...) the neural network is developed, which uses the radial basis function of the neural network. MBNN successfully does the task of providing analytical redundancy for the aircraft sensor. In this work, both detection and reconfiguration of a fault is done using neural networks. Hence, the control system becomes robust for handling sensor failures near steady state and reconfiguration is also faster. A generalized regression neural network, which is a type of radial basis network, is used for MBNN, which gives very efficient results for function approximation. An F8 aircraft model and C-Star controller, which improves its handling quality, are used for validation of the method involved. Models of F8 aircraft, C-Star controller, KBNNFD, and MBNN were developed using MATLAB/simulink. Successful implementation and simulation results are shown and discussed using Simulink. (shrink)

This paper proposes an adaptive neural network control approach for a direct-current system with full state constraints. To guarantee that state constraints always remain in the asymmetric time-varying constraint regions, the asymmetric time-varying Barrier Lyapunov Function is employed to structure an adaptive NN controller. As we all know that the constant constraint is only a special case of the time-varying constraint, hence, the proposed control method is more general for dealing with constraint problem as compared with the existing works (...) on DC systems. As far as we know, this system is the first studied situations with time-varying constraints. Using Lyapunov analysis, all signals in the closed-loop system are proved to be bounded and the constraints are not violated. In this paper, the effectiveness of the control method is demonstrated by simulation results. (shrink)

An ideal vision model accounts for behavior and neurophysiology in both naturalistic conditions and designed lab experiments. Unlike psychological theories, artificial neural networks (ANNs) actually perform visual tasks and generate testable predictions for arbitrary inputs. These advantages enable ANNs to engage the entire spectrum of the evidence. Failures of particular models drive progress in a vibrant ANN research program of human vision.

This paper proposes several recurrent neural network-based models for recognizing requisite and effectuation parts in Legal Texts. Firstly, we propose a modification of BiLSTM-CRF model that allows the use of external features to improve the performance of deep learning models in case large annotated corpora are not available. However, this model can only recognize RE parts which are not overlapped. Secondly, we propose two approaches for recognizing overlapping RE parts including the cascading approach which uses the sequence of BiLSTM-CRF (...) models and the unified model approach with the multilayer BiLSTM-CRF model and the multilayer BiLSTM-MLP-CRF model. Experimental results on two Japan law RRE datasets demonstrated advantages of our proposed models. For the Japanese National Pension Law dataset, our approaches obtained an \ score of 93.27% and exhibited a significant improvement compared to previous approaches. For the Japan Civil Code RRE dataset which is written in English, our approaches produced an \ score of 78.24% in recognizing RE parts that exhibited a significant improvement over strong baselines. In addition, using external features and in-domain pre-trained word embeddings also improved the performance of RRE systems. (shrink)

In their account of learning and behavior, the authors define an interactor as emitted behavior that operates on the environment, which excludes Pavlovian learning. A unified neural-network account of the operant-Pavlovian dichotomy favors interpreting neurons as interactors and synaptic efficacies as replicators. The latter interpretation implies that single-synapse change is inherently Lamarckian.

The subject matter of this book is the mental lexicon, that is, the way in which the form and meaning of words is stored by speakers of specific languages. This book attempts to narrow the gap between the results of experimental neurology and the concerns of theoretical linguistics in the area of lexical semantics. The prime goal as regards linguistic theory is to show how matters of lexical organization can be analysed and discussed within a neurologically informed framework that is (...) both adaptable and constrained. It combines the perspectives of distributed network modelling and linguistic semantics, and draws upon the accruing evidence from neuroimaging studies as regards the cortical regions involved. It engages with a number of controversial current issues in both disciplines. The book is intended as a tool for linguists interested in psychological adequacy and the latest advances in cognitive science. It provides a principled means of distinguishing those semantic features required by a mental lexicon that have a direct bearing on grammar from those that do not. It will appeal to researchers in neurolinguistics and lexical semantics."--p. 4 of cover. (shrink)

The ability to conduct logical reasoning is a fundamental aspect of intelligent behavior, and thus an important problem along the way to human-level artificial intelligence. Traditionally, symbolic methods from the field of knowledge representation and reasoning have been used to equip agents with capabilities that resemble human reasoning qualities. More recently, however, there has been an increasing interest in applying alternative approaches based on machine learning rather than logic-based formalisms to tackle this kind of tasks. Here, we make use of (...) state-of-the-art methods for training deep neural networks to devise a novel model that is closely coupled to symbolic reasoning methods, and thus able to learn how to effectively perform basic ontology reasoning. This term describes an important and at the same time very natural kind of problem settings where the rules for conducting reasoning are specified alongside with the actual information. Many problems in practice may be viewed as such reasoning tasks, which is why the presented approach is applicable to a plethora of important real-world problems. To demonstrate the effectiveness of the suggested method, we present the outcomes of several experiments that have been conducted on both toy datasets as well as real-world data, which show that our model learned to perform precise reasoning on a number of diverse inference tasks that require comprehensive deductive proficiencies. Furthermore, it turned out that the suggested model suffers much less from different obstacles that prohibit symbolic reasoning. (shrink)

The missing ingredients in efforts to develop neural networks and artificial intelligence (AI) that can emulate human intelligence have been the evolutionary processes of performing tasks at increased orders of hierarchical complexity. Stacked neural networks based on the Model of Hierarchical Complexity could emulate evolution's actual learning processes and behavioral reinforcement. Theoretically, this should result in stability and reduce certain programming demands. The eventual success of such methods begs questions of humans' survival in the face of (...) androids of superior intelligence and physical composition. These raise future moral questions worthy of speculation. (shrink)

Bowers et al. propose to use controlled behavioral experiments when evaluating deep neural networks as models of biological vision. We agree with the sentiment and draw parallels to the notion that “neuroscience needs behavior.” As a promising path forward, we suggest complementing image recognition tasks with increasingly realistic and well-controlled task environments that engage real-world object recognition behavior.

Reviews the contribution of neural network models in psychiatry and psychopathology, including diagnosis, pharmacotherapy and psychotherapy.

This paper focuses on modeling and control of a class of serial variable stiffness actuators based on level mechanisms for robotic applications. A multi-input multi-output complex nonlinear dynamic model is derived to fully describe SVSAs and the relative degree of the model is determined accordingly. Due to nonlinearity, high coupling, and parametric uncertainty of SVSAs, a neural network-based adaptive control strategy based on feedback linearization is proposed to handle system uncertainties. The feasibility of the proposed approach for position and (...) stiffness tracking of SVSAs is verified by simulation results. (shrink)

Cultural similarities and differences in facial expressions have been a controversial issue in the field of facial communications. A key step in addressing the debate regarding the cultural dependency of emotional expression is to characterize the visual features of specific facial expressions in individual cultures. Here we developed an image analysis framework for this purpose using convolutional neural networks that through training learned visual features critical for classification. We analyzed photographs of facial expressions derived from two databases, each (...) developed in a different country, in which corresponding emotion labels were available. While the CNNs reached high rates of correct results that were far above chance after training with each database, they showed many misclassifications when they analyzed faces from the database that was not used for training. These results suggest that facial features useful for classifying facial expressions differed between the databases. The selectivity of computational units in the CNNs to action units of the face varied across the facial expressions. Importantly, the AU selectivity often differed drastically between the CNNs trained with the different databases. Similarity and dissimilarity of these tuning profiles partly explained the pattern of misclassifications, suggesting that the AUs are important for characterizing the facial features and differ between the two countries. The AU tuning profiles, especially those reduced by principal component analysis, are compact summaries useful for comparisons across different databases, and thus might advance our understanding of universality vs. specificity of facial expressions across cultures. (shrink)

The paper deals with the quandary of the neutrality and transparency of technologies. First, I show how this problem is connected with the image of the opening of 'black boxes' that is pivotal to much of science and technology studies. Second, methodological and socio-political dimensions of the 'black box' metaphor are discussed. Third, I analyze three typical solutions to the problem of the neutrality of technologies outside and inside constructivist technology studies. It is demonstrated that despite their apparent differences, these (...) solutions are similar in their logic of conceptualizing technology as a neutral intermediary. Forth, I look for an alternative to this logic in the actor-network theory of Bruno Latour. Here technologies are conceived in terms of an eventful association of heterogeneous entities irreducible to its conditions of possibility. The construction of technologies is understood as mediation, or as a 'making-do' process where creators are surprised by their creations and vice versa. In Latour's actor-network, technologies are interpreted as opaque and non-neutral entities. Finally, I turn to some object-lessons from smart technologies powered by neural networks to demonstrate that these are empirical vindications of Latour's conception of technical mediation. Particular attention is paid to the opacity and (non)interpretability of machine learning algorithms. (shrink)