Warning: Humans cannot reliably detect speech deepfakes.

Speech deepfakes are artificial voices generated by machine learning models. Previous literature has highlighted deepfakes as one of the biggest security threats arising from progress in artificial intelligence due to their potential for misuse. However, studies investigating human detection capabilities are limited. We presented genuine and deepfake audio to n = 529 individuals and asked them to identify the deepfakes. We ran our experiments in English and Mandarin to understand if language affects detection performance and decision-making rationale. We found that detection capability is unreliable. Listeners only correctly spotted the deepfakes 73% of the time, and there was no difference in detectability between the two languages. Increasing listener awareness by providing examples of speech deepfakes only improves results slightly. As speech synthesis algorithms improve and become more realistic, we can expect the detection task to become harder. The difficulty of detecting speech deepfakes confirms their potential for misuse and signals that defenses against this threat are needed.

Deep Learning Applied to Raman Spectroscopy for the Detection of Microsatellite Instability/MMR Deficient Colorectal Cancer.

Defective DNA mismatch repair is one pathogenic pathway to colorectal cancer. It is characterised by microsatellite instability which provides a molecular biomarker for its detection. Clinical guidelines for universal testing of this biomarker are not met due to resource limitations; thus, there is interest in developing novel methods for its detection. Raman spectroscopy (RS) is an analytical tool able to interrogate the molecular vibrations of a sample to provide a unique biochemical fingerprint. The resulting datasets are complex and high-dimensional, making them an ideal candidate for deep learning, though this may be limited by small sample sizes. This study investigates the potential of using RS to distinguish between normal, microsatellite stable (MSS) and microsatellite unstable (MSI-H) adenocarcinoma in human colorectal samples and whether deep learning provides any benefit to this end over traditional machine learning models. A 1D convolutional neural network (CNN) was developed to discriminate between healthy, MSI-H and MSS in human tissue and compared to a principal component analysis-linear discriminant analysis (PCA-LDA) and a support vector machine (SVM) model. A nested cross-validation strategy was used to train 30 samples, 10 from each group, with a total of 1490 Raman spectra. The CNN achieved a sensitivity and specificity of 83% and 45% compared to PCA-LDA, which achieved a sensitivity and specificity of 82% and 51%, respectively. These are competitive with existing guidelines, despite the low sample size, speaking to the molecular discriminative power of RS combined with deep learning. A number of biochemical antecedents responsible for this discrimination are also explored, with Raman peaks associated with nucleic acids and collagen being implicated.

Inferring the location of neurons within an artificial network from their activity.

Inferring the connectivity of biological neural networks from neural activation data is an open problem. We propose that the analogous problem in artificial neural networks is more amenable to study and may illuminate the biological case. Here, we study the specific problem of assigning artificial neurons to locations in a network of known architecture, specifically the LeNet image classifier. We evaluate a supervised learning approach based on features derived from the eigenvectors of the activation correlation matrix. Experiments highlighted that for an image dataset to be effective for accurate localisation, it should fully activate the network and contain minimal confounding correlations. No single image dataset was found that resulted in perfect assignment, however perfect assignment was achieved using a concatenation of features from multiple image datasets.

Machine Learning of Raman Spectroscopy Data for Classifying Cancers: A Review of the Recent Literature.

Raman Spectroscopy has long been anticipated to augment clinical decision making, such as classifying oncological samples. Unfortunately, the complexity of Raman data has thus far inhibited their routine use in clinical settings. Traditional machine learning models have been used to help exploit this information, but recent advances in deep learning have the potential to improve the field. However, there are a number of potential pitfalls with both traditional and deep learning models. We conduct a literature review to ascertain the recent machine learning methods used to classify cancers using Raman spectral data. We find that while deep learning models are popular, and ostensibly outperform traditional learning models, there are many methodological considerations which may be leading to an over-estimation of performance; primarily, small sample sizes which compound sub-optimal choices regarding sampling and validation strategies. Amongst several recommendations is a call to collate large benchmark Raman datasets, similar to those that have helped transform digital pathology, which researchers can use to develop and refine deep learning models.

Fluorescence assay for simultaneous quantification of CFTR ion-channel function and plasma membrane proximity.

The cystic fibrosis transmembrane conductance regulator (CFTR) is a plasma membrane anion channel that plays a key role in controlling transepithelial fluid movement. Excessive activation results in intestinal fluid loss during secretory diarrheas, whereas CFTR mutations underlie cystic fibrosis (CF). Anion permeability depends both on how well CFTR channels work (permeation/gating) and on how many are present at the membrane. Recently, treatments with two drug classes targeting CFTR-one boosting ion-channel function (potentiators) and the other increasing plasma membrane density (correctors)-have provided significant health benefits to CF patients. Here, we present an image-based fluorescence assay that can rapidly and simultaneously estimate both CFTR ion-channel function and the protein's proximity to the membrane. We monitor F508del-CFTR, the most common CF-causing variant, and confirm rescue by low temperature, CFTR-targeting drugs and second-site revertant mutation R1070W. In addition, we characterize a panel of 62 CF-causing mutations. Our measurements correlate well with published data (electrophysiology and biochemistry), further confirming validity of the assay. Finally, we profile effects of acute treatment with approved potentiator drug VX-770 on the rare-mutation panel. Mapping the potentiation profile on CFTR structures raises mechanistic hypotheses on drug action, suggesting that VX-770 might allow an open-channel conformation with an alternative arrangement of domain interfaces. The assay is a valuable tool for investigation of CFTR molecular mechanisms, allowing accurate inferences on gating/permeation. In addition, by providing a two-dimensional characterization of the CFTR protein, it could better inform development of single-drug and precision therapies addressing the root cause of CF disease.

Coherence of achromatic, primary and basic classes of colour categories.

A range of explanations have been advanced for the systems of colour names found in different languages. Some explanations give special, fundamental status to a subset of colour categories. We argue that a subset of colour categories, if fundamental, will be coherent - meaning that a non-trivial criterion distinguishes them from the other colour categories. We test the coherence of subsets of achromatic (white, black and grey), primary (white, black, red, green, yellow, blue) and basic (primaries plus brown, orange, purple, pink and grey) colour categories in English. Criteria for defining colour categories were expressed in terms of behavioural, linguistic and geometric features derived from colour naming and linguistic usage data; and were discovered using machine learning methods. We find that achromatic and basic colour categories are coherent subsets but not primaries. These results support claims that the basic colour categories have special status, and undermine claims about the fundamental role of primaries in colour naming systems.

Reconciling the statistics of spectral reflectance and colour.

The spectral reflectance function of a surface specifies the fraction of the illumination reflected by it at each wavelength. Jointly with the illumination spectral density, this function determines the apparent colour of the surface. Models for the distribution of spectral reflectance functions in the natural environment are considered. The realism of the models is assessed in terms of the individual reflectance functions they generate, and in terms of the overall distribution of colours which they give rise to. Both realism assessments are made in comparison to empirical datasets. Previously described models (PCA- and fourier-based) of reflectance function statistics are evaluated, as are improved versions; and also a novel model, which synthesizes reflectance functions as a sum of sigmoid functions. Key model features for realism are identified. The new sigmoid-sum model is shown to be the most realistic, generating reflectance functions that are hard to distinguish from real ones, and accounting for the majority of colours found in natural images with the exception of an abundance of vegetation green and sky blue.

Limits on transfer learning from photographic image data to X-ray threat detection.

BACKGROUND: X-ray imaging is a crucial and ubiquitous tool for detecting threats to transport security, but interpretation of the images presents a logistical bottleneck. Recent advances in Deep Learning image classification offer hope of improving throughput through automation. However, Deep Learning methods require large quantities of labelled training data. While photographic data is cheap and plentiful, comparable training sets are seldom available for the X-ray domain. OBJECTIVE: To determine whether and to what extent it is feasible to exploit the availability of photo data to supplement the training of X-ray threat detectors. METHODS: A new dataset was collected, consisting of 1901 matched pairs of photo & X-ray images of 501 common objects. Of these, 258 pairs were of 69 objects considered threats in the context of aviation. This data was used to test a variety of transfer learning approaches. A simple model of threat cue availability was developed to understand the limits of this transferability. RESULTS: Appearance features learned from photos provide a useful basis for training classifiers. Some transfer from the photo to the X-ray domain is possible as ∼40% of danger cues are shared between the modalities, but the effectiveness of this transfer is limited since ∼60% of cues are not. CONCLUSIONS: Transfer learning is beneficial when X-ray data is very scarce-of the order of tens of training images in our experiments-but provides no significant benefit when hundreds or thousands of X-ray images are available.

Categorical colour geometry.

Ordinary language users group colours into categories that they refer to by a name e.g. pale green. Data on the colour categories of English speakers was collected using online crowd sourcing - 1,000 subjects produced 20,000 unconstrained names for 600 colour stimuli. From this data, using the framework of Information Geometry, a Riemannian metric was computed throughout the RGB cube. This is the first colour metric to have been computed from colour categorization data. In this categorical metric the distance between two close colours is determined by the difference in the distribution of names that the subject population applied to them. This contrasts with previous colour metrics which have been driven by stimulus discriminability, or acceptability of a colour match. The categorical metric is analysed and shown to be clearly different from discriminability-based metrics. Natural units of categorical length, area and volume are derived. These allow a count to be made of the number of categorically-distinct regions of categorically-similar colours that fit within colour space. Our analysis estimates that 27 such regions fit within the RGB cube, which agrees well with a previous estimate of 30 colours that can be identified by name by untrained subjects.

The Atlas Structure of Images.

Many operations of vision require image regions to be isolated and inter-related. This is challenging when they are different in detail and extent. Practical methods of Computer Vision approach this through the tools of downsampling, pyramids, cropping and patches. In this paper we develop an ideal geometric structure for this, compatible with the existing scale space model of image measurement. Its elements are apertures which view the image like fuzzy-edged portholes of frosted glass. We establish containment and cause/effect relations between apertures, and show that these link them into cross-scale atlases. Atlases formed of Gaussian apertures are shown to be a continuous version of the image pyramid used in Computer Vision, and allow various types of image description to naturally be expressed within their framework. We show that views through Gaussian apertures are approximately equivalent to the jets of derivative of Gaussian filter responses that form part of standard Scale Space theory. This supports a view of the simple cells of mammalian V1 as implementing a system of local views of the retinal image of varying extent and resolution. As a worked example we develop a keypoint descriptor scheme that outperforms previous schemes that do not make use of learning.

A Comparison of Thresholding Methods for Forensic Reconstruction Studies Using Fluorescent Powder Proxies for Trace Materials.

Image segmentation is a fundamental precursor to quantitative image analysis. At present, no standardised methodology exists for segmenting images of fluorescent proxies for trace evidence. Experiments evaluated (i) whether manual segmentation is reproducible within and between examiners (with three participants repeatedly tracing three images) (ii) whether manually defining a threshold level offers accurate and reproducible results (with 20 examiners segmenting 10 images), and (iii) whether a global thresholding algorithm might perform with similar accuracy, while offering improved reproducibility and efficiency (16 algorithms tested). Statistically significant differences were seen between examiners' traced outputs. Manually thresholding produced good accuracy on average (within ±1% of the expected values), but poor reproducibility (with multiple outliers). Three algorithms (Yen, MaxEntropy, and RenyiEntropy) offered similar accuracy, with improved reproducibility and efficiency. Together, these findings suggest that appropriate algorithms could perform thresholding tasks as part of a robust workflow for reconstruction studies employing fluorescent proxies for trace evidence.

A spatial frequency spectral peakedness model predicts discrimination performance of regularity in dot patterns.

Subjective assessments of spatial regularity are common in everyday life and also in science, for example in developmental biology. It has recently been shown that regularity is an adaptable visual dimension. It was proposed that regularity is coded via the peakedness of the distribution of neural responses across receptive field size. Here, we test this proposal for jittered square lattices of dots. We examine whether discriminability correlates with a simple peakedness measure across different presentation conditions (dot number, size, and average spacing). Using a filter-rectify-filter model, we determined responses across scale. Consistently, two peaks are present: a lower frequency peak corresponding to the dot spacing of the regular pattern and a higher frequency peak corresponding to the pattern element (dot). We define the "peakedness" of a particular presentation condition as the relative heights of these two peaks for a perfectly regular pattern constructed using the corresponding dot size, number and spacing. We conducted two psychophysical experiments in which observers judged relative regularity in a 2-alternative forced-choice task. In the first experiment we used a single reference pattern of intermediate regularity and, in the second, Thurstonian scaling of patterns covering the entire range of regularity. In both experiments discriminability was highly correlated with peakedness for a wide range of presentation conditions. This supports the hypothesis that regularity is coded via peakedness of the distribution of responses across scale.

Quantum dot conjugated nanobodies for multiplex imaging of protein dynamics at synapses.

Neurons communicate with each other through synapses, which show enrichment for specialized receptors. Although many studies have explored spatial enrichment and diffusion of these receptors in dissociated neurons using single particle tracking, much less is known about their dynamic properties at synapses in complex tissue like brain slices. Here we report the use of smaller and highly specific quantum dots conjugated with a recombinant single domain antibody fragment (VHH fragment) against green fluorescent protein to provide information on diffusion of adhesion molecules at the growth cone and neurotransmitter receptors at synapses. Our data reveals that QD-nanobodies can measure neurotransmitter receptor dynamics at both excitatory and inhibitory synapses in primary neuronal cultures as well as in ex vivo rat brain slices. We also demonstrate that this approach can be applied to tagging multiple proteins to simultaneously monitor their behavior. Thus, we provide a strategy for multiplex imaging of tagged membrane proteins to study their clustering, diffusion and transport both in vitro as well as in native tissue environments such as brain slices.

Machine Learning Based Localization and Classification with Atomic Magnetometers.

We demonstrate identification of position, material, orientation, and shape of objects imaged by a ^{85}Rb atomic magnetometer performing electromagnetic induction imaging supported by machine learning. Machine learning maximizes the information extracted from the images created by the magnetometer, demonstrating the use of hidden data. Localization 2.6 times better than the spatial resolution of the imaging system and successful classification up to 97% are obtained. This circumvents the need of solving the inverse problem and demonstrates the extension of machine learning to diffusive systems, such as low-frequency electrodynamics in media. Automated collection of task-relevant information from quantum-based electromagnetic imaging will have a relevant impact from biomedicine to security.

Steady-state EB cap size fluctuations are determined by stochastic microtubule growth and maturation.

Growing microtubules are protected from depolymerization by the presence of a GTP or GDP/Pi cap. End-binding proteins of the EB1 family bind to the stabilizing cap, allowing monitoring of its size in real time. The cap size has been shown to correlate with instantaneous microtubule stability. Here we have quantitatively characterized the properties of cap size fluctuations during steady-state growth and have developed a theory predicting their timescale and amplitude from the kinetics of microtubule growth and cap maturation. In contrast to growth speed fluctuations, cap size fluctuations show a characteristic timescale, which is defined by the lifetime of the cap sites. Growth fluctuations affect the amplitude of cap size fluctuations; however, cap size does not affect growth speed, indicating that microtubules are far from instability during most of their time of growth. Our theory provides the basis for a quantitative understanding of microtubule stability fluctuations during steady-state growth.

Detection of concealed cars in complex cargo X-ray imagery using Deep Learning.

BACKGROUND: Non-intrusive inspection systems based on X-ray radiography techniques are routinely used at transport hubs to ensure the conformity of cargo content with the supplied shipping manifest. As trade volumes increase and regulations become more stringent, manual inspection by trained operators is less and less viable due to low throughput. Machine vision techniques can assist operators in their task by automating parts of the inspection workflow. Since cars are routinely involved in trafficking, export fraud, and tax evasion schemes, they represent an attractive target for automated detection and flagging for subsequent inspection by operators. OBJECTIVE: Development and evaluation of a novel method for the automated detection of cars in complex X-ray cargo imagery. METHODS: X-ray cargo images from a stream-of-commerce dataset were classified using a window-based scheme. The limited number of car images was addressed by using an oversampling scheme. Different Convolutional Neural Network (CNN) architectures were compared with well-established bag of words approaches. In addition, robustness to concealment was evaluated by projection of objects into car images. RESULTS: CNN approaches outperformed all other methods evaluated, achieving 100% car image classification rate for a false positive rate of 1-in-454. Cars that were partially or completely obscured by other goods, a modus operandi frequently adopted by criminals, were correctly detected. CONCLUSIONS: We believe that this level of performance suggests that the method is suitable for deployment in the field. It is expected that the generic object detection workflow described can be extended to other object classes given the availability of suitable training data.

Automated X-ray image analysis for cargo security: Critical review and future promise.

We review the relatively immature field of automated image analysis for X-ray cargo imagery. There is increasing demand for automated analysis methods that can assist in the inspection and selection of containers, due to the ever-growing volumes of traded cargo and the increasing concerns that customs- and security-related threats are being smuggled across borders by organised crime and terrorist networks. We split the field into the classical pipeline of image preprocessing and image understanding. Preprocessing includes: image manipulation; quality improvement; Threat Image Projection (TIP); and material discrimination and segmentation. Image understanding includes: Automated Threat Detection (ATD); and Automated Contents Verification (ACV). We identify several gaps in the literature that need to be addressed and propose ideas for future research. Where the current literature is sparse we borrow from the single-view, multi-view, and CT X-ray baggage domains, which have some characteristics in common with X-ray cargo.

Measuring and correcting wobble in large-scale transmission radiography.

BACKGROUND: Large-scale transmission radiography scanners are used to image vehicles and cargo containers. Acquired images are inspected for threats by a human operator or a computer algorithm. To make accurate detections, it is important that image values are precise. However, due to the scale (∼5 m tall) of such systems, they can be mechanically unstable, causing the imaging array to wobble during a scan. This leads to an effective loss of precision in the captured image. OBJECTIVE: We consider the measurement of wobble and amelioration of the consequent loss of image precision. METHODS: Following our previous work, we use Beam Position Detectors (BPDs) to measure the cross-sectional profile of the X-ray beam, allowing for estimation, and thus correction, of wobble. We propose: (i) a model of image formation with a wobbling detector array; (ii) a method of wobble correction derived from this model; (iii) methods for calibrating sensor sensitivities and relative offsets; (iv) a Random Regression Forest based method for instantaneous estimation of detector wobble; and (v) using these estimates to apply corrections to captured images of difficult scenes. RESULTS: We show that these methods are able to correct for 87% of image error due wobble, and when applied to difficult images, a significant visible improvement in the intensity-windowed image quality is observed. CONCLUSIONS: The method improves the precision of wobble affected images, which should help improve detection of threats and the identification of different materials in the image.

Difference magnitude is not measured by discrimination steps for order of point patterns.

We have shown in previous work that the perception of order in point patterns is consistent with an interval scale structure (Protonotarios, Baum, Johnston, Hunter, & Griffin, 2014). The psychophysical scaling method used relies on the confusion between stimuli with similar levels of order, and the resulting discrimination scale is expressed in just-noticeable differences (jnds). As with other perceptual dimensions, an interesting question is whether suprathreshold (perceptual) differences are consistent with distances between stimuli on the discrimination scale. To test that, we collected discrimination data, and data based on comparison of perceptual differences. The stimuli were jittered square lattices of dots, covering the range from total disorder (Poisson) to perfect order (square lattice), roughly equally spaced on the discrimination scale. Observers picked the most ordered pattern from a pair, and the pair of patterns with the greatest difference in order from two pairs. Although the judgments of perceptual difference were found to be consistent with an interval scale, like the discrimination judgments, no common interval scale that could predict both sets of data was possible. In particular, the midpattern of the perceptual scale is 11 jnds away from the ordered end, and 5 jnds from the disordered end of the discrimination scale.

Neuronal activity mediated regulation of glutamate transporter GLT-1 surface diffusion in rat astrocytes in dissociated and slice cultures.

The astrocytic GLT-1 (or EAAT2) is the major glutamate transporter for clearing synaptic glutamate. While the diffusion dynamics of neurotransmitter receptors at the neuronal surface are well understood, far less is known regarding the surface trafficking of transporters in subcellular domains of the astrocyte membrane. Here, we have used live-cell imaging to study the mechanisms regulating GLT-1 surface diffusion in astrocytes in dissociated and brain slice cultures. Using GFP-time lapse imaging, we show that GLT-1 forms stable clusters that are dispersed rapidly and reversibly upon glutamate treatment in a transporter activity-dependent manner. Fluorescence recovery after photobleaching and single particle tracking using quantum dots revealed that clustered GLT-1 is more stable than diffuse GLT-1 and that glutamate increases GLT-1 surface diffusion in the astrocyte membrane. Interestingly, the two main GLT-1 isoforms expressed in the brain, GLT-1a and GLT-1b, are both found to be stabilized opposed to synapses under basal conditions, with GLT-1b more so. GLT-1 surface mobility is increased in proximity to activated synapses and alterations of neuronal activity can bidirectionally modulate the dynamics of both GLT-1 isoforms. Altogether, these data reveal that astrocytic GLT-1 surface mobility, via its transport activity, is modulated during neuronal firing, which may be a key process for shaping glutamate clearance and glutamatergic synaptic transmission. GLIA 2016;64:1252-1264.