Jan Leike (born 1986 or 1987) is an AI alignment researcher who has worked at DeepMind and OpenAI. He joined Anthropic in May 2024. == Education == Jan Leike obtained his undergraduate degree from the University of Freiburg in Germany. After earning a master's degree in computer science, he pursued a PhD in machine learning at the Australian National University under the supervision of Marcus Hutter. == Career == Leike made a six-month postdoctoral fellowship at the Future of Humanity Institute before joining DeepMind to focus on empirical AI safety research, where he collaborated with Shane Legg. === OpenAI === In 2021, Leike joined OpenAI. In June 2023, he and Ilya Sutskever became the co-leaders of the newly introduced "superalignment" project, which aimed to determine how to align future artificial superintelligences within four years to ensure their safety. This project involved automating AI alignment research using relatively advanced AI systems. At the time, Sutskever was OpenAI's Chief Scientist, and Leike was the Head of Alignment. Leike was featured in Time's list of the 100 most influential personalities in AI, both in 2023 and in 2024. In May 2024, Leike announced his resignation from OpenAI, following the departure of Sutskever, Daniel Kokotajlo and several other AI safety employees from the company. Leike wrote that "Over the past years, safety culture and processes have taken a backseat to shiny products", and that he "gradually lost trust" in OpenAI's leadership. In May 2024, Leike joined Anthropic, an AI company founded by former OpenAI employees.
Labeled data
Labeled data is a group of samples that have been tagged with one or more labels. Labeling typically takes a set of unlabeled data and augments each piece of it with informative tags called judgments. For example, a data label might indicate whether a photo contains a horse or a cow, which words were uttered in an audio recording, what type of action is being performed in a video, what the topic of a news article is, what the overall sentiment of a tweet is, or whether a dot in an X-ray is a tumor. Labels can be obtained by having humans make judgments about a given piece of unlabeled data. Labeled data is significantly more expensive to obtain than the raw unlabeled data. The quality of labeled data directly influences the performance of supervised machine learning models in operation, as these models learn from the provided labels. == Crowdsourced labeled data == In 2006, Fei-Fei Li, the co-director of the Stanford Human-Centered AI Institute, initiated research to improve the artificial intelligence models and algorithms for image recognition by significantly enlarging the training data. The researchers downloaded millions of images from the World Wide Web and a team of undergraduates started to apply labels for objects to each image. In 2007, Li outsourced the data labeling work on Amazon Mechanical Turk, an online marketplace for digital piece work. The 3.2 million images that were labeled by more than 49,000 workers formed the basis for ImageNet, one of the largest hand-labeled database for outline of object recognition. == Automated data labelling == After obtaining a labeled dataset, machine learning models can be applied to the data so that new unlabeled data can be presented to the model and a likely label can be guessed or predicted for that piece of unlabeled data. == Challenges == === Data-driven bias === Algorithmic decision-making is subject to programmer-driven bias as well as data-driven bias. Training data that relies on bias labeled data will result in prejudices and omissions in a predictive model, despite the machine learning algorithm being legitimate. The labeled data used to train a specific machine learning algorithm needs to be a statistically representative sample to not bias the results. For example, in facial recognition systems underrepresented groups are subsequently often misclassified if the labeled data available to train has not been representative of the population,. In 2018, a study by Joy Buolamwini and Timnit Gebru demonstrated that two facial analysis datasets that have been used to train facial recognition algorithms, IJB-A and Adience, are composed of 79.6% and 86.2% lighter skinned humans respectively. === Human error and inconsistency === Human annotators are prone to errors and biases when labeling data. This can lead to inconsistent labels and affect the quality of the data set. The inconsistency can affect the machine learning model's ability to generalize well. === Domain expertise === Certain fields, such as legal document analysis or medical imaging, require annotators with specialized domain knowledge. Without the expertise, the annotations or labeled data may be inaccurate, negatively impacting the machine learning model's performance in a real-world scenario.
Visual temporal attention
Visual temporal attention is a special case of visual attention that involves directing attention to specific instant of time. Similar to its spatial counterpart visual spatial attention, these attention modules have been widely implemented in video analytics in computer vision to provide enhanced performance and human interpretable explanation of deep learning models. As visual spatial attention mechanism allows human and/or computer vision systems to focus more on semantically more substantial regions in space, visual temporal attention modules enable machine learning algorithms to emphasize more on critical video frames in video analytics tasks, such as human action recognition. In convolutional neural network-based systems, the prioritization introduced by the attention mechanism is regularly implemented as a linear weighting layer with parameters determined by labeled training data. == Application in Action Recognition == Recent video segmentation algorithms often exploits both spatial and temporal attention mechanisms. Research in human action recognition has accelerated significantly since the introduction of powerful tools such as Convolutional Neural Networks (CNNs). However, effective methods for incorporation of temporal information into CNNs are still being actively explored. Motivated by the popular recurrent attention models in natural language processing, the Attention-aware Temporal Weighted CNN (ATW CNN) is proposed in videos, which embeds a visual attention model into a temporal weighted multi-stream CNN. This attention model is implemented as temporal weighting and it effectively boosts the recognition performance of video representations. Besides, each stream in the proposed ATW CNN framework is capable of end-to-end training, with both network parameters and temporal weights optimized by stochastic gradient descent (SGD) with back-propagation. Experimental results show that the ATW CNN attention mechanism contributes substantially to the performance gains with the more discriminative snippets by focusing on more relevant video segments. == Literature == Seibold VC, Balke J and Rolke B (2023): Temporal attention. Front. Cognit. 2:1168320. doi: 10.3389/fcogn.2023.1168320.
Egocentric vision
Egocentric vision or first-person vision is a sub-field of computer vision that entails analyzing images and videos captured by a wearable camera, which is typically worn on the head or on the chest and naturally approximates the visual field of the camera wearer. Consequently, visual data capture the part of the scene on which the user focuses to carry out the task at hand and offer a valuable perspective to understand the user's activities and their context in a naturalistic setting. The wearable camera looking forwards is often supplemented with a camera looking inward at the user's eye and able to measure a user's eye gaze, which is useful to reveal attention and to better understand the user's activity and intentions. == History == The idea of using a wearable camera to gather visual data from a first-person perspective dates back to the 70s, when Steve Mann invented "Digital Eye Glass", a device that, when worn, causes the human eye itself to effectively become both an electronic camera and a television display. Subsequently, wearable cameras were used for health-related applications in the context of Humanistic Intelligence and Wearable AI. Egocentric vision is best done from the point-of-eye, but may also be done by way of a neck-worn camera when eyeglasses would be in-the-way. This neck-worn variant was popularized by way of the Microsoft SenseCam in 2006 for experimental health research works. The interest of the computer vision community into the egocentric paradigm has been arising slowly entering the 2010s and it is rapidly growing in recent years, boosted by both the impressive advances in the field of wearable technology and by the increasing number of potential applications. The prototypical first-person vision system described by Kanade and Hebert, in 2012 is composed by three basic components: a localization component able to estimate the surrounding, a recognition component able to identify object and people, and an activity recognition component, able to provide information about the current activity of the user. Together, these three components provide a complete situational awareness of the user, which in turn can be used to provide assistance to the user or to the caregiver. Following this idea, the first computational techniques for egocentric analysis focused on hand-related activity recognition and social interaction analysis. Also, given the unconstrained nature of the video and the huge amount of data generated, temporal segmentation and summarization were among the first problems addressed. After almost ten years of egocentric vision (2007–2017), the field is still undergoing diversification. Emerging research topics include: Social saliency estimation Multi-agent egocentric vision systems Privacy preserving techniques and applications Attention-based activity analysis Social interaction analysis Hand pose analysis Ego graphical User Interfaces (EUI) Understanding social dynamics and attention Revisiting robotic vision and machine vision as egocentric sensing Activity forecasting Gaze prediction == Technical challenges == Today's wearable cameras are small and lightweight digital recording devices that can acquire images and videos automatically, without the user intervention, with different resolutions and frame rates, and from a first-person point of view. Therefore, wearable cameras are naturally primed to gather visual information from our everyday interactions since they offer an intimate perspective of the visual field of the camera wearer. Depending on the frame rate, it is common to distinguish between photo-cameras (also called lifelogging cameras) and video-cameras. The former (e.g., Narrative Clip and Microsoft SenseCam), are commonly worn on the chest, and are characterized by a very low frame rate (up to 2fpm) that allows to capture images over a long period of time without the need of recharging the battery. Consequently, they offer considerable potential for inferring knowledge about e.g. behaviour patterns, habits or lifestyle of the user. However, due to the low frame-rate and the free motion of the camera, temporally adjacent images typically present abrupt appearance changes so that motion features cannot be reliably estimated. The latter (e.g., Google Glass, GoPro), are commonly mounted on the head, and capture conventional video (around 35fps) that allows to capture fine temporal details of interactions. Consequently, they offer potential for in-depth analysis of daily or special activities. However, since the camera is moving with the wearer head, it becomes more difficult to estimate the global motion of the wearer and in the case of abrupt movements, the images can result blurred. In both cases, since the camera is worn in a naturalistic setting, visual data present a huge variability in terms of illumination conditions and object appearance. Moreover, the camera wearer is not visible in the image and what he/she is doing has to be inferred from the information in the visual field of the camera, implying that important information about the wearer, such for instance as pose or facial expression estimation, is not available. == Applications == A collection of studies published in a special theme issue of the American Journal of Preventive Medicine has demonstrated the potential of lifelogs captured through wearable cameras from a number of viewpoints. In particular, it has been shown that used as a tool for understanding and tracking lifestyle behaviour, lifelogs would enable the prevention of noncommunicable diseases associated to unhealthy trends and risky profiles (such as obesity and depression). In addition, used as a tool of re-memory cognitive training, lifelogs would enable the prevention of cognitive and functional decline in elderly people. More recently, egocentric cameras have been used to study human and animal cognition, human-human social interaction, human-robot interaction, human expertise in complex tasks. Other applications include navigation/assistive technologies for the blind, monitoring and assistance of industrial workflows, and augmented reality interfaces.
AppValley
AppValley is an independent American digital distribution service operated and trademarked by AppValley LLC. It serves as an alternative app store for the iOS mobile operating system, which allows users to download applications that are not available on the App Store, most commonly tweaked "++" apps, jailbreak apps, and apps including paid apps on the app store. == Legality == AppValley is among several services that violate enterprise developer certificates from Apple. The terms under which these are granted make clear that they are for companies who wish to distribute apps to their employees. AppValley uses these certificates to distribute software directly to non-employees, thereby bypassing the AppStore. AppValley's conduct had implications in U.S. sanctioned markets like Iran, Iraq, North Korea, Cuba, and Venezuela, which have all been subject to commercial sanctions. Among the software offered by AppValley and other services is pirated software, including paid apps on the app store and premium versions of Instagram, Spotify, Pokémon Go, and others. For instance, AppValley distributes an ad-free version of the music streaming app Spotify even on the free tier. == History == The website was founded in May 2017, releasing late that month with a very basic version of the app. There were less than 100 apps available for download at this time. On Jan 19, 2018, a new version dubbed AppValley 2.0 was released bringing dark mode, more categories, a search, and a much faster interface. On February 14, 2019, a Chinese partner "Jason Wu" allegedly took control of the main Twitter account and domain, causing the original AppValley developers to migrate to the domain app-valley.vip and the Twitter account handle @App_Valley_vip. As of September 2024, the app-valley.vip domain now redirects to appvalley.signulous.com. Today, AppValley continues to offer an alternative to Apple's App Store where app developers can publish their applications. == Features == AppValley is a mobile app installer which can also support iOS version that can be installed and downloaded on the mobile or the devices of the people who wish to get access to many different applications available. AppValley also contains apps that have been modified or tweaked for user preferences, and allows the user to by pass national restrictions on the use of apps, without having to resort to jailbreaking. As of June 2, 2020, there are over 1300 apps available for download.
Word error rate
Word error rate (WER) is a common metric of the performance of a speech recognition or machine translation system. The WER metric typically ranges from 0 to 1, where 0 indicates that the compared pieces of text are exactly identical, and 1 (or larger) indicates that they are completely different with no similarity. This way, a WER of 0.8 means that there is an 80% error rate for compared sentences. The general difficulty of measuring performance lies in the fact that the recognized word sequence can have a different length from the reference word sequence (supposedly the correct one). The WER is derived from the Levenshtein distance, working at the word level instead of the phoneme level. The WER is a valuable tool for comparing different systems as well as for evaluating improvements within one system. This kind of measurement, however, provides no details on the nature of translation errors and further work is therefore required to identify the main source(s) of error and to focus any research effort. This problem is solved by first aligning the recognized word sequence with the reference (spoken) word sequence using dynamic string alignment. Examination of this issue is seen through a theory called the power law that states the correlation between perplexity and word error rate. Word error rate can then be computed as: W E R = S + D + I N = S + D + I S + D + C {\displaystyle {\mathit {WER}}={\frac {S+D+I}{N}}={\frac {S+D+I}{S+D+C}}} where S is the number of substitutions, D is the number of deletions, I is the number of insertions, C is the number of correct words, N is the number of words in the reference (N=S+D+C) The intuition behind 'deletion' and 'insertion' is how to get from the reference to the hypothesis. So if we have the reference "This is wikipedia" and hypothesis "This _ wikipedia", we call it a deletion. Note that since N is the number of words in the reference, the word error rate can be larger than 1.0, namely if the number of insertions I is larger than the number of correct words C. When reporting the performance of a speech recognition system, sometimes word accuracy (WAcc) is used instead: W A c c = 1 − W E R = N − S − D − I N = C − I N {\displaystyle {\mathit {WAcc}}=1-{\mathit {WER}}={\frac {N-S-D-I}{N}}={\frac {C-I}{N}}} Since the WER can be larger than 1.0, the word accuracy can be smaller than 0.0. == Experiments == It is commonly believed that a lower word error rate shows superior accuracy in recognition of speech, compared with a higher word error rate. However, at least one study has shown that this may not be true. In a Microsoft Research experiment, it was shown that, if people were trained under "that matches the optimization objective for understanding", (Wang, Acero and Chelba, 2003) they would show a higher accuracy in understanding of language than other people who demonstrated a lower word error rate, showing that true understanding of spoken language relies on more than just high word recognition accuracy. == Other metrics == One problem with using a generic formula such as the one above, however, is that no account is taken of the effect that different types of error may have on the likelihood of successful outcome, e.g. some errors may be more disruptive than others and some may be corrected more easily than others. These factors are likely to be specific to the syntax being tested. A further problem is that, even with the best alignment, the formula cannot distinguish a substitution error from a combined deletion plus insertion error. Hunt (1990) has proposed the use of a weighted measure of performance accuracy where errors of substitution are weighted at unity but errors of deletion and insertion are both weighted only at 0.5, thus: W E R = S + 0.5 D + 0.5 I N {\displaystyle {\mathit {WER}}={\frac {S+0.5D+0.5I}{N}}} There is some debate, however, as to whether Hunt's formula may properly be used to assess the performance of a single system, as it was developed as a means of comparing more fairly competing candidate systems. A further complication is added by whether a given syntax allows for error correction and, if it does, how easy that process is for the user. There is thus some merit to the argument that performance metrics should be developed to suit the particular system being measured. Whichever metric is used, however, one major theoretical problem in assessing the performance of a system is deciding whether a word has been “mis-pronounced,” i.e. does the fault lie with the user or with the recogniser. This may be particularly relevant in a system which is designed to cope with non-native speakers of a given language or with strong regional accents. The pace at which words should be spoken during the measurement process is also a source of variability between subjects, as is the need for subjects to rest or take a breath. All such factors may need to be controlled in some way. For text dictation it is generally agreed that performance accuracy at a rate below 95% is not acceptable, but this again may be syntax and/or domain specific, e.g. whether there is time pressure on users to complete the task, whether there are alternative methods of completion, and so on. The term "Single Word Error Rate" is sometimes referred to as the percentage of incorrect recognitions for each different word in the system vocabulary. == Edit distance == The word error rate may also be referred to as the length normalized edit distance. The normalized edit distance between X and Y, d( X, Y ) is defined as the minimum of W( P ) / L ( P ), where P is an editing path between X and Y, W ( P ) is the sum of the weights of the elementary edit operations of P, and L(P) is the number of these operations (length of P).
Deep learning in photoacoustic imaging
Photoacoustic imaging (PA) is based on the photoacoustic effect, in which optical absorption causes a rise in temperature, which causes a subsequent rise in pressure via thermo-elastic expansion. This pressure rise propagates through the tissue and is sensed via ultrasonic transducers. Due to the proportionality between the optical absorption, the rise in temperature, and the rise in pressure, the ultrasound pressure wave signal can be used to quantify the original optical energy deposition within the tissue. Photoacoustic imaging has applications of deep learning in both photoacoustic computed tomography (PACT) and photoacoustic microscopy (PAM). PACT utilizes wide-field optical excitation and an array of unfocused ultrasound transducers. Similar to other computed tomography methods, the sample is imaged at multiple view angles, which are then used to perform an inverse reconstruction algorithm based on the detection geometry (typically through universal backprojection, modified delay-and-sum, or time reversal ) to elicit the initial pressure distribution within the tissue. PAM on the other hand uses focused ultrasound detection combined with weakly focused optical excitation (acoustic resolution PAM or AR-PAM) or tightly focused optical excitation (optical resolution PAM or OR-PAM). PAM typically captures images point-by-point via a mechanical raster scanning pattern. At each scanned point, the acoustic time-of-flight provides axial resolution while the acoustic focusing yields lateral resolution. == Applications of deep learning in PACT == The first application of deep learning in PACT was by Reiter et al. in which a deep neural network was trained to learn spatial impulse responses and locate photoacoustic point sources. The resulting mean axial and lateral point location errors on 2,412 of their randomly selected test images were 0.28 mm and 0.37 mm respectively. After this initial implementation, the applications of deep learning in PACT have branched out primarily into removing artifacts from acoustic reflections, sparse sampling, limited-view, and limited-bandwidth. There has also been some recent work in PACT toward using deep learning for wavefront localization. There have been networks based on fusion of information from two different reconstructions to improve the reconstruction using deep learning fusion based networks. === Using deep learning to locate photoacoustic point sources === Traditional photoacoustic beamforming techniques modeled photoacoustic wave propagation by using detector array geometry and the time-of-flight to account for differences in the PA signal arrival time. However, this technique failed to account for reverberant acoustic signals caused by acoustic reflection, resulting in acoustic reflection artifacts that corrupt the true photoacoustic point source location information. In Reiter et al., a convolutional neural network (similar to a simple VGG-16 style architecture) was used that took pre-beamformed photoacoustic data as input and outputted a classification result specifying the 2-D point source location. ==== Deep learning for PA wavefront localization ==== Johnstonbaugh et al. was able to localize the source of photoacoustic wavefronts with a deep neural network. The network used was an encoder-decoder style convolutional neural network. The encoder-decoder network was made of residual convolution, upsampling, and high field-of-view convolution modules. A Nyquist convolution layer and differentiable spatial-to-numerical transform layer were also used within the architecture. Simulated PA wavefronts served as the input for training the model. To create the wavefronts, the forward simulation of light propagation was done with the NIRFast toolbox and the light-diffusion approximation, while the forward simulation of sound propagation was done with the K-Wave toolbox. The simulated wavefronts were subjected to different scattering mediums and Gaussian noise. The output for the network was an artifact free heat map of the targets axial and lateral position. The network had a mean error rate of less than 30 microns when localizing target below 40 mm and had a mean error rate of 1.06 mm for localizing targets between 40 mm and 60 mm. With a slight modification to the network, the model was able to accommodate multi target localization. A validation experiment was performed in which pencil lead was submerged into an intralipid solution at a depth of 32 mm. The network was able to localize the lead's position when the solution had a reduced scattering coefficient of 0, 5, 10, and 15 cm−1. The results of the network show improvements over standard delay-and-sum or frequency-domain beamforming algorithms and Johnstonbaugh proposes that this technology could be used for optical wavefront shaping, circulating melanoma cell detection, and real-time vascular surgeries. === Removing acoustic reflection artifacts (in the presence of multiple sources and channel noise) === Building on the work of Reiter et al., Allman et al. utilized a full VGG-16 architecture to locate point sources and remove reflection artifacts within raw photoacoustic channel data (in the presence of multiple sources and channel noise). This utilization of deep learning trained on simulated data produced in the MATLAB k-wave library, and then later reaffirmed their results on experimental data. === Ill-posed PACT reconstruction === In PACT, tomographic reconstruction is performed, in which the projections from multiple solid angles are combined to form an image. When reconstruction methods like filtered backprojection or time reversal, are ill-posed inverse problems due to sampling under the Nyquist-Shannon's sampling requirement or with limited-bandwidth/view, the resulting reconstruction contains image artifacts. Traditionally these artifacts were removed with slow iterative methods like total variation minimization, but the advent of deep learning approaches has opened a new avenue that utilizes a priori knowledge from network training to remove artifacts. In the deep learning methods that seek to remove these sparse sampling, limited-bandwidth, and limited-view artifacts, the typical workflow involves first performing the ill-posed reconstruction technique to transform the pre-beamformed data into a 2-D representation of the initial pressure distribution that contains artifacts. Then, a convolutional neural network (CNN) is trained to remove the artifacts, in order to produce an artifact-free representation of the ground truth initial pressure distribution. ==== Using deep learning to remove sparse sampling artifacts ==== When the density of uniform tomographic view angles is under what is prescribed by the Nyquist-Shannon's sampling theorem, it is said that the imaging system is performing sparse sampling. Sparse sampling typically occurs as a way of keeping production costs low and improving image acquisition speed. The typical network architectures used to remove these sparse sampling artifacts are U-net and Fully Dense (FD) U-net. Both of these architectures contain a compression and decompression phase. The compression phase learns to compress the image to a latent representation that lacks the imaging artifacts and other details. The decompression phase then combines with information passed by the residual connections in order to add back image details without adding in the details associated with the artifacts. FD U-net modifies the original U-net architecture by including dense blocks that allow layers to utilize information learned by previous layers within the dense block. Another technique was proposed using a simple CNN based architecture for removal of artifacts and improving the k-wave image reconstruction. ==== Removing limited-view artifacts with deep learning ==== When a region of partial solid angles are not captured, generally due to geometric limitations, the image acquisition is said to have limited-view. As illustrated by the experiments of Davoudi et al., limited-view corruptions can be directly observed as missing information in the frequency domain of the reconstructed image. Limited-view, similar to sparse sampling, makes the initial reconstruction algorithm ill-posed. Prior to deep learning, the limited-view problem was addressed with complex hardware such as acoustic deflectors and full ring-shaped transducer arrays, as well as solutions like compressed sensing, weighted factor, and iterative filtered backprojection. The result of this ill-posed reconstruction is imaging artifacts that can be removed by CNNs. The deep learning algorithms used to remove limited-view artifacts include U-net and FD U-net, as well as generative adversarial networks (GANs) and volumetric versions of U-net. One GAN implementation of note improved upon U-net by using U-net as a generator and VGG as a discriminator, with the Wasserstein metric and gradient penalty to stabilize training (WGAN-GP). ==== Pixel-wise interpolation