Probabilistic modeling and inference are core tools in diverse fields including statistics, machine learning, computer vision, cognitive science, robotics, natural language processing, and artificial intelligence. To meet the functional requirements of applications, practitioners use a broad range of modeling techniques and approximate inference algorithms. However, implementing inference algorithms is often difficult and error prone. Gen simplifies the use of probabilistic modeling and inference, by providing modeling languages in which users express models, and high-level programming constructs that automate aspects of inference. Like some probabilistic programming research languages, Gen includes universal modeling languages that can represent any model, including models with stochastic structure, discrete and continuous random variables, and simulators. However, Gen is distinguished by the flexibility that it affords to users for customizing their inference algorithm. It is possible to use built-in algorithms that require only a couple lines of code, as well as develop custom algorithms that are more able to meet scalability and efficiency requirements. Gen’s flexible modeling and inference programming capabilities unify symbolic, neural, probabilistic, and simulation-based approaches to modeling and inference, including causal modeling, symbolic programming, deep learning, hierarchical Bayesiam modeling, graphics and physics engines, and planningand reinforcement learning.
Many ideas were based on a paper by Yan Ke, Derek Hoiem, and Rahul Sukthankar called “Computer Vision for Music Identification” (2005). In fact, even the Last.fm fingerprinter uses the code published by the authors of this paper. This is where I learned that audio identification is more about machine learning that it is about DSP. Many useful methods for extracting interesting features from audio streams are well-known and the problem is more about how to apply and index them the best way. The basic idea here is to treat audio as a spectral image and index the content of the image. I’ll explain this in more detail and how Chromaprint uses this in a following post. Another important paper for me was “Pairwise Boosted Audio Fingerprint” (2009) by Dalwon Jang, Chang D. Yoo, Sunil Lee, Sungwoong Kim and Ton Kalker (Ton Kalker is a co-author of a historically important paper “Robust Audio Hashing for Content Identification” (2001) published by Philips Research), which combined previous experiments of the authors with audio identification based on spectral centroid features and the indexing approach similar to the one suggested by Y. Ke, D. Hoiem and R. Sukthankar. For a long time this was the best solution I had and since it was actually not very hard to implement, the most time I spent on tweaking the configuration to get the best results. The last major change came after I learned about “chroma” features by reading the “Efficient Index-Based Audio Matching” (2008) by Frank Kurth and Meinard Müller. I’ve read more papers about chroma features later, but this was the first and also the most important one for me and some ideas about processing the feature vectors from it are implemented in Chromaprint. Chroma features are typically used for music identification, as opposed to audio file identification, but I tried to use them with the approach I already had implemented and it nicely improved the quality of the fingerprinting function and actually reduced complexity which allowed me to use much larger training data sets.
GPT-2 displays a broad set of capabilities, including the ability to generate conditional synthetic text samples of unprecedented quality, where we prime the model with an input and have it generate a lengthy continuation. In addition, GPT-2 outperforms other language models trained on specific domains (like Wikipedia, news, or books) without needing to use these domain-specific training datasets. On language tasks like question answering, reading comprehension, summarization, and translation, GPT-2 begins to learn these tasks from the raw text, using no task-specific training data. While scores on these downstream tasks are far from state-of-the-art, they suggest that the tasks can benefit from unsupervised techniques, given sufficient (unlabeled) data and compute.
In Hungary, Latvia, and Greece, travelers will be given an automated lie-detection test—by an animated AI border agent. The system, called iBorderCtrl, is part of a six-month pilot led by the Hungarian National Police at four different border crossing points. “We’re employing existing and proven technologies—as well as novel ones—to empower border agents to increase the accuracy and efficiency of border checks,” project coordinator George Boultadakis of European Dynamics in Luxembourg told the European Commission. “iBorderCtrl’s system will collect data that will move beyond biometrics and on to biomarkers of deceit.”
Some of these skilled lawyers did question whether their profession could ever entirely trust automation to make skilled legal decisions. For a small number, they suggested they would be sticking to “reliable” manual processes for the immediate future. However, most of the participants stressed that high-volume and low-risk contracts took up too much of their time, and felt it was incumbent on lawyers to automate work when, and where, possible. For them, the study was also a simple, practical demonstration of a not-so-scary AI future. However, lawyers also stressed that undue weight should not be put on legal AI alone. One participant, Justin Brown, stressed that humans must use new technology alongside their lawyerly instincts. He says: “Either working alone is inferior to the combination of both.”
That something else, call it imagination or call it dreaming, does not require validation with immediate reality. The closest incarnation we have today is the generative adversarial network (GAN). A GAN consists of two networks, a generator and a discriminator. One can consider a discriminator as a neural network that acts in concert with the objective function. That is, it validates an internal generator network with reality. The generator is an automation that recreates an approximation of reality. A GAN works using back-propagation and it does perform unsupervised learning. So perhaps unsupervised learn doesn’t require an objective function, however it may still need back-propagation.
Algorithmic music composition has developed a lot in the last few years, but the idea has a long history. In some sense, the first automatic music came from nature: Chinese windchimes, ancient Greek wind-powered Aeolian harps, or the Japanese water instrument suikinkutsu. But in the 1700s music became “algorithmic”: Musikalisches Würfelspiel, a game that generates short piano compositions from fragments, with choices made by dice.
Dice games, Markov chains, and RNNs aren’t the only ways to make algorithmic music. Some machine learning practitioners explore alternative approaches like hierarchical temporal memory, or principal components analysis. But I’m focusing on neural nets because they are responsible for most of the big changes recently. (Though even within the domain of neural nets there are some directions I’m leaving out that have fewer examples, such as restricted Boltzmann machines for composing 4-bar jazz licks, short variations on a single song, or hybrid RNN-RBM models, or hybrid autoencoder-LSTM models.)
wo summers ago, Courtenay Cotton led a workshop on machine learning that I attended with a New York–based group called the Women and Surveillance Initiative. It was a welcome introduction to the subject and a rare opportunity to cut through the hype to understand both the value of machine learning and the complications of this field of research. In our recent interview, Cotton, who now works as lead data scientist at n-Join, once again offered her clear thinking on machine learning and where it is headed.
As was the case with the mobile revolution, and the web before that, ML will cause us to rethink, restructure, displace, and consider new possibilities for virtually every experience we build. In the Google UX community, we’ve started an effort called “human-centered machine learning” (HCML) to help focus and guide that conversation. Using this lens, we look across products to see how ML can stay grounded in human needs while solving them in unique ways only possible through ML. Our team at Google works with UXers across the company to bring them up to speed on core ML concepts, understand how to integrate ML into the UX utility belt, and ensure ML and AI are built in inclusive ways. We’ve developed seven points to help designers navigate the new terrain of designing ML-driven products. Born out of our work with UX and AI teams at Google (and a healthy dose of trial and error), these points will help you put the user first, iterate quickly, and understand the unique opportunities ML creates.
Don’t expect Machine learning to figure out what problems to solve
Ask yourself if ML will address the problem in a unique way
Fake it with personal examples and wizards
Weigh the costs of false positives and false negatives
“The video, called “Alternative Face v1.1”, is the work of Mario Klingemann, a German artist. It plays audio from an NBC interview with Ms Conway through the mouth of Ms Hardy’s digital ghost. The video is wobbly and pixelated; a competent visual-effects shop could do much better. But Mr Klingemann did not fiddle with editing software to make it. Instead, he took only a few days to create the clip on a desktop computer using a generative adversarial network (GAN), a type of machine-learning algorithm. His computer spat it out automatically after being force fed old music videos of Ms Hardy. It is a recording of something that never happened.”
AlphaGo is made up of a number of relatively standard techniques: behavior cloning (supervised learning on human demonstration data), reinforcement learning (REINFORCE), value functions, and Monte Carlo Tree Search (MCTS). However, the way these components are combined is novel and not exactly standard. In particular, AlphaGo uses a SL (supervised learning) policy to initialize the learning of an RL (reinforcement learning) policy that gets perfected with self-play, which they then estimate a value function from, which then plugs into MCTS that (somewhat surprisingly) uses the (worse!, but more diverse) SL policy to sample rollouts. In addition, the policy/value nets are deep neural networks, so getting everything to work properly presents its own unique challenges (e.g. value function is trained in a tricky way to prevent overfitting). On all of these aspects, DeepMind has executed very well. That being said, AlphaGo does not by itself use any fundamental algorithmic breakthroughs in how we approach RL problems.
Until a few years ago, machine learning algorithms simply did not work very well on many meaningful tasks like recognizing objects or translation. Thus, when a machine learning algorithm failed to do the right thing, this was the rule, rather than the exception. Today, machine learning algorithms have advanced to the next stage of development: when presented with naturally occurring inputs, they can outperform humans. Machine learning has not yet reached true human-level performance, because when confronted by even a trivial adversary, most machine learning algorithms fail dramatically. In other words, we have reached the point where machine learning works, but may easily be broken.
In this paper, we demonstrated techniques for generating accessories in the form of eyeglass frames that, when printed and worn, can effectively fool state-of-the-art face-recognition systems. Our research builds on recent research in fooling machine-learning classifiers by perturbing inputs in an adversarial way, but does so with attention to two novel goals: the perturbations must be physically realizable and inconspicuous. We showed that our eyeglass frames enabled subjects to both dodge recognition and to impersonate others. We believe that our demonstration of techniques to realize these goals through printed eyeglass frames is both novel and important, and should inform future deliberations on the extent to which ML can be trusted in adversarial settings. Finally, we extended our work in two additional directions, first, to so-called black-box FRSs that can be queried but for which the internals are not known, and, second, to defeat state-of-the-art face detection systems.
