jm + ml   34

A history of the neural net/tank legend in AI, and other examples of reward hacking
@gwern: "A history of the neural net/tank legend in AI: https://t.co/2s4AOGMS3a (Feel free to suggest more sightings or examples of reward hacking!)"
gwern  history  ai  machine-learning  ml  genetic-algorithms  neural-networks  perceptron  learning  training  data  reward-hacking 
8 weeks ago by jm
Spotify’s Discover Weekly: How machine learning finds your new music
Not sure how accurate this is (it's not written by a Spotify employee), but seems pretty well researched -- according to this Discover Weekly is a mix of 3 different algorithms
discover-weekly  spotify  nlp  music  ai  ml  machine-learning 
8 weeks ago by jm
Universal adversarial perturbations
in today’s paper Moosavi-Dezfooli et al., show us how to create a _single_ perturbation that causes the vast majority of input images to be misclassified.
adversarial-classification  spam  image-recognition  ml  machine-learning  dnns  neural-networks  images  classification  perturbation  papers 
september 2017 by jm
The Dark Secret at the Heart of AI - MIT Technology Review
'The mysterious mind of [NVidia's self-driving car, driven by machine learning] points to a looming issue with artificial intelligence. The car’s underlying AI technology, known as deep learning, has proved very powerful at solving problems in recent years, and it has been widely deployed for tasks like image captioning, voice recognition, and language translation. There is now hope that the same techniques will be able to diagnose deadly diseases, make million-dollar trading decisions, and do countless other things to transform whole industries.

But this won’t happen—or shouldn’t happen—unless we find ways of making techniques like deep learning more understandable to their creators and accountable to their users. Otherwise it will be hard to predict when failures might occur—and it’s inevitable they will. That’s one reason Nvidia’s car is still experimental.

Already, mathematical models are being used to help determine who makes parole, who’s approved for a loan, and who gets hired for a job. If you could get access to these mathematical models, it would be possible to understand their reasoning. But banks, the military, employers, and others are now turning their attention to more complex machine-learning approaches that could make automated decision-making altogether inscrutable. Deep learning, the most common of these approaches, represents a fundamentally different way to program computers. “It is a problem that is already relevant, and it’s going to be much more relevant in the future,” says Tommi Jaakkola, a professor at MIT who works on applications of machine learning. “Whether it’s an investment decision, a medical decision, or maybe a military decision, you don’t want to just rely on a ‘black box’ method.”'
ai  algorithms  ml  machine-learning  legibility  explainability  deep-learning  nvidia 
may 2017 by jm
'What’s your ML Test Score? A rubric for ML production systems'
'Using machine learning in real-world production systems is complicated by a host of issues not found in small toy examples or even large offline research experiments. Testing and monitoring are key considerations for assessing the production-readiness of an ML system. But how much testing and monitoring is enough? We present an ML Test Score rubric based on a set of actionable tests to help quantify these issues.'

Google paper on testable machine learning systems.
machine-learning  testing  ml  papers  google 
april 2017 by jm
Zeynep Tufekci: Machine intelligence makes human morals more important | TED Talk | TED.com
Machine intelligence is here, and we're already using it to make subjective decisions. But the complex way AI grows and improves makes it hard to understand and even harder to control. In this cautionary talk, techno-sociologist Zeynep Tufekci explains how intelligent machines can fail in ways that don't fit human error patterns — and in ways we won't expect or be prepared for. "We cannot outsource our responsibilities to machines," she says. "We must hold on ever tighter to human values and human ethics."


More relevant now that nVidia are trialing ML-based self-driving cars in the US...
nvidia  ai  ml  machine-learning  scary  zeynep-tufekci  via:maciej  technology  ted-talks 
april 2017 by jm
Research Blog: Federated Learning: Collaborative Machine Learning without Centralized Training Data
Great stuff from Google - this is really nifty stuff for large-scale privacy-preserving machine learning usage:

It works like this: your device downloads the current model, improves it by learning from data on your phone, and then summarizes the changes as a small focused update. Only this update to the model is sent to the cloud, using encrypted communication, where it is immediately averaged with other user updates to improve the shared model. All the training data remains on your device, and no individual updates are stored in the cloud.

Federated Learning allows for smarter models, lower latency, and less power consumption, all while ensuring privacy. And this approach has another immediate benefit: in addition to providing an update to the shared model, the improved model on your phone can also be used immediately, powering experiences personalized by the way you use your phone.

Papers:
https://arxiv.org/pdf/1602.05629.pdf , https://arxiv.org/pdf/1610.05492.pdf
google  ml  machine-learning  training  federated-learning  gboard  models  privacy  data-privacy  data-protection 
april 2017 by jm
[1606.08813] European Union regulations on algorithmic decision-making and a "right to explanation"
We summarize the potential impact that the European Union's new General Data Protection Regulation will have on the routine use of machine learning algorithms. Slated to take effect as law across the EU in 2018, it will restrict automated individual decision-making (that is, algorithms that make decisions based on user-level predictors) which "significantly affect" users. The law will also effectively create a "right to explanation," whereby a user can ask for an explanation of an algorithmic decision that was made about them. We argue that while this law will pose large challenges for industry, it highlights opportunities for computer scientists to take the lead in designing algorithms and evaluation frameworks which avoid discrimination and enable explanation.


oh this'll be tricky.
algorithms  accountability  eu  gdpr  ml  machine-learning  via:daveb  europe  data-protection  right-to-explanation 
march 2017 by jm
'Rules of Machine Learning: Best Practices for ML Engineering' from Martin Zinkevich
'This document is intended to help those with a basic knowledge of machine learning get the benefit of best practices in machine learning from around Google. It presents a style for machine learning, similar to the Google C++ Style Guide and other popular guides to practical programming. If you have taken a class in machine learning, or built or worked on a machine­-learned model, then you have the necessary background to read this document.'

Full of good tips, if you wind up using ML in a production service.
machine-learning  ml  google  production  coding  best-practices 
january 2017 by jm
How a Machine Learns Prejudice - Scientific American
Agreed, this is a big issue.
If artificial intelligence takes over our lives, it probably won’t involve humans battling an army of robots that relentlessly apply Spock-like logic as they physically enslave us. Instead, the machine-learning algorithms that already let AI programs recommend a movie you’d like or recognize your friend’s face in a photo will likely be the same ones that one day deny you a loan, lead the police to your neighborhood or tell your doctor you need to go on a diet. And since humans create these algorithms, they're just as prone to biases that could lead to bad decisions—and worse outcomes.
These biases create some immediate concerns about our increasing reliance on artificially intelligent technology, as any AI system designed by humans to be absolutely "neutral" could still reinforce humans’ prejudicial thinking instead of seeing through it.
prejudice  bias  machine-learning  ml  data  training  race  racism  google  facebook 
january 2017 by jm
Here's Why Facebook's Trending Algorithm Keeps Promoting Fake News - BuzzFeed News
Kalina Bontcheva leads the EU-funded PHEME project working to compute the veracity of social media content. She said reducing the amount of human oversight for Trending heightens the likelihood of failures, and of the algorithm being fooled by people trying to game it.
“I think people are always going to try and outsmart these algorithms — we’ve seen this with search engine optimization,” she said. “I’m sure that once in a while there is going to be a very high-profile failure.”
Less human oversight means more reliance on the algorithm, which creates a new set of concerns, according to Kate Starbird, an assistant professor at the University of Washington who has been using machine learning and other technology to evaluate the accuracy of rumors and information during events such as the Boston bombings.
“[Facebook is] making an assumption that we’re more comfortable with a machine being biased than with a human being biased, because people don’t understand machines as well,” she said.
facebook  news  gaming  adversarial-classification  pheme  truth  social-media  algorithms  ml  machine-learning  media 
october 2016 by jm
Founder of Google X has no concept of how machine learning as policing tool risks reinforcing implicit bias
This is shocking:
At the end of the panel on artificial intelligence, a young black woman asked [Sebastian Thrun, CEO of the education startup Udacity, who is best known for founding Google X] whether bias in machine learning “could perpetuate structural inequality at a velocity much greater than perhaps humans can.” She offered the example of criminal justice, where “you have a machine learning tool that can identify criminals, and criminals may disproportionately be black because of other issues that have nothing to do with the intrinsic nature of these people, so the machine learns that black people are criminals, and that’s not necessarily the outcome that I think we want.”
In his reply, Thrun made it sound like her concern was one about political correctness, not unconscious bias. “Statistically what the machines do pick up are patterns and sometimes we don’t like these patterns. Sometimes they’re not politically correct,” Thrun said. “When we apply machine learning methods sometimes the truth we learn really surprises us, to be honest, and I think it’s good to have a dialogue about this.”


"the truth"! Jesus. We are fucked
google  googlex  bias  racism  implicit-bias  machine-learning  ml  sebastian-thrun  udacity  inequality  policing  crime 
october 2016 by jm
Image Synthesis from Yahoo's open_nsfw
What makes an image NSFW, according to Yahoo? I explore this question with a clever new visualization technique


Deep Dream applied to an NSFW classifier. This is a bit NSFW, as it happens
nsfw  yahoo  ml  deep-dream  images  porn 
october 2016 by jm
Remarks at the SASE Panel On The Moral Economy of Tech
Excellent talk. I love this analogy for ML applied to real-world data which affects people:
Treating the world as software promotes fantasies of control. And the best kind of control is control without responsibility. Our unique position as authors of software used by millions gives us power, but we don't accept that this should make us accountable. We're programmers—who else is going to write the software that runs the world? To put it plainly, we are surprised that people seem to get mad at us for trying to help. Fortunately we are smart people and have found a way out of this predicament. Instead of relying on algorithms, which we can be accused of manipulating for our benefit, we have turned to machine learning, an ingenious way of disclaiming responsibility for anything. Machine learning is like money laundering for bias. It's a clean, mathematical apparatus that gives the status quo the aura of logical inevitability. The numbers don't lie.


Particularly apposite today given Y Combinator's revelation that they use an AI bot to help 'sift admission applications', and don't know what criteria it's using: https://twitter.com/aprjoy/status/783032128653107200
culture  ethics  privacy  technology  surveillance  ml  machine-learning  bias  algorithms  software  control 
october 2016 by jm
Airflow/AMI/ASG nightly-packaging workflow
Some tantalising discussion on twitter of an Airflow + AMI + ASG workflow for ML packaging:

'We build models using Airflow. We deploy new models as AMIs where each AMI is model + scoring code. The AMI is hence a version of code + model at a point in time : #immutable_infrastructure. It's natural for Airflow to build & deploy the model+code with each Airflow DAG Run corresponding to a versioned AMI. if there's a problem, we can simply roll back to the previous AMI & identify the problematic model building Dag run. Since we use ASGs, Airflow can execute a rolling deploy of new AMIs. We could also have it do a validation & ASG rollback of the AMI if validation fails. Airflow is being used for reliable Model build+validation+deployment.'
ml  packaging  airflow  asg  ami  deployment  ops  infrastructure  rollback 
september 2016 by jm
How a Japanese cucumber farmer is using deep learning and TensorFlow
Unfortunately the usual ML problem arises at the end:
One of the current challenges with deep learning is that you need to have a large number of training datasets. To train the model, Makoto spent about three months taking 7,000 pictures of cucumbers sorted by his mother, but it’s probably not enough. "When I did a validation with the test images, the recognition accuracy exceeded 95%. But if you apply the system with real use cases, the accuracy drops down to about 70%. I suspect the neural network model has the issue of "overfitting" (the phenomenon in neural network where the model is trained to fit only to the small training dataset) because of the insufficient number of training images."


In other words, as with ML since we were using it in SpamAssassin, maintaining the training corpus becomes a really big problem. :(
google  machine-learning  tensorflow  cucumbers  deep-learning  ml 
september 2016 by jm
Hey Microsoft, the Internet Made My Bot Racist, Too
All machine learning algorithms strive to exaggerate and perpetuate the past. That is, after all, what they are learning from. The fundamental assumption of every machine learning algorithm is that the past is correct, and anything coming in the future will be, and should be, like the past. This is a fine assumption to make when you are Netflix trying to predict what movie you’ll like, but is immoral when applied to many other situations. For bots like mine and Microsoft’s, built for entertainment purposes, it can lead to embarrassment. But AI has started to be used in much more meaningful ways: predictive policing in Chicago, for example, has already led to widespread accusations of racial profiling.
This isn’t a little problem. This is a huge problem, and it demands a lot more attention then it’s getting now, particularly in the community of scientists and engineers who design and apply these algorithms. It’s one thing to get cursed out by an AI, but wholly another when one puts you in jail, denies you a mortgage, or decides to audit you.
machine-learning  ml  algorithms  future  society  microsoft 
march 2016 by jm
Fast Forward Labs: Fashion Goes Deep: Data Science at Lyst
this is more than just data science really -- this is proper machine learning, with deep learning and a convolutional neural network. serious business
lyst  machine-learning  data-science  ml  neural-networks  supervised-learning  unsupervised-learning  deep-learning 
december 2015 by jm
"Hidden Technical Debt in Machine-Learning Systems" [pdf]
Another great paper about from Google, talking about the tradeoffs that must be considered in practice over the long term with running a complex ML system in production.
technical-debt  ml  machine-learning  ops  software  production  papers  pdf  google 
december 2015 by jm
Control theory meets machine learning
'DB: Is there a difference between how control theorists and machine learning researchers think about robustness and error?

BR: In machine learning, we almost always model our errors as being random rather than worst-case. In some sense, random errors are actually much more benign than worst-case errors. [...] In machine learning, by assuming average-case performance, rather than worst-case, we can design predictive algorithms by averaging out the errors over large data sets. We want to be robust to fluctuations in the data, but only on average. This is much less restrictive than the worst-case restrictions in controls.

DB: So control theory is model-based and concerned with worst case. Machine learning is data based and concerned with average case. Is there a middle ground?

BR: I think there is! And I think there's an exciting opportunity here to understand how to combine robust control and reinforcement learning. Being able to build systems from data alone simplifies the engineering process, and has had several recent promising results. Guaranteeing that these systems won't behave catastrophically will enable us to actually deploy machine learning systems in a variety of applications with major impacts on our lives. It might enable safe autonomous vehicles that can navigate complex terrains. Or could assist us in diagnostics and treatments in health care. There are a lot of exciting possibilities, and that's why I'm excited about how to find a bridge between these two viewpoints.'
control-theory  interviews  machine-learning  ml  worst-case  self-driving-cars  cs 
november 2015 by jm
Mining High-Speed Data Streams: The Hoeffding Tree Algorithm
This paper proposes a decision tree learner for data streams, the Hoeffding Tree algorithm, which comes with the guarantee that the learned decision tree is asymptotically nearly identical to that of a non-incremental learner using infinitely many examples. This work constitutes a significant step in developing methodology suitable for modern ‘big data’ challenges and has initiated a lot of follow-up research. The Hoeffding Tree algorithm has been covered in various textbooks and is available in several public domain tools, including the WEKA Data Mining platform.
hoeffding-tree  algorithms  data-structures  streaming  streams  cep  decision-trees  ml  learning  papers 
august 2015 by jm
Recommender Systems (Machine Learning Summer School 2014 @ CMU)
Extremely authoritative slide deck on building a recommendation system, from Xavier Amatriain, Research/Engineering Manager at Netflix
netflix  recommendations  recommenders  ml  machine-learning  cmu  clustering  algorithms 
august 2015 by jm
The reusable holdout: Preserving validity in adaptive data analysis
Useful stats hack from Google: "We show how to safely reuse a holdout data set many times to validate the results of adaptively chosen analyses."
statistics  google  reusable-holdout  training  ml  machine-learning  data-analysis  holdout  corpus  sampling 
august 2015 by jm
Outlier Detection at Netflix | Hacker News
Excellent HN thread re automated anomaly detection in production, Q&A with the dev team
machine-learning  ml  remediation  anomaly-detection  netflix  ops  time-series  clustering 
july 2015 by jm
Top 10 data mining algorithms in plain English
This is a phenomenally useful ML/data-mining resource post -- 'the top 10 most influential data mining algorithms as voted on by 3 separate panels in [ICDM '06's] survey paper', but with a nice clear intro and description for each one. Here's the algorithms covered:
1. C4.5
2. k-means
3. Support vector machines
4. Apriori
5. EM
6. PageRank
7. AdaBoost
8. kNN
9. Naive Bayes
10. CART
svm  k-means  c4.5  apriori  em  pagerank  adaboost  knn  naive-bayes  cart  ml  data-mining  machine-learning  papers  algorithms  unsupervised  supervised 
may 2015 by jm
How to do named entity recognition: machine learning oversimplified
Good explanation of this NLP tokenization/feature-extraction technique. Example result: "Jimi/B-PER Hendrix/I-PER played/O at/O Woodstock/B-LOC ./O"
named-entities  feature-extraction  tokenization  nlp  ml  algorithms  machine-learning 
may 2015 by jm
Amazon Machine Learning
Upsides of this new AWS service:

* great UI and visualisations.

* solid choice of metric to evaluate the results. Maybe things moved on since I was working on it, but the use of AUC, false positives and false negatives was pretty new when I was working on it. (er, 10 years ago!)

Downsides:

* it could do with more support for unsupervised learning algorithms. Supervised learning means you need to provide training data, which in itself can be hard work. My experience with logistic regression in the past is that it requires very accurate training data, too -- its tolerance for misclassified training examples is poor.

* Also, in my experience, 80% of the hard work of using ML algorithms is writing good tokenisation and feature extraction algorithms. I don't see any help for that here unfortunately. (probably not that surprising as it requires really detailed knowledge of the input data to know what classes can be abbreviated into a single class, etc.)
amazon  aws  ml  machine-learning  auc  data-science 
april 2015 by jm
Twitter’s new anti-harassment filter
Twitter is calling it a “quality filter,” and it’s been rolling out to verified users running Twitter’s iOS app since last week. It appears to work much like a spam filter, except instead of hiding bots and copy-paste marketers, it screens “threats, offensive language, [and] duplicate content” out of your notifications feed.


via Nelson
via:nelson  harassment  spam  twitter  gamergame  abuse  ml 
april 2015 by jm
'Machine Learning: The High-Interest Credit Card of Technical Debt' [PDF]
Oh god yes. This is absolutely spot on, as you would expect from a Google paper -- at this stage they probably have accumulated more real-world ML-at-scale experience than anywhere else.

'Machine learning offers a fantastically powerful toolkit for building complex systems
quickly. This paper argues that it is dangerous to think of these quick wins
as coming for free. Using the framework of technical debt, we note that it is remarkably
easy to incur massive ongoing maintenance costs at the system level
when applying machine learning. The goal of this paper is highlight several machine
learning specific risk factors and design patterns to be avoided or refactored
where possible. These include boundary erosion, entanglement, hidden feedback
loops, undeclared consumers, data dependencies, changes in the external world,
and a variety of system-level anti-patterns.

[....]

'In this paper, we focus on the system-level interaction between machine learning code and larger systems
as an area where hidden technical debt may rapidly accumulate. At a system-level, a machine
learning model may subtly erode abstraction boundaries. It may be tempting to re-use input signals
in ways that create unintended tight coupling of otherwise disjoint systems. Machine learning
packages may often be treated as black boxes, resulting in large masses of “glue code” or calibration
layers that can lock in assumptions. Changes in the external world may make models or input
signals change behavior in unintended ways, ratcheting up maintenance cost and the burden of any
debt. Even monitoring that the system as a whole is operating as intended may be difficult without
careful design.

Indeed, a remarkable portion of real-world “machine learning” work is devoted to tackling issues
of this form. Paying down technical debt may initially appear less glamorous than research results
usually reported in academic ML conferences. But it is critical for long-term system health and
enables algorithmic advances and other cutting-edge improvements.'
machine-learning  ml  systems  ops  tech-debt  maintainance  google  papers  hidden-costs  development 
december 2014 by jm
Unsupervised machine learning
aka. "zero-shot learning". ok starting point
machine-learning  zero-shot  unsupervised  algorithms  ml 
may 2014 by jm
SAMOA, an open source platform for mining big data streams
Yahoo!'s streaming machine learning platform, built on Storm, implementing:

As a library, SAMOA contains state-of-the-art implementations of algorithms for distributed machine learning on streams. The first alpha release allows classification and clustering. For classification, we implemented a Vertical Hoeffding Tree (VHT), a distributed streaming version of decision trees tailored for sparse data (e.g., text). For clustering, we included a distributed algorithm based on CluStream. The library also includes meta-algorithms such as bagging.
storm  streaming  big-data  realtime  samoa  yahoo  machine-learning  ml  decision-trees  clustering  bagging  classification 
november 2013 by jm
"Machine Learning That Matters" [paper, PDF]
Great paper. This point particularly resonates: "It is easy to sit in your office and run a Weka algorithm on a data set you downloaded from the web. It is very hard to identify a problem for which machine learning may offer a solution, determine what data should be collected, select or extract relevant features, choose an appropriate learning method, select an evaluation method, interpret the results, involve domain experts, publicize the results to the relevant scientific community, persuade users to adopt the technique, and (only then) to truly have made a difference (see Figure 1). An ML researcher might well feel fatigued or daunted just contemplating this list of activities. However, each one is a necessary component of any research program that seeks to have a real impact on the world outside of machine learning."
machine-learning  ml  software  data  real-world  algorithms 
june 2012 by jm

related tags

abuse  accountability  adaboost  adversarial-classification  ai  airflow  algorithms  amazon  ami  anomaly-detection  apriori  asg  auc  aws  bagging  best-practices  bias  big-data  c4.5  cart  cep  classification  clustering  cmu  coding  control  control-theory  corpus  crime  cs  cucumbers  culture  d3  data  data-analysis  data-mining  data-privacy  data-protection  data-science  data-structures  dataviz  decision-trees  deep-dream  deep-learning  deployment  development  discover-weekly  dnns  dystopia  em  ethics  eu  europe  explainability  facebook  facial-recognition  feature-extraction  federated-learning  future  gamergame  gaming  gboard  gdpr  genetic-algorithms  google  googlex  gwern  harassment  hidden-costs  history  hoeffding-tree  holdout  image-recognition  images  implicit-bias  inequality  infrastructure  interviews  k-means  knn  learning  legibility  lyst  machine-learning  maintainance  media  microsoft  ml  models  music  naive-bayes  named-entities  netflix  neural-networks  news  nlp  nsfw  nvidia  ops  packaging  pagerank  papers  pdf  perceptron  perturbation  pheme  policing  porn  prejudice  privacy  production  race  racism  real-world  realtime  recommendations  recommenders  remediation  reusable-holdout  reward-hacking  right-to-explanation  rollback  samoa  sampling  scary  sebastian-thrun  self-driving-cars  social-media  society  software  spam  spotify  statistics  storm  streaming  streams  supervised  supervised-learning  surveillance  svm  systems  tech-debt  technical-debt  technology  ted-talks  tensorflow  testing  time-series  tokenization  training  truth  twitter  udacity  unsupervised  unsupervised-learning  us-politics  via:daveb  via:maciej  via:nelson  worst-case  yahoo  zero-shot  zeynep-tufekci 

Copy this bookmark:



description:


tags: