## Archive for the ‘Machine Learning’ Category

### MIT Intelligence Quest

Saturday, February 10th, 2018

MIT Intelligence Quest

From the webpage:

The MIT Intelligence Quest will advance the science and engineering of both human and machine intelligence. Launched on February 1, 2018, MIT IQ seeks to discover the foundations of human intelligence and drive the development of technological tools that can positively influence virtually every aspect of society.

The Institute’s culture of collaboration will encourage life scientists, computer scientists, social scientists, and engineers to join forces to investigate the societal implications of their work as they pursue hard problems lying beyond the current horizon of intelligence research. By uniting diverse fields and capitalizing on what they can teach each other, we seek to answer the deepest questions about intelligence.

We are setting out to answer two big questions: How does human intelligence work, in engineering terms? And how can we use that deep grasp of human intelligence to build wiser and more useful machines, to the benefit of society?

Drawing on MIT’s deep strengths and signature values, culture, and history, MIT IQ promises to make important contributions to understanding the nature of intelligence, and to harnessing it to make a better world.

The most refreshing aspect of the MIT Intelligence Quest page is that it ends a contact form.

That’s right, a contact form.

Unlike the ill-fated EU brain project that had pre-chosen approaches and had a roadmap for replicating a human brain. Are they still consuming funds with meetings, hotel rooms, etc.?

You know my mis-givings about creating intelligence in the absence of understanding our own.

On the other hand, mimicking how human intelligence works in bounded situations is a far more tractable problem.

Not too tractable but tractable enough to yield useful results.

### The Matrix Calculus You Need For Deep Learning

Wednesday, February 7th, 2018

Abstract:

This paper is an attempt to explain all the matrix calculus you need in order to understand the training of deep neural networks. We assume no math knowledge beyond what you learned in calculus 1, and provide links to help you refresh the necessary math where needed. Note that you do not need to understand this material before you start learning to train and use deep learning in practice; rather, this material is for those who are already familiar with the basics of neural networks, and wish to deepen their understanding of the underlying math. Don’t worry if you get stuck at some point along the way—just go back and reread the previous section, and try writing down and working through some examples. And if you’re still stuck, we’re happy to answer your questions in the Theory category at forums.fast.ai. Note: There is a reference section at the end of the paper summarizing all the key matrix calculus rules and terminology discussed here.

Here’s a recommendation for reading the paper:

(We teach in University of San Francisco’s MS in Data Science program and have other nefarious projects underway. You might know Terence as the creator of the ANTLR parser generator. For more material, see Jeremy’s fast.ai courses and University of San Francisco’s Data Institute in-person version of the deep learning course.

Apologies to Jeremy but I recognize ANTLR more quickly than I do Jeremy’s fast.ai courses. (Need to fix that.)

The paper runs thirty-three pages and as the authors say, most of it is unnecessary unless you want to understand what’s happening under the hood with deep learning.

Think of it as the difference between knowing how to drive a sports car and being able to work on a sports car.

With the latter set of skills, you can:

• tweak your sports car for maximum performance
• tweak someone else’s sports car for less performance
• detect someone tweaking your sports car

Read the paper, master the paper.

No test, just real world consequences that separate the prepared from the unprepared.

### Finally! A Main Stream Use for Deep Learning!

Tuesday, February 6th, 2018

Using deep learning to generate offensive license plates by Jonathan Nolis.

From the post:

If you’ve been on the internet for long enough you’ve seen quality content generated by deep learning algorithms. This includes algorithms trained on band names, video game titles, and Pokémon. As a data scientist who wants to keep up with modern tends in the field, I figured there would be no better way to learn how to use deep learning myself than to find a fun topic to generate text for. After having the desire to do this, I waited for a year before I found just the right data set to do it,

I happened to stumble on a list of banned license plates in Arizona. This list contains all of the personalized license plates that people requested but were denied by the Arizona Motor Vehicle Division. This dataset contained over 30,000 license plates which makes a great set of text for a deep learning algorithm. I included the data as text in my GitHub repository so other people can use it if they so choose. Unfortunately the data is from 2012, but I have an active Public Records Request to the state of Arizona for an updated list. I highly recommend you look through it, it’s very funny.

What a great idea! Not only are you learning deep learning but you are being offensive at the same time. A double-dipper!

A script for banging against your state license registration is left as an exercise for the reader.

A password generator using phonetics to spell offensive phrases for c-suite users would be nice.

Thursday, January 18th, 2018

From the post:

It’s a major milestone in the push to have search engines such as Bing and intelligent assistants such as Cortana interact with people and provide information in more natural ways, much like people communicate with each other.

A team at Microsoft Research Asia reached the human parity milestone using the Stanford Question Answering Dataset, known among researchers as SQuAD. It’s a machine reading comprehension dataset that is made up of questions about a set of Wikipedia articles.

According to the SQuAD leaderboard, on Jan. 3, Microsoft submitted a model that reached the score of 82.650 on the exact match portion. The human performance on the same set of questions and answers is 82.304. On Jan. 5, researchers with the Chinese e-commerce company Alibaba submitted a score of 82.440, also about the same as a human.

With machine reading comprehension, researchers say computers also would be able to quickly parse through information found in books and documents and provide people with the information they need most in an easily understandable way.

That would let drivers more easily find the answer they need in a dense car manual, saving time and effort in tense or difficult situations.

These tools also could let doctors, lawyers and other experts more quickly get through the drudgery of things like reading through large documents for specific medical findings or rarified legal precedent. The technology would augment their work and leave them with more time to apply the knowledge to focus on treating patients or formulating legal opinions.

Wait, wait! If you read the details about SQuAD, you realize how far Microsoft (or anyone else) is from “…reading through large documents for specific medical findings or rarified legal precedent….”

Stanford Question Answering Dataset (SQuAD) is a new reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles, where the answer to every question is a segment of text, or span, from the corresponding reading passage. With 100,000+ question-answer pairs on 500+ articles, SQuAD is significantly larger than previous reading comprehension datasets.

Not to take anything away from Microsoft Research Asia or the creators of SQuAD, but “…the answer to every question is a segment of text, or span, from the corresponding reading passage.” is a long way from synthesizing an answer from a long legal document.

The first hurdle is asking a question that can be scored against every “…segment of text, or span…” such that a relevant snippet of text can be found.

The second hurdle is the process of scoring snippets of text in order to retrieve the most useful one. That’s a mechanical process, not one that depends on the semantics of the underlying question or text.

There are other hurdles but those two suffice to show there is no “reading and answering questions” in the same sense we would apply to any human reader.

Click-bait headlines don’t serve the cause of advocating more AI research. On the contrary, a close reading of alleged progress leads to disappointment.

### Tutorial on Deep Generative Models (slides and video)

Wednesday, December 27th, 2017

Slides for: Tutorial on Deep Generative Models by Shakir Mohamed and Danilo Rezende.

Abstract:

This tutorial will be a review of recent advances in deep generative models. Generative models have a long history at UAI and recent methods have combined the generality of probabilistic reasoning with the scalability of deep learning to develop learning algorithms that have been applied to a wide variety of problems giving state-of-the-art results in image generation, text-to-speech synthesis, and image captioning, amongst many others. Advances in deep generative models are at the forefront of deep learning research because of the promise they offer for allowing data-efficient learning, and for model-based reinforcement learning. At the end of this tutorial, audience member will have a full understanding of the latest advances in generative modelling covering three of the active types of models: Markov models, latent variable models and implicit models, and how these models can be scaled to high dimensional data. The tutorial will expose many questions that remain in this area, and for which thereremains a great deal of opportunity from members of the UAI community.

Deep sledding on the latest developments in deep generative models (August 2017 presentation) that ends with a bibliography starting on slide 84 of 96.

Depending on how much time has passed since the tutorial, try searching the topics as they are covered, keep a bibliography of your finds and compare it to that of the authors.

Sunday, December 24th, 2017

From the post:

Earlier this month at an undisclosed location in the Middle East, computers using special algorithms helped intelligence analysts identify objects in a video feed from a small ScanEagle drone over the battlefield.

A few days into the trials, the computer identified objects – people, cars, types of building – correctly about 60 percent of the time. Just over a week on the job – and a handful of on-the-fly software updates later – the machine’s accuracy improved to around 80 percent. Next month, when its creators send the technology back to war with more software and hardware updates, they believe it will become even more accurate.

It’s an early win for a small team of just 12 people who started working on the project in April. Over the next year, they plan to expand the project to help automate the analysis of video feeds coming from large drones – and that’s just the beginning.

“What we’re setting the stage for is a future of human-machine teaming,” said Air Force Lt. Gen. John N.T. “Jack” Shanahan, director for defense intelligence for warfighter support, the Pentagon general who is overseeing the effort. Shanahan believes the concept will revolutionize the way the military fights.

So you will recognize Air Force Lt. Gen. John N.T. “Jack” Shanahan (Nvidia conference):

From the Nvidia conference:

Don’t change the culture. Unleash the culture.

That was the message one young officer gave Lt. General John “Jack” Shanahan — the Pentagon’s director for defense for warfighter support — who is hustling to put artificial intelligence and machine learning to work for the U.S. Defense Department.

Highlighting the growing role AI is playing in security, intelligence and defense, Shanahan spoke Wednesday during a keynote address about his team’s use of GPU-driven deep learning at our GPU Technology Conference in Washington.

Shanahan leads Project Maven, an effort launched in April to put machine learning and AI to work, starting with efforts to turn the countless hours of aerial video surveillance collected by the U.S. military into actionable intelligence.

There are at least two market opportunity for adversarial learning. The most obvious one is testing a competitor’s algorithm so it performs less well than yours on “… people, cars, types of building….”

The less obvious market requires US sales of AI-enabled weapon systems to its client states. Client states have an interest in verifying the quality of AI-enabled weapon systems, not to mention non-client states who will be interested in defeating such systems.

For any of those markets, weaponizing adversarial learning and developing a reputation for the same can’t start too soon. Is your anti-AI research department hiring?

### Is it a vehicle? A helicopter? No, it’s a rifle! Messing with Machine Learning

Wednesday, December 20th, 2017

Partial Information Attacks on Real-world AI

From the post:

We’ve developed a query-efficient approach for finding adversarial examples for black-box machine learning classifiers. We can even produce adversarial examples in the partial information black-box setting, where the attacker only gets access to “scores” for a small number of likely classes, as is the case with commercial services such as Google Cloud Vision (GCV).

The post is a quick read (est. 2 minutes) with references but you really need to see:

Abstract:

Current neural network-based image classifiers are susceptible to adversarial examples, even in the black-box setting, where the attacker is limited to query access without access to gradients. Previous methods — substitute networks and coordinate-based finite-difference methods — are either unreliable or query-inefficient, making these methods impractical for certain problems.

We introduce a new method for reliably generating adversarial examples under more restricted, practical black-box threat models. First, we apply natural evolution strategies to perform black-box attacks using two to three orders of magnitude fewer queries than previous methods. Second, we introduce a new algorithm to perform targeted adversarial attacks in the partial-information setting, where the attacker only has access to a limited number of target classes. Using these techniques, we successfully perform the first targeted adversarial attack against a commercially deployed machine learning system, the Google Cloud Vision API, in the partial information setting.

The paper contains this example:

How does it go? Seeing is believing!

Defeating image classifiers will be an exploding market for jewel merchants, bankers, diplomats, and others with reasons to avoid being captured by modern image classification systems.

### Was that Stevie Nicks or Tacotron 2.0? ML Singing in 2018

Tuesday, December 19th, 2017

[S]amim @samim tweeted:

In 2018, machine learning based singing vocal synthesisers will go mainstream. It will transform the music industry beyond recognition.

Abstract:

This paper describes Tacotron 2, a neural network architecture for speech synthesis directly from text. The system is composed of a recurrent sequence-to-sequence feature prediction network that maps character embeddings to mel-scale spectrograms, followed by a modified WaveNet model acting as a vocoder to synthesize timedomain waveforms from those spectrograms. Our model achieves a mean opinion score (MOS) of 4.53 comparable to a MOS of 4.58 for professionally recorded speech. To validate our design choices, we present ablation studies of key components of our system and evaluate the impact of using mel spectrograms as the input to WaveNet instead of linguistic, duration, and F0 features. We further demonstrate that using a compact acoustic intermediate representation enables significant simplification of the WaveNet architecture.

and,

Audio samples from “Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions”

Try the samples before dismissing the prediction of machine learning singing in 2018.

I have a different question:

What is in your test set for ML singing?

Among my top picks, Stevie Nicks, Janis Joplin, and of course, Grace Slick.

### Statistics vs. Machine Learning Dictionary (flat text vs. topic map)

Saturday, December 16th, 2017

Data science terminology (UBC Master of Data Science)

From the webpage:

This document is intended to help students navigate the large amount of jargon, terminology, and acronyms encountered in the MDS program and beyond. There is also an accompanying blog post.

Stat-ML dictionary

This section covers terms that have different meanings in different contexts, specifically statistics vs. machine learning (ML).
… (emphasis in original)

Gasp! You don’t mean that the same words have different meanings in machine learning and statistics!

Even more shocking, some words/acronyms, have the same meaning!

Never fear, a human reader can use this document to distinguish the usages.

Automated processors, not so much.

If these terms were treated as occurrences of topics, where the topics had the respective scopes of statistics and machine-learning, then for any scoped document, an enhanced view with the correct definition for the unsteady reader could be supplied.

Static markup of legacy documents is not required as annotations can be added as a document is streamed to a reader. Opening the potential, of course, for different annotations depending upon the skill and interest of the reader.

If for each term/subject, more properties than the scope of statistics or machine-learning or both were supplied, users of the topic map could search on those properties to match terms not included here. Such as which type of bias (in statistics) does bias mean in your paper? A casually written Wikipedia article reports twelve and with refinement, the number could be higher.

Flat text is far easier to write than a topic map but tasks every reader with re-discovering the distinctions already known to the author of the document.

Imagine your office, department, agency’s vocabulary and its definitions captured and then used to annotate internal or external documentation for your staff.

Instead of very new staffer asking (hopefully), what do we mean by (your common term), the definition appears with a mouse-over in a document.

Are you capturing the soft knowledge of your staff?

### IJCAI – Proceedings 1969-2016 Treasure Trove of AI Papers

Tuesday, December 12th, 2017

IJCAI – Proceedings 1969-2016

International Joint Conferences on Artificial Intelligence is a non-profit corporation founded in California, in 1969 for scientific and educational purposes, including dissemination of information on Artificial Intelligence at conferences in which cutting-edge scientific results are presented and through dissemination of materials presented at these meetings in form of Proceedings, books, video recordings, and other educational materials. IJCAI conferences present premier international gatherings of AI researchers and practitioners. IJCAI conferences were held biennially in odd-numbered years since 1969. They are sponsored jointly by International Joint Conferences on Artificial Intelligence Organization (IJCAI), and the national AI societie(s) of the host nation(s).

While looking for a paper on automatic concept formulation for Jack Park, I found this archive of prior International Joint Conferences on Artificial Intelligence proceedings.

The latest proceedings, 2016, runs six volumes and approximately 4276 pages.

Enjoy!

### CatBoost: Yandex’s machine learning algorithm (here be Russians)

Thursday, December 7th, 2017

From the post:

Russia’s Internet giant Yandex has launched CatBoost, an open source machine learning service. The algorithm has already been integrated by the European Organization for Nuclear Research to analyze data from the Large Hadron Collider, the world’s most sophisticated experimental facility.

Machine learning helps make decisions by analyzing data and can be used in many different areas, including music choice and facial recognition. Yandex, one of Russia’s leading tech companies, has made its advanced machine learning algorithm, CatBoost, available free of charge for developers around the globe.

“This is the first Russian machine learning technology that’s an open source,” said Mikhail Bilenko, Yandex’s head of machine intelligence and research.

I called out the Russian origin of the CatBoost algorithm, not because I have any nationalistic tendencies but you can find frothing paranoids in U.S. government agencies and their familiars who do. In those cases, avoid CatBoost.

If you work in saner environments, or need to use categorical data (read not converted to numbers), give CatBoost a close look!

Enjoy!

### AlphaZero: Mastering Unambiguous, Low-Dimensional Data

Wednesday, December 6th, 2017

Abstract:

The game of chess is the most widely-studied domain in the history of artificial intelligence. The strongest programs are based on a combination of sophisticated search techniques, domain-specific adaptations, and handcrafted evaluation functions that have been refined by human experts over several decades. In contrast, the AlphaGo Zero program recently achieved superhuman performance in the game of Go, by tabula rasa reinforcement learning from games of self-play. In this paper, we generalise this approach into a single AlphaZero algorithm that can achieve, tabula rasa, superhuman performance in many challenging domains. Starting from random play, and given no domain knowledge except the game rules, AlphaZero achieved within 24 hours a superhuman level of play in the games of chess and shogi (Japanese chess) as well as Go, and convincingly defeated a world-champion program in each case.

The achievements by the AlphaZero team and their algorithm merit joyous celebration.

Joyous celebration recognizing AlphaZero masters unambiguous, low-dimensional data governed by deterministic rules that define the outcomes for any state, more quickly and completely than any human.

Chess, Shogi and Go appear complex to humans due to the large number of potential outcomes. But every outcome is the result of the application of deterministic rules to unambiguous, low-dimensional data. Something that AlphaZero excels at doing.

What hasn’t been shown is equivalent performance on ambiguous, high-dimensional data, governed by partially (if that) known rules, for a limited set of sub-cases. For those cases, well, you need a human being.

That’s not to take anything away from the AlphaZero team, but to recognize the strengths of AlphaZero and to avoid its application where it is weak.

### 23 Deep Learning Papers To Get You Started — Part 1 (Reading Slowly)

Saturday, November 25th, 2017

23 Deep Learning Papers To Get You Started — Part 1 by Rupak Kr. Thakur.

Deep Learning has probably been the single-most discussed topic in the academia and industry in recent times. Today, it is no longer exclusive to an elite group of scientists. Its widespread applications warrants that people from all disciplines have an understanding of the underlying concepts, so as to be able to better apply these techniques in their field of work. As a result of which, MOOCs, certifications and bootcamps have flourished. People have generally preferred the hands-on learning experiences. However, there is a considerable population who still give in to the charm of learning the subject the traditional way — through research papers.

Reading research papers can be pretty time-consuming, especially since there are hordes of publications available nowadays, as Andrew Ng said at an AI conference, recently, along with encouraging people to use the existing research output to build truly transformative solutions across industries.

In this series of blog posts, I’ll try to condense the learnings from some really important papers into 15–20 min reads, without missing out on any key formulas or explanations. The blog posts are written, keeping in mind the people, who want to learn basic concepts and applications of deep learning, but can’t spend too much time scouring through the vast literature available. Each part of the blog will broadly cater to a theme and will introduce related key papers, along with suggesting some great papers for additional reading.

In the first part, we’ll explore papers related to CNNs — an important network architecture in deep learning. Let’s get started!

The start of what promises to be a great series on deep learning!

While the posts will extract the concepts and important points of the papers, I suggest you download the papers and map the summaries back to the papers themselves.

It will be good practice on reading original research, not to mention re-enforcing what you have learned from the posts.

In my reading, I will be looking for ways to influence deep learning towards one answer or another.

Whatever they may say about “facts” in public, no sane client asks for advice without an opinion on the range of acceptable answers.

Imagine you found ISIS content on Twitter has no measurable impact on ISIS recruiting. Would any intelligence agency would ask you for deep learning services again?

### Hackers! 90% of Federal IT Managers Aiming for Their Own Feet!

Tuesday, November 14th, 2017

Most Powerful Applications:

• 90% of Feds say AI could help prepare agencies for real-world cyber attack scenarios and 87% say it would improve the efficiency of the Federal cyber security workforce
• 91% say their agency could utilize AI to monitor human activity and deter insider threats, including detecting suspicious elements and large amounts of data being downloaded, and analyzing risky user behavior
• (emphasis in original)

One sure conclusion from this report, 90% of Feds don’t know AIs mistake turtles for rifles, 90% of the time. The adversarial example literature is full of such cases and getting more robust by the day.

The trap federal IT managers have fallen into is a familiar one. To solve an entirely human problem, a shortage of qualified labor, they want mechanize the required task, even if it means a lower qualify end result. Human problems are solved poorly, if at all, by mechanized solutions.

Opposed by lowest common denominator AI systems, hackers will be all but running the mints as cybersecurity AI systems spread across the federal government. “Ghost” federal installations will appear on agency records for confirmation of FedEx/UPS shipments. The possibilities are endless.

If you are a state or local government or even a federal IT manager, letting hackers run wild isn’t a foregone conclusion.

You could pattern your compensation packages after West Coast start-ups, along with similar perks. Expensive but do you want an OMB type data leak on your record?

### Is That a Turtle in Your Pocket or Are You Just Glad To See Me?

Thursday, November 9th, 2017

Apologies to Mae West for spoiling her famous line from Sexette:

Is that a gun in your pocket, or are you just glad to see me?

Seems appropriate since Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok have created a 3-D turtle that is mistaken by neural networks as a rifle.

You can find the details in: Synthesizing Robust Adversarial Examples.

Abstract:

Neural network-based classifiers parallel or exceed human-level accuracy on many common tasks and are used in practical systems. Yet, neural networks are susceptible to adversarial examples, carefully perturbed inputs that cause networks to misbehave in arbitrarily chosen ways. When generated with standard methods, these examples do not consistently fool a classifier in the physical world due to viewpoint shifts, camera noise, and other natural transformations. Adversarial examples generated using standard techniques require complete control over direct input to the classifier, which is impossible in many real-world systems.

We introduce the first method for constructing real-world 3D objects that consistently fool a neural network across a wide distribution of angles and viewpoints. We present a general-purpose algorithm for generating adversarial examples that are robust across any chosen distribution of transformations. We demonstrate its application in two dimensions, producing adversarial images that are robust to noise, distortion, and affine transformation. Finally, we apply the algorithm to produce arbitrary physical 3D-printed adversarial objects, demonstrating that our approach works end-to-end in the real world. Our results show that adversarial examples are a practical concern for real-world systems.

All in good fun until you remember neural networks feed classification decisions to humans who make fire/no fire decisions and soon, fire/no fire decisions will be made by autonomous systems. Errors in classification decisions such as turtle vs. rifle will have deadly results.

What are the stakes in your neural net classification system? How easily can it be fooled by adversaries?

### Metasploit for Machine Learning: Deep-Pwning

Thursday, November 9th, 2017

Metasploit for Machine Learning: Deep-Pwning

From the post:

Deep-pwning is a lightweight framework for experimenting with machine learning models with the goal of evaluating their robustness against a motivated adversary.

Note that deep-pwning in its current state is no where close to maturity or completion. It is meant to be experimented with, expanded upon, and extended by you. Only then can we help it truly become the goto penetration testing toolkit for statistical machine learning models.

### Metasploit for Machine Learning: Background

Researchers have found that it is surprisingly trivial to trick a machine learning model (classifier, clusterer, regressor etc.) into making an objectively wrong decisions. This field of research is called Adversarial Machine Learning. It is not hyperbole to claim that any motivated attacker can bypass any machine learning system, given enough information and time. However, this issue is often overlooked when architects and engineers design and build machine learning systems. The consequences are worrying when these systems are put into use in critical scenarios, such as in the medical, transportation, financial, or security-related fields.

Hence, when one is evaluating the efficacy of applications using machine learning, their malleability in an adversarial setting should be measured alongside the system’s precision and recall.

(emphasis in original)

As motivation for a deep dive into machine learning, looming reliance on machine learning to compensate for a shortage of cybersecurity defender talent is hard to beat. (Why Machine Learning will Boost Cyber Security Defenses amid Talent Shortfall)

Reducing cybersecurity to the level of machine learning is nearly as inviting as use of an older, less secure version of MINIX by Intel. If you are going to take advantage of a Berkeley software license, at least get the best stuff. Yes?

Machine learning is of growing importance, but since classifiers can be fooled into identifying a 3-D turtle as a rifle, it hasn’t reached human levels of robustness.

Or to put that differently, when was the last time you identified a turtle as a rifle?

Turtle vs. rifle is a distinction few of us would miss in language, even without additional properties, as in a topic map. But thinking of their properties or characteristics, maybe a fruitful way to understand why they can be confused.

Or even planning for their confusion and communicating that plan to others.

### Cheap Tracking of Public Officials/Police

Thursday, October 12th, 2017

The use of license plate readers by law enforcement and others is on the rise. Such readers record the location of your license plate at a particular time and place. They also relieve public bodies of large sums of money.

How I replicated an $86 million project in 57 lines of code by Tait Brown details how he used open source software to create a “…good enough…” license plate reader for far less than the ticket price of$86 million.

Brown has an amusing (read unrealistic) good Samaritan scenario for his less expensive/more extensive surveillance system:

While it’s easy to get caught up in the Orwellian nature of an “always on” network of license plate snitchers, there are many positive applications of this technology. Imagine a passive system scanning fellow motorists for an abductors car that automatically alerts authorities and family members to their current location and direction.

The Teslas vehicles are already brimming with cameras and sensors with the ability to receive OTA updates — imagine turning them into a virtual fleet of good samaritans. Ubers and Lyft drivers could also be outfitted with these devices to dramatically increase the coverage area.

Using open source technology and existing components, it seems possible to offer a solution that provides a much higher rate of return — for an investment much less than $86M. The better use of Brown’s less expensive/more extensive surveillance system is tracking police and public official cars. Invite them to the gold fish bowl they have created for all the rest of us. A great public data resource for testing testimony about the presence/absence of police officers at crime scenes, protests, long rides to the police station and public officials consorting with co-conspirators. ACLU calls for government to monitor itself, reflect an unhealthy confidence in governmental integrity. Only a close watch on government by citizens enables governmental integrity. ### Machine Translation and Automated Analysis of Cuneiform Languages Monday, October 2nd, 2017 Machine Translation and Automated Analysis of Cuneiform Languages From the webpage: The MTAAC project develops and applies new computerized methods to translate and analyze the contents of some 67,000 highly standardized administrative documents from southern Mesopotamia (ancient Iraq) from the 21st century BC. Our methodology, which combines machine learning with statistical and neural machine translation technologies, can then be applied to other ancient languages. This methodology, the translations, and the historical, social and economic data extracted from them, will be offered to the public in open access. A recently funded (March 2017) project that strikes a number of resonances with me! “Open access” and cuneiform isn’t an unheard of combination but many remember when access to cuneiform primary materials was a matter of whim and caprice. There are dark pockets where such practices continue but projects like MTAAC are hard on their heels. The use of machine learning and automated analysis have the potential, when all extant cuneiform texts (multiple projects such as this one) are available, to provide a firm basis for grammars, lexicons, translations. Do read: Machine Translation and Automated Analysis of the Sumerian Language by Émilie Pagé-Perron, Maria Sukhareva, Ilya Khait, Christian Chiarcos, for more details about the project. There’s more to data science than taking advantage of sex-starved neurotics with under five second attention spans and twitchy mouse fingers. ### BuzzFeed News Searches For Hidden Spy Planes Monday, August 7th, 2017 BuzzFeed News Trained A Computer To Search For Hidden Spy Planes. This Is What We Found. From the post: Data and R code for the analysis supporting this August 7, 2017 BuzzFeed News post on identifying potential surveillance aircraft. Supporting files are in this GitHub repository. Awesome! This is what data journalism is about! While Musk and others are wringing their hands over AI, BuzzFeed uses machine learning to out government spy planes. How cool is that? So, what are some of the headlines from The New York Times today? 1. Scientists Fear Trump Will Dismiss Climate Change Report 2. What Music Do Americans Love the Most? 50 Detailed Fan Maps 3. Partisan C.I.A. Chief Heartens Trump and Worries the Agency 4. North Korea Warns U.S. of Retaliation Over Sanctions 5. Industries Are Left in the Lurch by Trump’s Stalled Trade Plans 6. White House Won’t Say Who Is on Its Deregulation Teams 7. Wells Fargo Faces New Inquiry Over Insurance Refunds 8. Take the Generic, Patients Are Told. Until They Are Not. 9.$78,000 of Debt for a Harvard Theater Degree
10. Investigators in Israel Turn Up the Heat on Netanyahu

Four out of ten stories are about our accidental president (1, 3, 5, 6) The other six (2, 4, 7, 8, 9, 10), offer no actionable information.

Not a word about government spy planes.

Why isn’t The New York Times pressing the government hard?

Or perhaps the easier question: Why are you still reading The New York Times?

### Overlap – Attacking on Machine Learning Models

Saturday, August 5th, 2017

Abstract:

Deep neural network-based classifiers are known to be vulnerable to adversarial examples that can fool them into misclassifying their input through the addition of small-magnitude perturbations. However, recent studies have demonstrated that such adversarial examples are not very effective in the physical world–they either completely fail to cause misclassification or only work in restricted cases where a relatively complex image is perturbed and printed on paper. In this paper we propose a new attack algorithm–Robust Physical Perturbations (RP2)– that generates perturbations by taking images under different conditions into account. Our algorithm can create spatially-constrained perturbations that mimic vandalism or art to reduce the likelihood of detection by a casual observer. We show that adversarial examples generated by RP2 achieve high success rates under various conditions for real road sign recognition by using an evaluation methodology that captures physical world conditions. We physically realized and evaluated two attacks, one that causes a Stop sign to be misclassified as a Speed Limit sign in 100% of the testing conditions, and one that causes a Right Turn sign to be misclassified as either a Stop or Added Lane sign in 100% of the testing conditions.

I was struck by the image used for this paper in a tweet:

I recognized this as an “overlapping” markup problem before discovering the authors were attacking machine learning models. On overlapping markup, see: Towards the unification of formats for overlapping markup by Paolo Marinelli, Fabio Vitali, Stefano Zacchiroli, or more recently, It’s more than just overlap: Text As Graph – Refining our notion of what text really is—this time for sure! by Ronald Haentjens Dekker and David J. Birnbaum.

From the conclusion:

In this paper, we introduced Robust Physical Perturbations (RP2), an algorithm that generates robust, physically realizable adversarial perturbations. Previous algorithms assume that the inputs of DNNs can be modified digitally to achieve misclassification, but such an assumption is infeasible, as an attacker with control over DNN inputs can simply replace it with an input of his choice. Therefore, adversarial attack algorithms must apply perturbations physically, and in doing so, need to account for new challenges such as a changing viewpoint due to distances, camera angles, different lighting conditions, and occlusion of the sign. Furthermore, fabrication of a perturbation introduces a new source of error due to a limited color gamut in printers.

We use RP2 to create two types of perturbations: subtle perturbations, which are small, undetectable changes to the entire sign, and camouflage perturbations, which are visible perturbations in the shape of graffiti or art. When the Stop sign was overlayed with a print out, subtle perturbations fooled the classifier 100% of the time under different physical conditions. When only the perturbations were added to the sign, the classifier was fooled by camouflage graffiti and art perturbations 66.7% and 100% of the time respectively under different physical conditions. Finally, when an untargeted poster-printed camouflage perturbation was overlayed on a Right Turn sign, the classifier was fooled 100% of the time. In future work, we plan to test our algorithm further by varying some of the other conditions we did not consider in this paper, such as sign occlusion.

Excellent work but my question: Is the inability of the classifier to recognize overlapping images similar to the issues encountered as overlapping markup?

To be sure overlapping markup is in part an artifice of unimaginative XML rules, since overlapping texts are far more common than non-overlapping texts. Especially when talking about critical editions or even differing analysis of the same text.

But beyond syntax, there is the subtlety of treating separate “layers” or stacks of a text as separate and yet tracking the relationship between two or more such stacks, when arbitrary additions or deletions can occur in any of them. Additions and deletions that must be accounted for across all layers/stacks.

I don’t have a solution to offer but pose the question of layers of recognition in hopes that machine learning models can capitalize on the lessons learned about a very similar problem with overlapping markup.

### Neuroscience-Inspired Artificial Intelligence

Saturday, August 5th, 2017

Neuroscience-Inspired Artificial Intelligence by Demis Hassabis, Dharshan Kumaran, Christopher Summerfield, and Matthew Botvinick.

Abstract:

The fields of neuroscience and artificial intelligence (AI) have a long and intertwined history. In more recent times, however, communication and collaboration between the two fields has become less commonplace. In this article, we argue that better understanding biological brains could play a vital role in building intelligent machines. We survey historical interactions between the AI and neuroscience fields and emphasize current advances in AI that have been inspired by the study of neural computation in humans and other animals. We conclude by highlighting shared themes that may be key for advancing future research in both fields.

Extremely rich article with nearly four (4) pages of citations.

Reading this paper closely and chasing the citations is a non-trivial task but you will be prepared understand and/or participate in the next big neuroscience/AI breakthrough.

Enjoy!

### Mistaken Location of Creativity in “Machine Creativity Beats Some Modern Art”

Friday, June 30th, 2017

Machine Creativity Beats Some Modern Art

From the post:

Creativity is one of the great challenges for machine intelligence. There is no shortage of evidence showing how machines can match and even outperform humans in vast areas of endeavor, such as face and object recognition, doodling, image synthesis, language translation, a vast variety of games such as chess and Go, and so on. But when it comes to creativity, the machines lag well behind.

Not through lack of effort. For example, machines have learned to recognize artistic style, separate it from the content of an image, and then apply it to other images. That makes it possible to convert any photograph into the style of Van Gogh’s Starry Night, for instance. But while this and other work provides important insight into the nature of artistic style, it doesn’t count as creativity. So the challenge remains to find ways of exploiting machine intelligence for creative purposes.

Today, we get some insight into progress in this area thanks to the work of Ahmed Elgammal at the Art & AI Laboratory at Rutgers University in New Jersey, along with colleagues at Facebook’s AI labs and elsewhere.
… (emphasis in original)

This summary of CAN: Creative Adversarial Networks, Generating “Art” by Learning About Styles and Deviating from Style Norms by Ahmed Elgammal, Bingchen Liu, Mohamed Elhoseiny, Marian Mazzone, repeats a mistake made by the authors, that is the misplacement of creativity.

Creativity, indeed, even art itself, is easily argued to reside in the viewer (reader) and not the creator at all.

To illustrate, I quote a long passage from Stanley Fish’s How to Recognize a Poem When You See One below but a quick summary/reminder goes like this:

Fish was teaching back to back classes in the same classroom and for the first class, wrote a list of authors on the blackboard. After the first class ended but before the second class, a poetry class, arrived, he enclosed the list of authors in a rectangle and wrote a page number, as though the list was from a book. When the second class arrived, he asked them to interpret the “poem” that was on the board. Which they proceeded to do. Where would you locate creativity in that situation?

The longer and better written start of the story (by Fish):

[1] Last time I sketched out an argument by which meanings are the property neither of fixed and stable texts nor of free and independent readers but of interpretive communities that are responsible both for the shape of a reader’s activities and for the texts those activities produce. In this lecture I propose to extend that argument so as to account not only for the meanings a poem might be said to have but for the fact of its being recognized as a poem in the first place. And once again I would like to begin with an anecdote.

[2] In the summer of 1971 I was teaching two courses under the joint auspices of the Linguistic Institute of America and the English Department of the State University of New York at Buffalo. I taught these courses in the morning and in the same room. At 9:30 I would meet a group of students who were interested in the relationship between linguistics and literary criticism. Our nominal subject was stylistics but our concerns were finally theoretical and extended to the presuppositions and assumptions which underlie both linguistic and literary practice. At 11:00 these students were replaced by another group whose concerns were exclusively literary and were in fact confined to English religious poetry of the seventeenth century. These students had been learning how to identify Christian symbols and how to recognize typological patterns and how to move from the observation of these symbols and patterns to the specification of a poetic intention that was usually didactic or homiletic. On the day I am thinking about, the only connection between the two classes was an assignment given to the first which was still on the blackboard at the beginning of the second. It read:

Jacobs-Rosenbaum
Levin
Thorne
Hayes
Ohman (?)

[3] I am sure that many of you will already have recognized the names on this list, but for the sake of the record, allow me to identify them. Roderick Jacobs and Peter Rosenbaum are two linguists who have coauthored a number of textbooks and coedited a number of anthologies. Samuel Levin is a linguist who was one of the first to apply the operations of transformational grammar to literary texts. J. P. Thorne is a linguist at Edinburgh who, like Levin, was attempting to extend the rules of transformational grammar to the notorious ir-regularities of poetic language. Curtis Hayes is a linguist who was then using transformational grammar in order to establish an objective basis for his intuitive impression that the language of Gibbon’s Decline and Fall of the Roman Empire is more complex than the language of Hemingway’s novels. And Richard Ohmann is the literary critic who, more than any other, was responsible for introducing the vocabulary of transformational grammar to the literary community. Ohmann’s name was spelled as you see it here because I could not remember whether it contained one or two n’s. In other words, the question mark in parenthesis signified nothing more than a faulty memory and a desire on my part to appear scrupulous. The fact that the names appeared in a list that was arranged vertically, and that Levin, Thorne, and Hayes formed a column that was more or less centered in relation to the paired names of Jacobs and Rosenbaum, was similarly accidental and was evidence only of a certain compulsiveness if, indeed, it was evidence of anything at all.

[5] In addition to specifying significances for the words of the poem and relating those significances to one another, the students began to discern larger structural patterns. It was noted that of the six names in the poem three–Jacobs, Rosenbaum, and Levin–are Hebrew, two–Thorne and Hayes–are Christian, and one–Ohman–is ambiguous, the ambiguity being marked in the poem itself (as the phrase goes) by the question mark in parenthesis. This division was seen as a reflection of the basic distinction between the old dis-pensation and the new, the law of sin and the law of love. That distinction, however, is blurred and finally dissolved by the typological perspective which invests the old testament events and heroes with new testament meanings. The structure of the poem, my students concluded, is therefore a double one, establishing and undermining its basic pattern (Hebrew vs. Christian) at the same time. In this context there is finally no pressure to resolve the ambiguity of Ohman since the two possible readings–the name is Hebrew, the name is Christian–are both authorized by the reconciling presence in the poem of Jesus Christ. Finally, I must report that one student took to counting letters and found, to no one’s surprise, that the most prominent letters in the poem were S, O, N.

The account by Fish isn’t long and is highly recommended if you are interested in this issue.

If readers/viewers interpret images as art, is the “creativity” of the process that brought it into being even meaningful? Or does polling of viewers measure their appreciation of an image as art, without regard to the process that created it? Exactly what are we measuring when polling such viewers?

By Fish’s account, such a poll tells us a great deal about the viewers but nothing about the creator of the art.

FYI, that same lesson applies to column headers, metadata keys, and indeed, data itself. Which means the “meaning” of what you wrote may be obvious to you, but not to anyone else.

Topic maps can increase your odds of being understood or discovering the understanding captured by others.

### If You Don’t Think “Working For The Man” Is All That Weird

Saturday, June 17th, 2017

From the post:

Financial services jobs go in and out of fashion. In 2001 equity research for internet companies was all the rage. In 2006, structuring collateralised debt obligations (CDOs) was the thing. In 2010, credit traders were popular. In 2014, compliance professionals were it. In 2017, it’s all about machine learning and big data. If you can get in here, your future in finance will be assured.

J.P. Morgan’s quantitative investing and derivatives strategy team, led Marko Kolanovic and Rajesh T. Krishnamachari, has just issued the most comprehensive report ever on big data and machine learning in financial services.

Titled, ‘Big Data and AI Strategies’ and subheaded, ‘Machine Learning and Alternative Data Approach to Investing’, the report says that machine learning will become crucial to the future functioning of markets. Analysts, portfolio managers, traders and chief investment officers all need to become familiar with machine learning techniques. If they don’t they’ll be left behind: traditional data sources like quarterly earnings and GDP figures will become increasingly irrelevant as managers using newer datasets and methods will be able to predict them in advance and to trade ahead of their release.

At 280 pages, the report is too long to cover in detail, but we’ve pulled out the most salient points for you below.

How important is Sarah’s post and the report by J.P. Morgan?

Let put it this way: Sarah’s post is the first business type post I have saved as a complete webpage so I can clean it up and print without all the clutter. This year. Perhaps last year as well. It’s that important.

Sarah’s post is a quick guide to the languages, talents and tools you will need to start “working for the man.”

It that catches your interest, then Sarah’s post is pure gold.

Enjoy!

PS: I’m still working on a link for the full 280 page report. The switchboard is down for the weekend so I will be following up with J.P. Morgan on Monday next.

### Deep Learning – Dodging The NSA

Monday, May 29th, 2017

Ivanov’s motivation for local deep learning hardware came from monthly AWS bills.

You may suffer from those or be training on data sets you’d rather not share with the NSA.

For whatever reason, follow these detailed descriptions to build your own deep learning box.

Caution: If more than a month or more has lapsed from this post and your starting to build a system, check all the update links. Hardware and prices change rapidly.

Saturday, March 4th, 2017

Chip Huyen, who teaches CS 20SI: “TensorFlow for Deep Learning Research” @Standford, has posted code examples for the class, along with a chatbot, developed for one of the assignments.

A neural chatbot using sequence to sequence model with attentional decoder. This is a fully functional chatbot.

This is based on Google Translate Tensorflow model https://github.com/tensorflow/models/blob/master/tutorials/rnn/translate/

Sequence to sequence model by Cho et al.(2014)

Created by Chip Huyen as the starter code for assignment 3, class CS 20SI: “TensorFlow for Deep Learning Research” cs20si.stanford.edu

The detailed assignment handout and information on training time can be found at http://web.stanford.edu/class/cs20si/assignments/a3.pdf

Dialogue is lacking but this chatbot could be trained to appear to government forces as a live “jihadist” following and conversing with other “jihadists.” Who may themselves be chatbots.

Unlike the expense of pilots for a fleet of drones, a single user could “pilot” a group of chatbots, creating an over-sized impression in cyberspace. The deeper the modeling of human jihadists, the harder it will be to distinguish virtual jihadists.

I say “jihadists” for headline effect. You could create interacting chatbots for right/left wing hate groups, gun owners, churches, etc., in short, anyone seeking to dilute surveillance.

(Unlike the ACLU or EFF, I don’t concede there are any legitimate reasons for government surveillance. The dangers of government surveillance far exceed any possible crime it could prevent. Government surveillance is the question. The answer is NO.)

CS 20SI: Tensorflow for Deep Learning Research

From the webpage:

Tensorflow is a powerful open-source software library for machine learning developed by researchers at Google Brain. It has many pre-built functions to ease the task of building different neural networks. Tensorflow allows distribution of computation across different computers, as well as multiple CPUs and GPUs within a single machine. TensorFlow provides a Python API, as well as a less documented C++ API. For this course, we will be using Python.

This course will cover the fundamentals and contemporary usage of the Tensorflow library for deep learning research. We aim to help students understand the graphical computational model of Tensorflow, explore the functions it has to offer, and learn how to build and structure models best suited for a deep learning project. Through the course, students will use Tensorflow to build models of different complexity, from simple linear/logistic regression to convolutional neural network and recurrent neural networks with LSTM to solve tasks such as word embeddings, translation, optical character recognition. Students will also learn best practices to structure a model and manage research experiments.

Enjoy!

### Meet Fenton (my data crunching machine)

Saturday, February 25th, 2017

From the post:

As you might be aware, I have been experimenting with AWS as a remote GPU-enabled machine for a while, configuring Jupyter Notebook to use it as a backend. It seemed to work fine, although costs did build over time, and I had to always keep in mind to shut it off, alongside with a couple of other limitations. Long story short, around 3 months ago I decided to build my own machine learning rig.

My idea in a nutshell was to build a machine that would only act as a server, being accessible from anywhere to me, always ready to unleash its computational powers on whichever task I’d be working on. Although this setup did take some time to assess, assemble and configure, it has been working flawlessly ever since, and I am very happy with it.

This is the most crucial part. After serious consideration and leveraging the budget I decided to invest into EVGA GeForce GTX 1080 8GB card backed by Nvidia GTX 1080 GPU. It is really snappy (and expensive), and in this particular case it only takes 15 minutes to run — 3 times faster than a g2.2xlarge AWS machine! If you still feel hesitant, think of it this way: the faster your model runs, the more experiments you can carry out over the same period of time.
… (emphasis in original)

Total for this GPU rig? £1562.26

If you are interested in comparing the performance of a Beowulf cluster, see: A Homemade Beowulf Cluster: Part 1, Hardware Assembly and A Homemade Beowulf Cluster: Part 2, Machine Configuration.

Either way, you are going to have enough processing power that your skill and not hardware limits are going to be the limiting factor.

### Aerial Informatics and Robotics Platform [simulator]

Thursday, February 16th, 2017

Aerial Informatics and Robotics Platform (Microsoft)

From the webpage:

Machine learning is becoming an increasingly important artificial intelligence approach to building autonomous and robotic systems. One of the key challenges with machine learning is the need for many samples — the amount of data needed to learn useful behaviors is prohibitively high. In addition, the robotic system is often non-operational during the training phase. This requires debugging to occur in real-world experiments with an unpredictable robot.

The Aerial Informatics and Robotics platform solves for these two problems: the large data needs for training, and the ability to debug in a simulator. It will provide realistic simulation tools for designers and developers to seamlessly generate the copious amounts of training data they need. In addition, the platform leverages recent advances in physics and perception computation to create accurate, real-world simulations. Together, this realism, based on efficiently generated ground truth data, enables the study and execution of complex missions that might be time-consuming and/or risky in the real-world. For example, collisions in a simulator cost virtually nothing, yet provide actionable information for improving the design.

Open source simulator from Microsoft for drones.

How very cool!

Imagine training your drone to search for breaches of the Dakota Access pipeline.

Or how to react when it encounters hostile drones.

Enjoy!

### DeepBach: a Steerable Model for Bach chorales generation

Wednesday, December 14th, 2016

Abstract:

The composition of polyphonic chorale music in the style of J.S Bach has represented a major challenge in automatic music composition over the last decades. The art of Bach chorales composition involves combining four-part harmony with characteristic rhythmic patterns and typical melodic movements to produce musical phrases which begin, evolve and end (cadences) in a harmonious way. To our knowledge, no model so far was able to solve all these problems simultaneously using an agnostic machine-learning approach. This paper introduces DeepBach, a statistical model aimed at modeling polyphonic music and specifically four parts, hymn-like pieces. We claim that, after being trained on the chorale harmonizations by Johann Sebastian Bach, our model is capable of generating highly convincing chorales in the style of Bach. We evaluate how indistinguishable our generated chorales are from existing Bach chorales with a listening test. The results corroborate our claim. A key strength of DeepBach is that it is agnostic and flexible. Users can constrain the generation by imposing some notes, rhythms or cadences in the generated score. This allows users to reharmonize user-defined melodies. DeepBach’s generation is fast, making it usable for interactive music composition applications. Several generation examples are provided and discussed from a musical point of view.

Take this with you on January 20, 2017 in case you tire of playing #DisruptJ20 Twitter Game (guessing XQuery/XPath definitions). Unlikely I know but anything can happen.

Deeply impressive work.

You can hear samples at:

http://www.flow-machines.com/deepbach-steerable-model-bach-choralesgeneration/

https://github.com/SonyCSL-Paris/DeepBach

Makes me curious about the composition of “like” works for composers who left smaller corpora.

### Comparing Symbolic Deep Learning Frameworks

Thursday, December 8th, 2016

Deep Learning Part 1: Comparison of Symbolic Deep Learning Frameworks by Anusua Trivedi.

From the post:

This blog series is based on my upcoming talk on re-usability of Deep Learning Models at the Hadoop+Strata World Conference in Singapore. This blog series will be in several parts – where I describe my experiences and go deep into the reasons behind my choices.

Deep learning is an emerging field of research, which has its application across multiple domains. I try to show how transfer learning and fine tuning strategy leads to re-usability of the same Convolution Neural Network model in different disjoint domains. Application of this model across various different domains brings value to using this fine-tuned model.

In this blog (Part1), I describe and compare the commonly used open-source deep learning frameworks. I dive deep into different pros and cons for each framework, and discuss why I chose Theano for my work.

Your mileage may vary but a great starting place!

### Four Experiments in Handwriting with a Neural Network

Tuesday, December 6th, 2016

While the handwriting experiments are compelling and entertaining, the author’s have a more profound goal for this activity:

The black box reputation of machine learning models is well deserved, but we believe part of that reputation has been born from the programming context into which they have been locked into. The experience of having an easily inspectable model available in the same programming context as the interactive visualization environment (here, javascript) proved to be very productive for prototyping and exploring new ideas for this post.

As we are able to move them more and more into the same programming context that user interface work is done, we believe we will see richer modes of human-ai interactions flourish. This could have a marked impact on debugging and building models, for sure, but also in how the models are used. Machine learning research typically seeks to mimic and substitute humans, and increasingly it’s able to. What seems less explored is using machine learning to augment humans. This sort of complicated human-machine interaction is best explored when the full capabilities of the model are available in the user interface context.

Setting up a search alert for future work from these authors!