217 karma · joined October 13, 2016
White collar crime, like fraud and insider trading, sees as much automated detection software as street crime does. But even if it did not, let's assume we focus our attention on street crime, then your reasoning is a fallacy of relative privation: Y is a worse problem than X, so attempting to solve X is bad/unimportant/makes solving Y less likely.
As to the degree this is "encoding racial bias", I feel this is passing the buck to the computer. The computer is merely calculating a probability distribution on historical data the best it possibly can (and likely, more accurate, effective, and with more objectivity than humans ever can). Impartiality or partiality does not start or end with a computer program, it starts with human policies.
One does not solve police racism by creating a model that adheres to your view of an ideal world. This argumentation is naive. It does not change anything, but perhaps makes you feel good. You either solve it at its core, or you don't solve it at all.
One should also distinguish between heavy-crime areas and police bias. Some areas, where a certain race may be overrepresented, are just more prone to street crimes. This is not bias, but fact. Don't cripple your model, just because you don't agree with the facts.
As for white collar crime vs. street crime; automated systems to detect fraud, insider trading, scams, and theft, are also in place. These just don't benefit from capacity planning / more police presence on the streets.
The Minority Report angle is journalistic nonsense / click-bait, akin to invoking Terminator when talking about Google's AI efforts. This is about areas of crime, not pre-crime for individuals.
These systems can adapt. If an area sees a drop in crime rate, then that area requires less police presence. It is not just trained on a single snapshot.
Finally, such Crime Prediction software can be used by human police officers to make better decisions. It is far off from self-driving police cars without any human input.
Note, I am not saying algorithmic bias does not exist, or that this is not an important issue to solve, as we start automating more processes in society. But then, let's discuss it on fair and equal ground, and leave human bias against "human-biased" systems out. We can talk about this mathematically now, or at least, in a non-emotional rational manner.
Aside: Other comments here are very America-centric, and detract from the topic at hand. It's like posting news articles on manual SWAT-raids gone wrong in NY, when discussing predictive policing software in California.
Agreed on the clever marketing, but at least they put their money (expensive dataset) where their mouth is (release it to reverse engineers the world over).
Fully Homomorphic Encryption challenges would be interesting, but it would disqualify our current state-of-the-art algorithms, and reduce the playing field to a handful of people who know how to write algo's that work with Fully Homomorphic Encryption (if any competitor at all is allowed to work on this, and not too busy working for the NSA).
And simple linear transforms on already anonymized features are not so easy to reverse engineer as you may think. Just try it on a few datasets from UCI.
"""
https://arxiv.org/abs/1508.06574
"An encryption scheme is said to be homomorphic
if certain mathematical operations can be applied
directly to the cipher text in such a way that
decrypting the result renders the same answer as
applying the function to the original unencrypted
data."
The function = GradientBoostingRegressor
the cipher text = X_encrypted
original data = X
same answer = mean absolute error
"""
import numpy as np
from sklearn.metrics import mean_absolute_error
from sklearn.ensemble import GradientBoostingRegressor
# Replicability
np.random.seed(0)
# Create a data set with 1000 samples and 3 features
X = np.random.randint(0, 60, (1000,3))
# Create ground truth (the product of the three
# features - 100) / 11
y = (np.prod(X, axis=1) - 100) / 11.
# Encrypt y
y_encrypted = y + 20
# Encrypt X
X_encrypted = X * -0.5
# Init our model
rgr = GradientBoostingRegressor(random_state=42)
# Fit model on first 500 unencrypted features
rgr.fit(X[:500], y[:500])
# Predict the remaining 500 features
preds = rgr.predict(X[500:])
# Fit model on first 500 encrypted features
rgr.fit(X_encrypted[:500], y[:500])
# Predict the remaining encrypted features and decrypt
preds_decrypted = rgr.predict(X_encrypted[500:]) - 20
# Evaluate both functions
print(mean_absolute_error(preds, y[500:]))
print(mean_absolute_error(preds_decrypted, y[500:]))
#>>> 323.09
#>>> 323.72I bet they show potential investors their returns. But to the public? That may be both iffy and unnecessary.
Kaggle competitions also see this. https://www.boozallen.com/content/dam/boozallen/documents/20... [page 93] shows a graph where a simple average of the top models in a competition gives an ensemble model with higher accuracy than any of the individual competitors.
I think that is largely the beauty of Numerai. Using adversarial agents to build a "collaborative" model. Disclaimer: Won a few bitcoin on Numerai.
- Numerai does not have the ability itself, it harvests predictions from many others.
- To assume that better than random guessing is not possible, is to assume that the stock market is operating at maximum efficiency, and that all the hedge funds that are profitable have very lucky random number generators, instead of well-paid quants.
- The co-founder of perhaps the biggest money making hedge fund, Renaissance Technologies, has invested in Numerai. http://www.wsj.com/articles/renaissance-technologies-hedge-f...
It's a bit like saying you don't trust any SEO company that does not rank #1 for SEO. Fair, sure. But the reality is a bit more grey. Disclosure: Won a few bitcoin on Numerai.
Caruana showed the cartoon of the difference between a statistician and a machine learning practitioner by showing a cliff. The statistician carefully inches to the edge, stomping her feet to see if the ground is still stable, then 10 meters before the edge she stops and draws her conclusions. The machine learning practitioner dives headfirst from the cliff, with a parachute that reads "cross-validation".
See also:
http://norvig.com/chomsky.html On Chomsky and the Two Cultures of Statistical Learning.
And http://projecteuclid.org/euclid.ss/1009213726 Statistical Modeling: The Two Cultures by Leo Breiman.
and this joke:
> Norvig teamed up with a Stanford statistician to prove that statisticians, data scientists and mathematicians think the same way. They hypothesized that, if they all received the same dataset, worked on it, and came back together, they’d find they all independently used the same techniques. So, they got a very large dataset and shared it between them.
> Norvig used the whole dataset and built a complex predictive model. The statistician took a 1% sample of the dataset, discarded the rest, and showed that the data met certain assumptions.
> The mathematician, believe it or not, didn’t even look at the dataset. Rather, he proved the characteristics of various formulas that could (in theory) be applied to the data.
2. This platform will provide a test bed for AI algo's and help "democratize" AI. One does not need to set up their own platform. One can compare approaches. One can learn from other implementations running on a common ground. Being resource constrained forces one to be more creative and this paves the way for more energy-friendly methods. Sure, a high school student will not dominate the power houses like DeepMind et al. But the high school student can get up and running in a week or few days.
3. https://arxiv.org/abs/1608.08225 Physics and Deep Learning are well entwined. Deep Learning certainly is a big idea, up there with VC theory and boosting. It exists for decades now, I agree the more recent incarnation was made possible with more computing power and better bigger datasets, and relies less on new tricks. Yet, still tricks are being invented in the recent years that have majorly contributed to better generalization. Dropout being one. DL, and especially the relevant Deep Reinforcement Learning, is not just GPU's, but a lot of new (and budding) theory. One can run Random Forests (and other approaches) on CUDA too. The Neural Turing Machine heralded a whole new, intellectually deep and stimulating, field in Deep Learning, and we haven't seen the best of it yet. There are also fields, like vision, where the other approaches are significantly underperforming relative to DL. Try to train a SVM or RF on ImageNet. Also, one is not required to use DL for your agent. Experiment with the classic approaches and see which is better (https://arxiv.org/abs/1603.04119).
I actually shared some of your concern, not for AI research, but for game development. I thought it was very hard to actually get a good job in that field. Then the mobile game market started booming, and indy developers could make a living. AI has got the backing of all the major players in industry. Instead of pipe dreams and philosophical meanderings, we have actually business-value adding working models now. It is not going anywhere soon.
If all else fails, you remain a good coder or data analyst with a lot of automation skills.
What would you propose as an alternative? If nothing, fine, but how can we relate, when we only have a single best (or if you prefer: flawed) thing?
> It can't handle modelling itself
It can add its internal states to the environment, and hence model these internal states.
> the large dimensionality of its hypothesis space
Its bounded by how many compressors / programs are available to it. Calculating the length of these programs that are consistent with the environment is feasible.
> handle imprecise stimulus information
I'm not sure if you mean imprecise stimuli here or imprecise sensors.
> can be made arbitrarily "stupid"
This can be seen as a flaw, or as a simple property (or even a feature). Something that can be optimal, and "stupid", does not detract much from its ability to be optimal.
> before we get to mere computational limitations or the uncomputability issue
Just because it is uncomputable, does not mean the theory is flawed. Sure, it is not practical, and we like practical things, but it is still valuable to have such a theory. Especially when approximations do yield practical applications.
Marcus Hutter on these issues: http://hunch.net/?cat=14
The OP just has not heard of any accomplishments, but anyone with a little expertise in deep (reinforcement) learning knows about the major contributions to the field by Schmidhuber.
Using this criterion you are using popular media and fields you know not much about, to brush away the accomplishments of respectable scientists. Don't base your skepticism on your own lack of knowledge: that makes it selective -- You can not cut through the bullshit, if you don't know how to wield a sword.
Those are not mutually exclusive. If you want to actually claim that Chris McKinstry was only insane, and his project without any scientific merit, you should make the case for it. Brushing projects away, only because the conceiver of them was mentally ill, is not nice at all.
https://twitter.com/ClintonGlobal/status/651606553884258304
(A Unilever and Clinton Foundation cooperation found in https://wikileaks.org/podesta-emails/emailid/1596)
and some older abuse charges brought up against people connected with Unilever.
See also: https://webcache.googleusercontent.com/search?q=cache:AfaAe2...
Also consider the possibility that the pedo symbology was inspired by the ice cream logo (and not the other way around).
But most likely just a (meaningful) coincidence. Hearts and spirals are common design elements after all.
“hotdog” = boy
“pizza” = girl
“cheese” = little girl
“pasta” = little boy
“ice cream” = male prostitute
“walnut” = person of colour
“map” = semen
“sauce” = orgy
The FBI document detailing the symbols can be found here: https://wikileaks.org/wiki/FBI_pedophile_symbols
The similarities were rather striking. One pizza place even changed their logo shortly after the comparisons were drawn. This only added more fuel to the fire. http://i.imgur.com/GIRxh8r.jpg
Myself being a huge fan of conspiracy theory I've seen this before: the minds become focussed on finding evidence, discarding Occam's razor, and mining for all the nuggets. What remains are coincidences that appear too good to be random. Compare with the "backwards playing record hidden messages" hype. Once someone points it out to you, your mind starts hearing it too.
P.S.: The "hashes" found in the .jpg images were a dead end / the result of pareidolia.
> “I’ve known Jeff for fifteen years. Terrific guy,” Trump told New York Magazine for a 2002 profile of Epstein. “He’s a lot of fun to be with. It is even said that he likes beautiful women as much as I do, and many of them are on the younger side. No doubt about it — Jeffrey enjoys his social life.”
http://dailycaller.com/2016/10/09/the-friendship-between-tru...
Going meta-conspiracy. Perhaps the pizzagate stuff was a psyops pushed to detract from other, more relevant, stories.
The state actors responsible for this want the US to know that they have these records. That they can attack/disturb/flood the personal computers, social networks, and family of these soldiers / military assets.
Also note the apparent warning in the Shadow Brokers auction:
> You see what cryptolockers and stuxnet can do. You see free files we give for free. You see attacks on banks and SWIFT in news. Maybe there is Equation Group version of cryptolocker+stuxnet for banks and financial systems? If Equation Group lose control of cyber weapons, who else lose or find cyber weapons? If electronic data go bye bye where leave Wealthy Elites? Maybe with dumb cattle?
> Post CP is an acronym in which CP can have a variety of meanings. Initially used as an acronym for 4chan posters to indirectly mention or request a child pornography thread, in later years the acronym became commonly associated with the posting of other subjects with the same first letters (C & P, most notably “Cheese Pizza”) to parody prohibited request threads or to troll those who hope to find child pornography.
We Are Anonymous - Inside the Hacker World of LulzSec, Anonymous and the Global Cyber Insurgency (2013) Parmy Olson
> There was also "u jelly?," a way of asking if someone was jealous, and "cheese pizza," or "CP", slang for child porn. More shrewd 4chan users would start discussion threads about literal cheese pizza, including photos of pizzas, and add hidden links to a child porn archive within the image code -- accessed by opening the pizza images in a text program instead of an image viewer.
If a computer program spots an irrelevant unhelpful correlation its generalization error will noticeable go up. If it is an irrelevant "helpful" correlation it means there is a problem with the data (such as leakage), not with the algorithm. If there is a problem with the data, all bets are of, both for black and white box models.
A blackbox model will probably not find that being an Armenian alone will lead to more crashes. Being non-linear in nature it will find interactions (young male Armenians are more likely to crash than young males in general). If the importance of such a feature is not significant enough to distinguish it from noise, then regularization may automatically remove it.
Even if we observe that 99 out 100 Armenians crash their cars, and you decline someone a loan, because he/she is Armenian, you may just have discriminated against the 1 Armenian who is a safe driver. Young male drivers who drive safely have a worse time getting loans, because their group (the set of young male drivers) spoiled it for them. So their only hope of getting a loan is you adding more features (like nationality), to be able to distinguish them as safe drivers, not removing them and lumping them into the status quo.
Only if the cost * probability of a PR disaster outweighs the cost of a poor credit risk model, is a bank economically incentivized to not discriminate.
Then again, even if you take care to not directly discriminate, you will probably indirectly discriminate. For instance, in your credit risk model, remove the `gender` column from the features, and use the remaining features to try to predict it. If performance is better than random guessing, you are using proxy features (like `income`) to discriminate on `gender`. You will find that nearly every feature you use is correlated with `race`. Now what? Throw away all these features and let your competition eat your lunch?
Along the lines of what Pedro Domingos said [2], you can not solve the problem of discrimination by making poorer performing machine learning models that adhere to your view of what is ideal. Discrimination won't disappear because you made a model that makes you feel good. Want no discrimination of women? Work on closing the wage gap. Don't cripple your statistically correct ML models or sweep the discrimination under the rug, covered by correlated variables.
It is not so much carelessness, as it is the nature of the beast. And banks remain in business by how much they can trust their customers first and foremost, trustworthiness by customers is a second (and customer trust is very much malleable: It is the perception of trust, not objective trust like "can we trust this customer to pay back their loan").
It also depends on how you (mathematically) define "fairness" [3]. You can define fairness in ways that still allow you to discriminate.
[1] A new car built by my company leaves somewhere traveling at 60 mph. The rear differential locks up. The car crashes and burns with everyone trapped inside. Now, should we initiate a recall? Take the number of vehicles in the field, A, multiply by the probable rate of failure, B, multiply by the average out-of-court settlement, C. A times B times C equals X. If X is less than the cost of a recall, we don't do one.
[2] https://www.youtube.com/watch?v=furfdqtdAvc
[3] https://algorithmicfairness.wordpress.com/2016/09/26/on-the-... https://arxiv.org/abs/1609.07236
> Testosterone plays a significant role in the arousal of these behavioral manifestations in the brain centers involved in aggression and on the development of the muscular system that enables their realization. There is evidence that testosterone levels are higher in individuals with aggressive behavior, such as prisoners who have committed violent crimes.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3693622/
> Inmates who had committed personal crimes of sex and violence had higher testosterone levels than inmates who had committed property crimes of burglary, theft, and drugs. Inmates with higher testosterone levels also violated more rules in prison, especially rules involving overt confrontation.
http://www.sciencedirect.com/science/article/pii/01918869940...
Though the connection can be said to be weak, and definitely not the only factor (high testosterone alone is not sufficient for criminal offending), it is there.
> In this case, the features they are finding don't seem to make any sense. A slight smile in the criminals seems more likely to be due to the way that set of photos are taken
From the paper: "We stress that the criminal face images in Sc are normal ID photos not police mugshots."
> by the fact the criminal set came from a single police department (in a single geographical area)
Subset Sc contains ID photos of 730 criminals, of which 330 are published as wanted suspects by the ministry of public security of China and by the departments of public security for the provinces of Guangdong, Jiangsu, Liaoning, etc.; the others are provided by a city police department in China under a confidentiality agreement.
> if it included a single "family"-gang of criminals it is likely that would have been enough to taint the features.
Family resemblance is an interesting one. But unlikely to significantly affect the accuracy difference between proper labeling and random labeling (they'd all need to be related)
Overfit is sufficiently ruled out (to me), but leakage is not. Unfortunately it is not possible to replicate this study (even if the dataset was available, the implementation details are scarce). Differently sized raw ID pictures, or compression artifacts, could lead to near undetectable leakage for outsiders. I would probably not give this paper my stamp of approval, even if it was on an uncontroversial subject, but it is not abysmally bad.
I do think one has to be careful to separate moral concerns from technical concerns. Sure, this all feels very wrong to me, and should be taken into account when creating new regulation for ML systems, but the research itself (apart from the small sample size, and vague data gathering methods) is sufficiently solid for debate. Maybe we don't want to admit that phrenology can have a measurable impact on behavior, but that is wishful thinking, not science. Like you said: 'a link between increased testosterone and different facial features' exists, and I just sourced you that a link between criminal behavior and testosterone exists. Logic would deem us to conclude that different facial features are indicative of different criminal behavior, no matter the bizarre, scary, immoral research that supports it.
Yes, we have to try a load of stuff, especially in deep learning, where the old feature engineering seems to be replaced with architecture engineering [1].
Yet, the stuff to try is usually well-known [2]. You should not be trying stuff completely at random (unless you do a random gridsearch for hyperparams [3]).
> we chose 3 layers of 10, 15, and 11 nodes, instead of 3 layers of 10, 15 and 12 nodes!
Hinton advises to start with a wide net, ensure that the net is able to learn the function, then add dropout, smaller layers, and/or regularization to reduce the overfit. This should avoid scenario's like the above.
Wanting to have the perfect model for a specific problem mathematically calculated may be an unrealistic demand. If we view machine learning as compression (using the shortest program possible to describe a function/distribution), per Kolmogorov Complexity, we can't compute this shortest program, and if we happen to find it, we can't know we did. The maths is for theory. Intuition is for the practical application. See intuition more as guided guessing, instead of random guessing: with experience you should know what are the flaws and benefits of the different models and architectures for particular problems. You can limit your search to sane narrow ranges.
[1] https://smerity.com/articles/2016/architectures_are_the_new_...
[2] Finally, suppose you want to train an LDNN. Rumor has it that it’s very difficult to do so, that it is “black magic” that requires years of experience. And while it is true that experience helps quite a bit, the amount of “trickery” is surprisingly limited ---- one needs be on the lookout for only a small number well-known pitfalls. - http://yyue.blogspot.com.br/2015/01/a-brief-overview-of-deep...
[3] http://www.jmlr.org/papers/v13/bergstra12a.html https://people.eecs.berkeley.edu/~kjamieson/hyperband.html
I worked in SEO before, which had far more elements of "black magic". Perhaps SEO helps with the transition to ML, because you are basically reverse engineering a model (Google's search engine) / crafting input to get a higher ranked output. It's feature engineering, experimentation, and debugging all-in-one.
And front-end development of the old days... debugging old javascript or IE6 render bugs makes ML debugging pale in comparison. You had to make a broken model work, without being able to repair it.
As for the long debugging cycles in ML. John Langford coined "sub-linear debugging": Output enough intermediate information to quickly know if you introduced a major bug or hit upon a significant improvement [1]. Machine learning competitions are not so much won by skill, but by the teams iterating faster and more efficiently: Those who try more (failed) experiments hit upon more successful experiments. No Neural Net researcher should let all nets finish training, before drawing conclusions/estimates on learning process.
Sure, the ML field is relatively new, and computer programming has a longer history of proper debugging and testing. It is difficult to do monitoring on feedback-looped models running in production, yet no more difficult than control theory ;). And proper practices are being developed as we speak [2]. The author will probably write a randomization script to avoid malordered samples automatically in the future.
[1] http://www.machinedlearnings.com/2013/06/productivity-is-abo...