A new theory suggests that dreams' illogical logic has an important purpose.
Overfitting<p>The goal of machine learning is to supply an algorithm with a data set, a "training set," in which patterns can be recognized and from which predictions that apply to other unseen data sets can be derived.</p><p>If machine learning learns its training set too well, it merely spits out a prediction that precisely — and uselessly — matches that data instead of underlying patterns within it that could serve as predictions likely to be true of other thus-far unseen data. In such a case, the algorithm describes what the data set <em>is</em> rather than what it <em>means</em>. This is called "overfitting."</p><img type="lazy-image" data-runner-src="https://assets.rebelmouse.io/eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpbWFnZSI6Imh0dHBzOi8vYXNzZXRzLnJibC5tcy8yNDc4NTQ4Ni9vcmlnaW4uanBnIiwiZXhwaXJlc19hdCI6MTY2NDM4NDk1Mn0.bMHbBbt7Nz0vmmQ8fdBKaO-Ycpme5eOCxbjPLEHq9XQ/img.jpg?width=980" id="5049a" class="rm-shortcode" data-rm-shortcode-id="f9a6823125e01f4d69ce13d1eef84486" data-rm-shortcode-name="rebelmouse-image" />
The value of noise<p>To keep machine learning from becoming too fixated on the specific data points in the set being analyzed, programmers may introduce extra, unrelated data as noise or corrupted inputs that are less self-similar than the real data being analyzed.</p><p>This noise typically has nothing to do with the project at hand. It's there, metaphorically speaking, to "distract" and even confuse the algorithm, forcing it to step back a bit to a vantage point at which patterns in the data may be more readily perceived and not drawn from the specific details within the data set.</p><p>Unfortunately, overfitting also occurs a lot in the real world as people race to draw conclusions from insufficient data points — xkcd has a fun example of how this can happen with <a href="https://xkcd.com/1122/" target="_blank">election "facts."</a></p><p>(In machine learning, there's also "underfitting," where an algorithm is too simple to track enough aspects of the data set to glean its patterns.)</p><img type="lazy-image" data-runner-src="https://assets.rebelmouse.io/eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpbWFnZSI6Imh0dHBzOi8vYXNzZXRzLnJibC5tcy8yNDc4NTQ5My9vcmlnaW4uanBnIiwiZXhwaXJlc19hdCI6MTYyMDE5NjY1M30.iS2bq7WEQLeS34zNFPnXwzAZZn9blCyI-KVuXmcHI6o/img.jpg?width=980" id="cd486" class="rm-shortcode" data-rm-shortcode-id="c49cfbbffceb00e3f37f00e0fef859d9" data-rm-shortcode-name="rebelmouse-image" />
Credit: agsandrew/Adobe Stock
Nightly noise<p>There remains a lot we don't know about how much storage space our noggins contain. However, it's obvious that if the brain remembered absolutely everything we experienced in every detail, that would be an awful lot to remember. So it seems the brain consolidates experiences as we dream. To do this, it must make sense of them. It must have a system for figuring out what's important enough to remember and what's unimportant enough to forget rather than just dumping the whole thing into our long-term memory.</p><p>Performing such a wholesale dump would be an awful lot like overfitting: simply documenting what we've experienced without sorting through it to ascertain its meaning.</p><p>This is where the new theory, the <a href="https://arxiv.org/pdf/2007.09560.pdf" target="_blank">Overfitting Brain Hypothesis</a> (OBH) proposed by Erik Hoel of Tufts University, comes in. Suggesting that perhaps the brain's sleeping analysis of experiences is akin to machine learning, he proposes that the illogical narratives in dreams are the biological equivalent of the noise programmers inject into algorithms to keep them from overfitting their data. He says that this may supply just enough off-pattern nonsense to force our brains to see the forest and not the trees in our daily data, our experiences.</p><p>Our experiences, of course, are delivered to us as sensory input, so Hoel suggests that dreams are sensory-input noise, biologically-realistic noise injection with a narrative twist:</p><p style="margin-left: 20px;">"Specifically, there is good evidence that dreams are based on the stochastic percolation of signals through the hierarchical structure of the cortex, activating the default-mode network. Note that there is growing evidence that most of these signals originate in a top-down manner, meaning that the 'corrupted inputs' will bear statistical similarities to the models and representations of the brain. In other words, they are derived from a stochastic exploration of the hierarchical structure of the brain. This leads to the kind structured hallucinations that are common during dreams."</p><p>Put plainly, our dreams are just realistic enough to engross us and carry us along, but are just different enough from our experiences —our "training set" — to effectively serve as noise.</p><p>It's an interesting theory.</p><p>Obviously, we don't know the extent to which our biological mental process actually resemble the comparatively simpler, man-made machine learning. Still, the OBH is worth thinking about, maybe at least more worth thinking about than whatever <em>that</em> was last night.</p>
Textual analysis of social media posts finds users' anxiety and suicide-risk levels are rising, among other negative trends.
A study finds 1.8 billion trees and shrubs in the Sahara desert.
- AI analysis of satellite images sees trees and shrubs where human eyes can't.
- At the western edge of the Sahara is more significant vegetation than previously suspected.
- Machine learning trained to recognize trees completed the detailed study in hours.
Why this matters<img type="lazy-image" data-runner-src="https://assets.rebelmouse.io/eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpbWFnZSI6Imh0dHBzOi8vYXNzZXRzLnJibC5tcy8yNDU2MDQ1OC9vcmlnaW4uanBnIiwiZXhwaXJlc19hdCI6MTYzOTkyODg5NX0.O3S2DRTyAxh-JZqxGKj9KkC6ndZAloEh4hKhpcyeFDQ/img.jpg?width=980" id="3770d" class="rm-shortcode" data-rm-shortcode-id="3c27b79d4c0600fb6ebb82e650cabec0" data-rm-shortcode-name="rebelmouse-image" />
Area in which trees were located
Credit: University of Copenhagen<p>As important as trees are in fighting climate change, scientists need to know what trees there are, and where, and the study's finding represents a significant addition to the global tree inventory.</p><p>The vegetation Brandt and his colleagues have identified is in the Western Sahara, a region of about 1.3 million square kilometers that includes the desert, <a href="https://en.wikipedia.org/wiki/Sahel" target="_blank">the Sahel</a>, and the <a href="https://www.sciencedirect.com/topics/agricultural-and-biological-sciences/subhumid-zones" target="_blank" rel="noopener noreferrer">sub-humid zones</a> of West Africa.</p><p>These trees and shrubs have been left out of previous tabulations of carbon-processing worldwide forests. Says Brandt, "Trees outside of forested areas are usually not included in climate models, and we know very little about their carbon stocks. They are basically a white spot on maps and an unknown component in the global carbon cycle."</p><p>In addition to being valuable climate-change information, the research can help facilitate strategic development of the region in which the vegetation grows due to a greater understanding of local ecosystems.</p>
Trained for trees<img type="lazy-image" data-runner-src="https://assets.rebelmouse.io/eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpbWFnZSI6Imh0dHBzOi8vYXNzZXRzLnJibC5tcy8yNDU2MDQ3MC9vcmlnaW4uanBnIiwiZXhwaXJlc19hdCI6MTYzNTk5NTI3NH0.fR-n1I2DHBIRPLvXv4g0PVM8ciZwSLWorBUUw2wc-Vk/img.jpg?width=980" id="e02c0" class="rm-shortcode" data-rm-shortcode-id="79955b13661dca8b6e19007935129af1" data-rm-shortcode-name="rebelmouse-image" />
Credit: Martin Brandt/University of Copenhagen<p>There's been an assumption that there's hardly enough vegetation outside of forested areas to be worth counting in areas such as this one. As a result the study represents the first time a significant number of trees — likely in the hundreds of millions when shrubs are subtracted from the overall figure — have been catalogued in the drylands region.</p><p>Members of the university's Department of Computer Science trained a machine-learning module to recognize trees by feeding it thousands of pictures of them. This training left the AI be capable of spotting trees in the tiny details of satellite images supplied by NASA. The task took the AI just hours — it would take a human years to perform an equivalent analysis.</p><p>"This technology has enormous potential when it comes to documenting changes on a global scale and ultimately, in contributing towards global climate goals," says co-author Christian Igel. "It is a motivation for us to develop this type of beneficial artificial intelligence."</p><p>"Indeed," says Brandt says, "I think it marks the beginning of a new scientific era."</p>
Looking ahead and beyond<p>The researchers hope to further refine their AI to provide a more detailed accounting of the trees it identifies in satellite photos.</p><p>The study's senior author, Rasmus Fensholt, says, "we are also interested in using satellites to determine tree species, as tree types are significant in relation to their value to local populations who use wood resources as part of their livelihoods. Trees and their fruit are consumed by both livestock and humans, and when preserved in the fields, trees have a positive effect on crop yields because they improve the balance of water and nutrients."</p><p>Ahead is an expansion of the team's tree hunt to a larger area of Africa, with the long-term goal being the creation of a more comprehensive and accurate global database of trees that grow beyond the boundaries of forests.</p>
Can we stop a rogue AI by teaching it ethics? That might be easier said than done.
- One way we might prevent AI from going rogue is by teaching our machines ethics so they don't cause problems.
- The questions of what we should, or even can, teach computers remains unknown.
- How we pick the values artificial intelligence follows might be the most important thing.
What effect does how we build the machine have on what ethics the machine can follow?<iframe width="730" height="430" src="https://www.youtube.com/embed/IHE63fxpHCg" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe><p> <strong><br> </strong>Humans are really good at explaining ethical problems and discussing potential solutions. Some of us are very good at teaching entire systems of ethics to other people. However, we tend to do this using language rather than code. We also teach people with learning capabilities similar to us rather than to a machine with different abilities. Shifting from people to machines may introduce some limitations. <br> <br> Many different methods of machine learning could be applied to ethical theory. The trouble is, they may prove to be very capable of absorbing one moral stance and utterly incapable of handling another. </p><p>Reinforcement learning (RL) is a way to teach a machine to do something by having it maximize a reward signal. Through trial and error, the machine is eventually able to learn how to get as much reward as possible efficiently. With its built-in tendency to maximize what is defined as good, this system clearly lends itself to utilitarianism, with its goal of maximizing the total happiness, and other consequentialist ethical systems. How to use it to effectively teach a different ethical system remains unknown. <br> <br> Alternatively, apprenticeship or imitation learning allows a programmer to give a computer a long list of data or an exemplar to observe and allow the machine to infer values and preferences from it. Thinkers concerned with the alignment problem often argue that this could teach a machine our preferences and values through action rather than idealized language. It would just require us to show the machine a moral exemplar and tell it to copy what they do. The idea has more than a few similarities to <a href="https://bigthink.com/scotty-hendricks/virtue-ethics-the-moral-system-you-have-never-heard-of-but-have-probably-used" target="_self">virtue ethics</a>. </p><p>The problem of who is a moral exemplar for other people remains unsolved, and who, if anybody, we should have computers try to emulate is equally up for debate. <br> <br> At the same time, there are some moral theories that we don't know how to teach to machines. Deontological theories, known for creating universal rules to stick to all the time, typically rely on a moral agent to apply reason to the situation they find themselves in along particular lines. No machine in existence is currently able to do that. Even the more limited idea of rights, and the concept that they should not be violated no matter what any optimization tendency says, might prove challenging to code into a machine, given how specific and clearly defined you'd have to make these rights.</p><p>After discussing these problems, Gabriel notes that:<br> <br> "In the light of these considerations, it seems possible that the methods we use to build artificial agents may influence the kind of values or principles we are able encode."<br> <br> This is a very real problem. After all, if you have a super AI, wouldn't you want to teach it ethics with the learning technique best suited for how you built it? What do you do if that technique can't teach it anything besides utilitarianism very well but you've decided virtue ethics is the right way to go? </p>
Machine learning is a powerful and imperfect tool that should not go unmonitored.
- When you harness the power and potential of machine learning, there are also some drastic downsides that you've got to manage.
- Deploying machine learning, you face the risk that it be discriminatory, biased, inequitable, exploitative, or opaque.
- In this article, I cover six ways that machine learning threatens social justice and reach an incisive conclusion: The remedy is to take on machine learning standardization as a form of social activism.
Here are six ways machine learning threatens social justice<img type="lazy-image" data-runner-src="https://assets.rebelmouse.io/eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpbWFnZSI6Imh0dHBzOi8vYXNzZXRzLnJibC5tcy8yNDUyMDgxNC9vcmlnaW4uanBnIiwiZXhwaXJlc19hdCI6MTY0MzM0NjgxOH0.zHvEEsYGbNA-lnkq4nss7vwVkZlrKkuKf0XASf7A7Jg/img.jpg?width=980" id="05f07" class="rm-shortcode" data-rm-shortcode-id="a7089b6621166f5a2df77d975f8b9f74" data-rm-shortcode-name="rebelmouse-image" />
Credit: metamorworks via Shutterstock<p><strong></strong><strong>1) </strong><strong>Blatantly discriminatory models</strong> are predictive models that base decisions partly or entirely on a protected class. Protected classes include race, religion, national origin, gender, gender identity, sexual orientation, pregnancy, and disability status. By taking one of these characteristics as an input, the model's outputs – and the decisions driven by the model – are based at least in part on membership in a protected class. Although models rarely do so directly, there is <a href="https://www.youtube.com/watch?v=eSlzy1x6Fy0" target="_blank">precedent</a> and <a href="https://www.youtube.com/watch?v=wfpNN8ASIq4" target="_blank">support</a> for doing so.</p><p>This would mean that a model could explicitly hinder, for example, black defendants for being black. So, imagine sitting across from a person being evaluated for a job, a loan, or even parole. When they ask you how the decision process works, you inform them, "For one thing, our algorithm penalized your score by seven points because you're black." This may sound shocking and sensationalistic, but I'm only literally describing what the model would do, mechanically, if race were permitted as a model input. </p><p><strong>2) Machine bias</strong>. Even when protected classes are not provided as a direct model input, we find, in some cases, that model predictions are still inequitable. This is because other variables end up serving as proxies to protected classes. This is <a href="https://coursera.org/share/51350b8fb12a5937bbddc0e53a4f207d" target="_blank" rel="noopener noreferrer">a bit complicated</a>, since it turns out that models that are fair in one sense are unfair in another. </p><p>For example, some crime risk models succeed in flagging both black and white defendants with equal precision – each flag tells the same probabilistic story, regardless of race – and yet the models falsely flag black defendants more often than white ones. A crime-risk model called COMPAS, which is sold to law enforcement across the US, falsely flags white defendants at a rate of 23.5%, and Black defendants at 44.9%. In other words, black defendants who don't deserve it are <a href="https://coursera.org/share/df6e6ba7108980bb7eeae0ba22123ac1" target="_blank" rel="noopener noreferrer">erroneously flagged almost twice as much</a> as white defendants who don't deserve it.</p><p><strong>3) Inferring sensitive attributes</strong>—predicting pregnancy and beyond. Machine learning predicts sensitive information about individuals, such as sexual orientation, whether they're pregnant, whether they'll quit their job, and whether they're going to die. Researchers have shown that it is possible to <a href="https://youtu.be/aNwvXhcq9hk" target="_blank" rel="noopener noreferrer">predict race based on Facebook likes</a>. These predictive models deliver dynamite.</p><p>In a particularly extraordinary case, officials in China use facial recognition to <a href="https://www.nytimes.com/2019/04/14/technology/china-surveillance-artificial-intelligence-racial-profiling.html" target="_blank" rel="noopener noreferrer">identify and track the Uighurs, a minority ethnic group</a> systematically oppressed by the government. This is the first known case of a government using machine learning to profile by ethnicity. One Chinese start-up valued at more than $1 billion said its software could recognize "sensitive groups of people." It's website said, "If originally one Uighur lives in a neighborhood, and within 20 days six Uighurs appear, it immediately sends alarms" to law enforcement.</p>