#neural network datasets

[aiweirdness]

Botober 2024

Back by popular demand, here are some AI-generated drawing prompts to use in this, the spooky month of October!

Longtime AI Weirdness readers may recognize many of these - that's because there are throwbacks to very tiny language models, circa 2017-2018. (There are 7 tiny models each contributing a few groups of prompts; feel free to guess what they were trained on and then check your answers at aiweirdness.com)

[d0nutzgg]

Tonight I am hunting down venomous and nonvenomous snake pictures that are under the creative commons of specific breeds in order to create one of the most advanced, in depth datasets of different venomous and nonvenomous snakes as well as a test set that will include snakes from both sides of all species. I love snakes a lot and really, all reptiles. It is definitely tedious work, as I have to make sure each picture is cleared before I can use it (ethically), but I am making a lot of progress! I have species such as the King Cobra, Inland Taipan, and Eyelash Pit Viper among just a few! Wikimedia Commons has been a huge help!

I'm super excited.

Hope your nights are going good. I am still not feeling good but jamming + virtual snake hunting is keeping me busy!

[jcmarchi]

Hyperrealistic Deepfakes: A Growing Threat to Truth and Reality

New Post has been published on https://thedigitalinsider.com/hyperrealistic-deepfakes-a-growing-threat-to-truth-and-reality/

Hyperrealistic Deepfakes: A Growing Threat to Truth and Reality

In an era where technology evolves at an exceptionally fast pace, deepfakes have emerged as a controversial and potentially dangerous innovation. These hyperrealistic digital forgeries, created using advanced Artificial Intelligence (AI) techniques like Generative Adversarial Networks (GANs), can mimic real-life appearances and movements with supernatural accuracy.

Initially, deepfakes were a niche application, but they have quickly gained prominence, blurring the lines between reality and fiction. While the entertainment industry uses deepfakes for visual effects and creative storytelling, the darker implications are alarming. Hyperrealistic deepfakes can undermine the integrity of information, erode public trust, and disrupt social and political systems. They are gradually becoming tools to spread misinformation, manipulate political outcomes, and damage personal reputations.

The Origins and Evolution of Deepfakes

Deepfakes utilize advanced AI techniques to create incredibly realistic and convincing digital forgeries. These techniques involve training neural networks on large datasets of images and videos, enabling them to generate synthetic media that closely mimics real-life appearances and movements. The advent of GANs in 2014 marked a significant milestone, allowing the creation of more sophisticated and hyperrealistic deepfakes.

GANs consist of two neural networks, the generator and the discriminator, working in tandem. The generator creates fake images while the discriminator attempts to distinguish between real and fake images. Through this adversarial process, both networks improve, leading to the creation of highly realistic synthetic media.

Recent advancements in machine learning techniques, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), have further enhanced the realism of deepfakes. These advancements allow for better temporal coherence, meaning synthesized videos are smoother and more consistent over time.

The spike in deepfake quality is primarily due to advancements in AI algorithms, more extensive training datasets, and increased computational power. Deepfakes can now replicate not just facial features and expressions but also minute details like skin texture, eye movements, and subtle gestures. The availability of vast amounts of high-resolution data, coupled with powerful GPUs and cloud computing, has also accelerated the development of hyperrealistic deepfakes.

The Dual-Edged Sword of Technology

While the technology behind deepfakes has legitimate and beneficial applications in entertainment, education, and even medicine, its potential for misuse is alarming. Hyperrealistic deepfakes can be weaponized in several ways, including political manipulation, misinformation, cybersecurity threats, and reputation damage.

For instance, deepfakes can create false statements or actions by public figures, potentially influencing elections and undermining democratic processes. They can also spread misinformation, making it nearly impossible to distinguish between genuine and fake content. Deepfakes can bypass security systems that rely on biometric data, posing a significant threat to personal and organizational security. Additionally, individuals and organizations can suffer immense harm from deepfakes that depict them in compromising or defamatory situations.

Real-World Impact and Psychological Consequences

Several high-profile cases have demonstrated the potential for harm from hyperrealistic deepfakes. The deepfake video created by filmmaker Jordan Peele and released by BuzzFeed showed former President Barack Obama appearing to say derogatory remarks about Donald Trump. This video was created to raise awareness about the potential dangers of deepfakes and how they can be used to spread disinformation.

Likewise, another deepfake video depicted Mark Zuckerberg boasting about having control over users’ data, suggesting a scenario where data control translates to power. This video, created as part of an art installation, was intended to critique the power held by tech giants.

Similarly, the Nancy Pelosi video in 2019, though not a deepfake, points out how easy it is to spread misleading content and the potential consequences. In 2021, a series of deepfake videos featuring actor Tom Cruise went viral on TikTok, demonstrating the power of hyperrealistic deepfakes to capture public attention and go viral. These cases illustrate the psychological and societal implications of deepfakes, including the erosion of trust in digital media and the potential for increased polarization and conflict.

Psychological and Societal Implications

Beyond the immediate threats to individuals and institutions, hyperrealistic deepfakes have broader psychological and societal implications. The erosion of trust in digital media can lead to a phenomenon known as the “liar’s dividend,” where the mere possibility of content being fake can be used to dismiss genuine evidence.

As deepfakes become more prevalent, public trust in media sources may diminish. People may become skeptical of all digital content, undermining the credibility of legitimate news organizations. This distrust can aggravate societal divisions and polarize communities. When people cannot agree on basic facts, constructive dialogue and problem-solving become increasingly difficult.

In addition, misinformation and fake news, amplified by deepfakes, can deepen existing societal rifts, leading to increased polarization and conflict. This can make it harder for communities to come together and address shared challenges.

Legal and Ethical Challenges

The rise of hyperrealistic deepfakes presents new challenges for legal systems worldwide. Legislators and law enforcement agencies must make efforts to define and regulate digital forgeries, balancing the need for security with the protection of free speech and privacy rights.

Making effective legislation to combat deepfakes is complex. Laws must be precise enough to target malicious actors without hindering innovation or infringing on free speech. This requires careful consideration and collaboration among legal experts, technologists, and policymakers. For instance, the United States passed the DEEPFAKES Accountability Act, making it illegal to create or distribute deepfakes without disclosing their artificial nature. Similarly, several other countries, such as China and the European Union, are coming up with strict and comprehensive AI regulations to avoid problems.

Combating the Deepfake Threat

Addressing the threat of hyperrealistic deepfakes requires a multifaceted approach involving technological, legal, and societal measures.

Technological solutions include detection algorithms that can identify deepfakes by analyzing inconsistencies in lighting, shadows, and facial movements, digital watermarking to verify the authenticity of media, and blockchain technology to provide a decentralized and immutable record of media provenance.

Legal and regulatory measures include passing laws to address the creation and distribution of deepfakes and establishing dedicated regulatory bodies to monitor and respond to deepfake-related incidents.

Societal and educational initiatives include media literacy programs to help individuals critically evaluate content and public awareness campaigns to inform citizens about deepfakes. Moreover, collaboration among governments, tech companies, academia, and civil society is essential to combat the deepfake threat effectively.

The Bottom Line

Hyperrealistic deepfakes pose a significant threat to our perception of truth and reality. While they offer exciting possibilities in entertainment and education, their potential for misuse is alarming. To combat this threat, a multifaceted approach involving advanced detection technologies, robust legal frameworks, and comprehensive public awareness is essential.

By encouraging collaboration among technologists, policymakers, and society, we can mitigate the risks and preserve the integrity of information in the digital age. It is a collective effort to ensure that innovation does not come at the cost of trust and truth.

[dberga]

Publication in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

https://ieeexplore.ieee.org/document/10356628

A New Framework for Evaluating Image Quality Including Deep Learning Task Performances as A Proxy

iquaflow is a framework that provides a set of tools to assess image quality. The user can add custom metrics that can be easily integrated and a set of unsupervised methods is offered by default. Furthermore, iquaflow measures quality by using the performance of AI models trained on the images as a proxy. This also helps to easily make studies of performance degradation of several modifications of the original dataset, for instance, with images reconstructed after different levels of lossy compression; satellite images would be a use case example, since they are commonly compressed before downloading to the ground. In this situation, the optimization problem involves finding images that, while being compressed to their smallest possible file size, still maintain sufficient quality to meet the required performance of the deep learning algorithms. Thus, a study with iquaflow is suitable for such case. All this development is wrapped in Mlflow : an interactive tool used to visualize and summarize the results. This document describes different use cases and provides links to their respective repositories. To ease the creation of new studies, we include a cookiecutter repository. The source code, issue tracker and aforementioned repositories are all hosted on GitHub.

https://github.com/satellogic/iquaflow

[river-taxbird]

There is no such thing as AI.

How to help the non technical and less online people in your life navigate the latest techbro grift.

I've seen other people say stuff to this effect but it's worth reiterating. Today in class, my professor was talking about a news article where a celebrity's likeness was used in an ai image without their permission. Then she mentioned a guest lecture about how AI is going to help finance professionals. Then I pointed out, those two things aren't really related.

The term AI is being used to obfuscate details about multiple semi-related technologies.

Traditionally in sci-fi, AI means artificial general intelligence like Data from star trek, or the terminator. This, I shouldn't need to say, doesn't exist. Techbros use the term AI to trick investors into funding their projects. It's largely a grift.

What is the term AI being used to obfuscate?

If you want to help the less online and less tech literate people in your life navigate the hype around AI, the best way to do it is to encourage them to change their language around AI topics.

By calling these technologies what they really are, and encouraging the people around us to know the real names, we can help lift the veil, kill the hype, and keep people safe from scams. Here are some starting points, which I am just pulling from Wikipedia. I'd highly encourage you to do your own research.

Machine learning (ML): is an umbrella term for solving problems for which development of algorithms by human programmers would be cost-prohibitive, and instead the problems are solved by helping machines "discover" their "own" algorithms, without needing to be explicitly told what to do by any human-developed algorithms. (This is the basis of most technologically people call AI)

Language model: (LM or LLM) is a probabilistic model of a natural language that can generate probabilities of a series of words, based on text corpora in one or multiple languages it was trained on. (This would be your ChatGPT.)

Generative adversarial network (GAN): is a class of machine learning framework and a prominent framework for approaching generative AI. In a GAN, two neural networks contest with each other in the form of a zero-sum game, where one agent's gain is another agent's loss. (This is the source of some AI images and deepfakes.)

Diffusion Models: Models that generate the probability distribution of a given dataset. In image generation, a neural network is trained to denoise images with added gaussian noise by learning to remove the noise. After the training is complete, it can then be used for image generation by starting with a random noise image and denoise that. (This is the more common technology behind AI images, including Dall-E and Stable Diffusion. I added this one to the post after as it was brought to my attention it is now more common than GANs.)

I know these terms are more technical, but they are also more accurate, and they can easily be explained in a way non-technical people can understand. The grifters are using language to give this technology its power, so we can use language to take it's power away and let people see it for what it really is.

[gothhabiba]

On the one hand, people who take a hardline stance on “AI art is not art” are clearly saying something naïve and indefensible (as though any process cannot be used to make art? as though artistry cannot still be involved in the set-up of the parameters and the choice of data set and the framing of the result? as though “AI” means any one thing? you’re going to have a real hard time with process music, poetry cut-up methods, &c.).

But all of this (as well as takes that what's really needed is a crackdown on IP) are a distraction from a vital issue—namely that this is technology used to create and sort enormous databases of images, and the uses to which this technology is put in a police state are obvious: it's used in service of surveillance, incarceration, criminalisation, and the furthering of violence against criminalised people.

Of course we've long known that datasets are not "neutral" and that racist data will provide racist outcomes, and we've long known that the problem goes beyond the datasets (even carefully vetting datasets does not necessarily control for social factors). With regards to "predictive policing," this suggests that criminalisation of supposed leftist "radicals" and racialised people (and the concepts creating these two groups overlap significantly; [link 1], [link 2]) is not a problem, but intentional—a process is built so that it always finds people "suspicious" or "guilty," but because it is based on an "algorithm" or "machine learning" or so-called "AI" (processes that people tend to understand murkily, if at all), they can be presented as innocent and neutral. These are things that have been brought up repeatedly with regards to "automatic" processes and things that trawl the web to produce large datasets in the recent past (e.g. facial recognition technology), so their almost complete absence from the discourse wrt "AI art" confuses me.

Abeba Birhane's thread here, summarizing this paper (h/t @thingsthatmakeyouacey) explains how the LAION-400M dataset was sourced/created, how it is filtered, and how images are retrieved from it (for this reason it's a good beginner explanation of what large-scale datasets and large neural networks are 'doing'). She goes into how racist, misogynistic, and sexually violent content is returned (and racist mis-categorisations are made) as a result of every one of those processes. She also brings up issues of privacy, how individuals' data is stored in datasets (even after the individual deletes it from where it was originally posted), and how it may be stored associated with metadata which the poster did not intend to make public. This paper (h/t thingsthatmakeyouacey [link]) looks at the ImageNet-ILSVRC-2012 dataset to discuss "the landscape of harm and threats both the society at large and individuals face due to uncritical and ill-considered dataset curation practices" including the inclusion of non-consensual pornography in the dataset.

Of course (again) this is nothing that hasn't already been happening with large social media websites or with "big data" (Birhane notes that "On the one hand LAION-400M has opened a door that allows us to get a glimpse into the world of large scale datasets; these kinds of datasets remain hidden inside BigTech corps"). And there's no un-creating the technology behind this—resistance will have to be directed towards demolishing the police / carceral / imperial state as a whole. But all criticism of "AI" art can't be dismissed as always revolving around an anti-intellectual lack of knowledge of art history or else a reactionary desire to strengthen IP law (as though that would ever benefit small creators at the expense of large corporations...).

[foone]

So here's the thing about AI art, and why it seems to be connected to a bunch of unethical scumbags despite being an ethically neutral technology on its own. After the readmore, cause long. Tl;dr: capitalism

The problem is competition. More generally, the problem is capitalism.

So the kind of AI art we're seeing these days is based on something called "deep learning", a type of machine learning based on neural networks. How they work exactly isn't important, but one aspect in general is: they have to be trained.

The way it works is that if you want your AI to be able to generate X, you have to be able to train it on a lot of X. The more, the better. It gets better and better at generating something the more it has seen it. Too small a training dataset and it will do a bad job of generating it.

So you need to feed your hungry AI as much as you can. Now, say you've got two AI projects starting up:

Project A wants to do this ethically. They generate their own content to train the AI on, and they seek out datasets that allow them to be used in AI training systems. They avoid misusing any public data that doesn't explicitly give consent for the data to be used for AI training.

Meanwhile, Project B has no interest in the ethics of what they're doing, so long as it makes them money. So they don't shy away from scraping entire websites of user-submitted content and stuffing it into their AI. DeviantArt, Flickr, Tumblr? It's all the same to them. Shove it in!

Now let's fast forward a couple months of these two projects doing this. They both go to demo their project to potential investors and the public art large.

Which one do you think has a better-trained AI? the one with the smaller, ethically-obtained dataset? Or the one with the much larger dataset that they "found" somewhere after it fell off a truck?

It's gonna be the second one, every time. So they get the money, they get the attention, they get to keep growing as more and more data gets stuffed into it.

And this has a follow-on effect: we've just pre-selected AI projects for being run by amoral bastards, remember. So when someone is like "hey can we use this AI to make NFTs?" or "Hey can your AI help us detect illegal immigrants by scanning Facebook selfies?", of course they're gonna say "yeah, if you pay us enough".

So while the technology is not, in itself, immoral or unethical, the situations around how it gets used in capitalism definitely are. That external influence heavily affects how it gets used, and who "wins" in this field. And it won't be the good guys.

An important follow-up: this is focusing on the production side of AI, but obviously even if you had an AI art generator trained on entirely ethically sourced data, it could still be used unethically: it could put artists out of work, by replacing their labor with cheaper machine labor. Again, this is not a problem of the technology itself: it's a problem of capitalism. If artists weren't competing to survive, the existence of cheap AI art would not be a threat.

I just feel it's important to point this out, because I sometimes see people defending the existence of AI Art from a sort of abstract perspective. Yes, if you separate it completely from the society we live in, it's a neutral or even good technology. Unfortunately, we still live in a world ruled by capitalism, and it only makes sense to analyze AI Art from a perspective of having to continue to live in capitalism alongside it.

If you want ideologically pure AI Art, feel free to rise up, lose your chains, overthrow the bourgeoisie, and all that. But it's naive to defend it as just a neutral technology like any other when it's being wielded in capitalism; ie overwhelmingly negatively in impact.

[shirecorn]

so I binged all of your stuff I could find about your Skyscraper gods lore, and I'm curious - how, if at all, do creatures like changelings, the sirens/the other realm in general, and Discord fit into your AU?

Discord is actually from another universe! He is not a giant god thing, but he came to equestria and decided he liked it here and he could really mess this place up. After being imprisoned in stone for a thousand years and witnessing the civilization from a town square (I forget where he was put tbh) he became both super angry and somewhat paradoxically attached to this reality. When he was unstoned he got super vengeful. But then meeting fluttershy and befriending her helped him become even more attached to this place. It's not his home, though, and he only comes around to help save it because his friends are here.

Also he does not have a set appearance and shifts his shape with every step and gesture. Its a bit dizzying to watch, like the early examples of neural network stuff from 2015. Like you're looking at a tiger through an "AI" (I HATE that we call this BS ai, we should go back to saying neural net and machine learning, or better yet Machine Bias Training) anyway looking at a tiger but through a filter of an AI that has been trained to see horses everywhere. And by the time the filter has switched to a dataset of tigers, his actual form has turned into a snake.

This is my vision and I have done like. 10 drawings in order to capture it but the video editing to get there is beyond my current skillset so I keep procrastinating on starting

---

Changelings are insect-related creatures and are as far from mammals as you can get. I need to do a couple more drawings before I'm satisfied because I came up with a new concept after the first art, and just need to draw it.

---

is the other realm equestria girls? I dont watch EQ so I dont know how to fit it. I know there are sirens in a flashback in the pony show, in uhh Season 7 episode 25 or 26

[chryza]

Idk tho I kinda liked AI art as an artist like I wish there was a generator from a clean dataset because it was fun to get ideas and things that I would draw or incorporate into other ideas for the fun of it. It was hardly a replacement for the actual work of art or writing but as an experimental step towards the future of technology as it relates to art I find it fascinating. Especially once you start looking into the really experimental stuff like people generating faces while telling the computer to slowly subtract faces from its input and watching the computer produce an idea of a human without the correct parameters. Highlighting why the program kept going back to the same features over and over despite being given vastly different prompts. I was truly so excited for the future of neural network artwork and it’s all been just….destroyed.

[mariacallous]

How will AI be used in health care settings?

Artificial intelligence (AI) shows tremendous promise for applications in health care. Tools such as machine learning algorithms, artificial neural networks, and generative AI (e.g., Large Language Models) have the potential to aid with tasks such as diagnosis, treatment planning, and resource management. Advocates have suggested that these tools could benefit large numbers of people by increasing access to health care services (especially for populations that are currently underserved), reducing costs, and improving quality of care.

This enthusiasm has driven the burgeoning development and trial application of AI in health care by some of the largest players in the tech industry. To give just two examples, Google Research has been rapidly testing and improving upon its “Med-PaLM” tool, and NVIDIA recently announced a partnership with Hippocratic AI that aims to deploy virtual health care assistants for a variety of tasks to address a current shortfall in the supply in the workforce.

What are some challenges or potential negative consequences to using AI in health care?

Technology adoption can happen rapidly, exponentially going from prototypes used by a small number of researchers to products affecting the lives of millions or even billions of people. Given the significant impact health care system changes could have on Americans’ health as well as on the U.S. economy, it is essential to preemptively identify potential pitfalls before scaleup takes place and carefully consider policy actions that can address them.

One area of concern arises from the recognition that the ultimate impact of AI on health outcomes will be shaped not only by the sophistication of the technological tools themselves but also by external “human factors.” Broadly speaking, human factors could blunt the positive impacts of AI tools in health care—or even introduce unintended, negative consequences—in two ways:

If developers train AI tools with data that don’t sufficiently mirror diversity in the populations in which they will be deployed. Even tools that are effective in the aggregate could create disparate outcomes. For example, if the datasets used to train AI have gaps, they can cause AI to provide responses that are lower quality for some users and situations. This might lead to the tool systematically providing less accurate recommendations for some groups of users or experiencing “catastrophic failures” more frequently for some groups, such as failure to identify symptoms in time for effective treatment or even recommending courses of treatment that could result in harm.  

If patterns of AI use systematically differ across groups. There may be an initial skepticism among many potential users to trust AI for consequential decisions that affect their health. Attitudes may differ within the population based on attributes such as age and familiarity with technology, which could affect who uses AI tools, understands and interprets the AI’s output, and adheres to treatment recommendations. Further, people’s impressions of AI health care tools will be shaped over time based on their own experiences and what they learn from others.

In recent research, we used simulation modeling to study a large range of different of hypothetical populations of users and AI health care tool specifications. We found that social conditions such as initial attitudes toward AI tools within a population and how people change their attitudes over time can potentially:

Lead to a modestly accurate AI tool having a negative impact on population health. This can occur because people’s experiences with an AI tool may be filtered through their expectations and then shared with others. For example, if an AI tool’s capabilities are objectively positive—in expectation, the AI won’t give recommendations that are harmful or completely ineffective—but sufficiently lower than expectations, users who are disappointed will lose trust in the tool. This could make them less likely to seek future treatment or adhere to recommendations if they do and lead them to pass along negative perceptions of the tool to friends, family, and others with whom they interact.

Create health disparities even after the introduction of a high-performing and unbiased AI tool (i.e., that performs equally well for all users). Specifically, when there are initial differences between groups within the population in their trust of AI-based health care—for example because of one group’s systematically negative previous experiences with health care or due to the AI tool being poorly communicated to one group—differential use patterns alone can translate into meaningful differences in health patterns across groups. These use patterns can also exacerbate differential effects on health across groups when AI training deficiencies cause a tool to provide better quality recommendations for some users than others.

Barriers to positive health impacts associated with systematic and shifting use patterns are largely beyond individual developers’ direct control but can be overcome with strategically designed policies and practices.

What could a regulatory framework for AI in health care look like?

Disregarding how human factors intersect with AI-powered health care tools can create outcomes that are costly in terms of life, health, and resources. There is also the potential that without careful oversight and forethought, AI tools can maintain or exacerbate existing health disparities or even introduce new ones. Guarding against negative consequences will require specific policies and ongoing, coordinated action that goes beyond the usual scope of individual product development. Based on our research, we suggest that any regulatory framework for AI in health care should accomplish three aims:

Ensure that AI tools are rigorously tested before they are made fully available to the public and are subject to regular scrutiny afterward. Those developing AI tools for use in health care should carefully consider whether the training data are matched to the tasks that the tools will perform and representative of the full population of eventual users. Characteristics of users to consider include (but are certainly not limited to) age, gender, culture, ethnicity, socioeconomic status, education, and language fluency. Policies should encourage and support developers in investing time and resources into pre- and post-launch assessments, including:

pilot tests to assess performance across a wide variety of groups that might experience disparate impact before large-scale application

monitoring whether and to what extent disparate use patterns and outcomes are observed after release

identifying appropriate corrective action if issues are found.

Require that users be clearly informed about what tools can do and what they cannot. Neither health care workers nor patients are likely to have extensive training or sophisticated understanding of the technical underpinnings of AI tools. It will be essential that plain-language use instructions, cautionary warnings, or other features designed to inform appropriate application boundaries are built into tools. Without these features, users’ expectations of AI capabilities might be inaccurate, with negative effects on health outcomes. For example, a recent report outlines how overreliance on AI tools by inexperienced mushroom foragers has led to cases of poisoning; it is easy to imagine how this might be a harbinger of patients misdiagnosing themselves with health care tools that are made publicly available and missing critical treatment or advocating for treatment that is contraindicated. Similarly, tools used by health care professionals should be supported by rigorous use protocols. Although advanced tools will likely provide accurate guidance an overwhelming majority of the time, they can also experience catastrophic failures (such as those referred to as “hallucinations” in the AI field), so it is critical for trained human users to be in the loop when making key decisions.

Proactively protect against medical misinformation. False or misleading claims about health and health care—whether the result of ignorance or malicious intent—have proliferated in digital spaces and become harder for the average person to distinguish from reliable information. This type of misinformation about health care AI tools presents a serious threat, potentially leading to mistrust or misapplication of these tools. To discourage misinformation, guardrails should be put in place to ensure consistent transparency about what data are used and how that continuous verification of training data accuracy takes place.

How can regulation of AI in health care keep pace with rapidly changing conditions?

In addition to developers of tools themselves, there are important opportunities for unaffiliated researchers to study the impact of AI health care tools as they are introduced and recommend adjustments to any regulatory framework. Two examples of what this work might contribute are:

Social scientists can learn more about how people think about and engage with AI tools, as well as how perceptions and behaviors change over time. Rigorous data collection and qualitative and quantitative analyses can shed light on these questions, improving understanding of how individuals, communities, and society adapt to shifts in the health care landscape.

Systems scientists can consider the co-evolution of AI tools and human behavior over time. Building on or tangential to recent research, systems science can be used to explore the complex interactions that determine how multiple health care AI tools deployed across diverse settings might affect long-term health trends. Using longitudinal data collected as AI tools come into widespread use, prospective simulation models can provide timely guidance on how policies might need to be course corrected.

[apas-95]

i think a lot of people like. skimmed through posts last year about bias in AI datasets and came to the opinion that neural networks are just an inherently evil form of technology, to the point i'm now seeing posts that have 'AI is implemented in more tech' as a straightforward, unclarified Bad Outcome

[jcmarchi]

AI’s Inner Dialogue: How Self-Reflection Enhances Chatbots and Virtual Assistants

New Post has been published on https://thedigitalinsider.com/ais-inner-dialogue-how-self-reflection-enhances-chatbots-and-virtual-assistants/

AI’s Inner Dialogue: How Self-Reflection Enhances Chatbots and Virtual Assistants

Recently, Artificial Intelligence (AI) chatbots and virtual assistants have become indispensable, transforming our interactions with digital platforms and services. These intelligent systems can understand natural language and adapt to context. They are ubiquitous in our daily lives, whether as customer service bots on websites or voice-activated assistants on our smartphones. However, an often-overlooked aspect called self-reflection is behind their extraordinary abilities. Like humans, these digital companions can benefit significantly from introspection, analyzing their processes, biases, and decision-making.

This self-awareness is not merely a theoretical concept but a practical necessity for AI to progress into more effective and ethical tools. Recognizing the importance of self-reflection in AI can lead to powerful technological advancements that are also responsible and empathetic to human needs and values. This empowerment of AI systems through self-reflection leads to a future where AI is not just a tool, but a partner in our digital interactions.

Understanding Self-Reflection in AI Systems

Self-reflection in AI is the capability of AI systems to introspect and analyze their own processes, decisions, and underlying mechanisms. This involves evaluating internal processes, biases, assumptions, and performance metrics to understand how specific outputs are derived from input data. It includes deciphering neural network layers, feature extraction methods, and decision-making pathways.

Self-reflection is particularly vital for chatbots and virtual assistants. These AI systems directly engage with users, making it essential for them to adapt and improve based on user interactions. Self-reflective chatbots can adapt to user preferences, context, and conversational nuances, learning from past interactions to offer more personalized and relevant responses. They can also recognize and address biases inherent in their training data or assumptions made during inference, actively working towards fairness and reducing unintended discrimination.

Incorporating self-reflection into chatbots and virtual assistants yields several benefits. First, it enhances their understanding of language, context, and user intent, increasing response accuracy. Secondly, chatbots can make adequate decisions and avoid potentially harmful outcomes by analyzing and addressing biases. Lastly, self-reflection enables chatbots to accumulate knowledge over time, augmenting their capabilities beyond their initial training, thus enabling long-term learning and improvement. This continuous self-improvement is vital for resilience in novel situations and maintaining relevance in a rapidly evolving technological world.

The Inner Dialogue: How AI Systems Think

AI systems, such as chatbots and virtual assistants, simulate a thought process that involves complex modeling and learning mechanisms. These systems rely heavily on neural networks to process vast amounts of information. During training, neural networks learn patterns from extensive datasets. These networks propagate forward when encountering new input data, such as a user query. This process computes an output, and if the result is incorrect, backward propagation adjusts the network’s weights to minimize errors. Neurons within these networks apply activation functions to their inputs, introducing non-linearity that enables the system to capture complex relationships.

AI models, particularly chatbots, learn from interactions through various learning paradigms, for example:

In supervised learning, chatbots learn from labeled examples, such as historical conversations, to map inputs to outputs.

Reinforcement learning involves chatbots receiving rewards (positive or negative) based on their responses, allowing them to adjust their behavior to maximize rewards over time.

Transfer learning utilizes pre-trained models like GPT that have learned general language understanding. Fine-tuning these models adapts them to tasks such as generating chatbot responses.

It is essential to balance adaptability and consistency for chatbots. They must adapt to diverse user queries, contexts, and tones, continually learning from each interaction to improve future responses. However, maintaining consistency in behavior and personality is equally important. In other words, chatbots should avoid drastic changes in personality and refrain from contradicting themselves to ensure a coherent and reliable user experience.

Enhancing User Experience Through Self-Reflection

Enhancing the user experience through self-reflection involves several vital aspects contributing to chatbots and virtual assistants’ effectiveness and ethical behavior. Firstly, self-reflective chatbots excel in personalization and context awareness by maintaining user profiles and remembering preferences and past interactions. This personalized approach enhances user satisfaction, making them feel valued and understood. By analyzing contextual cues such as previous messages and user intent, self-reflective chatbots deliver more relevant and meaningful answers, enhancing the overall user experience.

Another vital aspect of self-reflection in chatbots is reducing bias and improving fairness. Self-reflective chatbots actively detect biased responses related to gender, race, or other sensitive attributes and adjust their behavior accordingly to avoid perpetuating harmful stereotypes. This emphasis on reducing bias through self-reflection reassures the audience about the ethical implications of AI, making them feel more confident in its use.

Furthermore, self-reflection empowers chatbots to handle ambiguity and uncertainty in user queries effectively. Ambiguity is a common challenge chatbots face, but self-reflection enables them to seek clarifications or provide context-aware responses that enhance understanding.

Case Studies: Successful Implementations of Self-Reflective AI Systems

Google’s BERT and Transformer models have significantly improved natural language understanding by employing self-reflective pre-training on extensive text data. This allows them to understand context in both directions, enhancing language processing capabilities.

Similarly, OpenAI’s GPT series demonstrates the effectiveness of self-reflection in AI. These models learn from various Internet texts during pre-training and can adapt to multiple tasks through fine-tuning. Their introspective ability to train data and use context is key to their adaptability and high performance across different applications.

Likewise, Microsoft’s ChatGPT and Copilot utilize self-reflection to enhance user interactions and task performance. ChatGPT generates conversational responses by adapting to user input and context, reflecting on its training data and interactions. Similarly, Copilot assists developers with code suggestions and explanations, improving their suggestions through self-reflection based on user feedback and interactions.

Other notable examples include Amazon’s Alexa, which uses self-reflection to personalize user experiences, and IBM’s Watson, which leverages self-reflection to enhance its diagnostic capabilities in healthcare.

These case studies exemplify the transformative impact of self-reflective AI, enhancing capabilities and fostering continuous improvement.

Ethical Considerations and Challenges

Ethical considerations and challenges are significant in the development of self-reflective AI systems. Transparency and accountability are at the forefront, necessitating explainable systems that can justify their decisions. This transparency is essential for users to comprehend the rationale behind a chatbot’s responses, while auditability ensures traceability and accountability for those decisions.

Equally important is the establishment of guardrails for self-reflection. These boundaries are essential to prevent chatbots from straying too far from their designed behavior, ensuring consistency and reliability in their interactions.

Human oversight is another aspect, with human reviewers playing a pivotal role in identifying and correcting harmful patterns in chatbot behavior, such as bias or offensive language. This emphasis on human oversight in self-reflective AI systems provides the audience with a sense of security, knowing that humans are still in control.

Lastly, it is critical to avoid harmful feedback loops. Self-reflective AI must proactively address bias amplification, particularly if learning from biased data.

The Bottom Line

In conclusion, self-reflection plays a pivotal role in enhancing AI systems’ capabilities and ethical behavior, particularly chatbots and virtual assistants. By introspecting and analyzing their processes, biases, and decision-making, these systems can improve response accuracy, reduce bias, and foster inclusivity.

Successful implementations of self-reflective AI, such as Google’s BERT and OpenAI’s GPT series, demonstrate this approach’s transformative impact. However, ethical considerations and challenges, including transparency, accountability, and guardrails, demand following responsible AI development and deployment practices.

[drosera-sundews]

On AI and art theft

This one’s for you @drawsoneverything​

So starting off with a disclaimer. I do not know how deviantart’s Dreamup works. I never worked there. I just know how AI works in general.

Shortest version: AIs (or neural networks, which I assume Dreamup is) are programs that mimic human neuron cells -and thus human learning processes-to a degree. A big difference between a neural network and a regular computer program is that neural networks require training. Like a human would require learning.

A neural network (the simple ones, at least) consist of a few ‘layers’ which contain many ‘nodes’. At least 3 are required, an input and output layer, plus one extra layer in between. Imagine it as something like this:

Naturally, this will all be computer code. But this is the basic anatomy of your simplest neural networks.

Each node is connected to each node, like so:

Not gonna draw all of it but you get the gist, ey. They are all connected much like -once again- neurons in the brain.

Now once you make an empty neural network like this it can’t do anything. It’s there but it’s useless and it’s interchangeable with every other untrained neural network (barring amount of layers and numbers of nodes).

What’s next is thinking of the task you want your network to perform. For example, to train it to recognize hand written letters. In this case, you’d have 26 nodes in your output layer, one for each letter of the alphabet. 

Now we need some training data

Perfect. Now imagine that’s a dataset with 5000 people’s handwriting. You need quite a lot of data to properly train an AI.

The idea is we want to input an image of a letter in any handwriting, and we want the AI to fire one of it’s output neurons, namely the one corresponding to the letter it ‘sees’.

The images need to be translated to numerical values, in order to be put in the input layer. This can be done, for example, by translating each pixel of the image in a numerical value, and having each node in the input layer be a ‘pixel’. In a human, these would be the neurons in our retina, or the back of your eye.

Translation of the image is going to look somewhat like this:

With the red numbers going into the input layer as the starting values.

We can now input our images. Something complex is going to happen. Every node is going to look at the values of the input nodes and nodes before them, and gain a value based on that, following arbitrary, random patterns. You don’t need to know the specifics here, just know that now that we have input, the nodes (or neurons) can fire.

Wow, the AI is just super wrong.

That’s okay tho, we can simply run it again until it hits the right answer by chance.

This is what the training is for. You need a dataset that’s annotated. In this case a human has to look at the inputs and identify them, so they can tell the AI if it’s right or wrong. If you do this enough times, the AI will learn to do it on it’s own.

And now here’s the catch. With every new piece of input data the machine guesses right, the values Thiof the nodes in the middle layers are changed. Like a maze, the paths to the right exit become clearer with every time the maze is completed. And while the values of the in and output layers are changed with every run, the middle layers remain. This is where the learning happens.

This is a trained neural network. And this is why it doesn’t matter that deviantart changed their policy to opt-in instead of opt-out. Those few days were enough. Once you train a neural network it doesn’t need it’s training data anymore.

Worst thing is, the middle layers are very much a black box. We don’t know what happens in there. The AI learns to categorize things on learned criteria, but what those are we can not know.

This is a very shortened down version of a very complex process ofc. Naturally, an art generator like DreamUp is going to be much more sophisticated. But I recon it follows the same rules as the simpler neural networks, which gives us something to work with.

Once a neural network is trained it cannot be ‘untrained’. Once you taint the dataset it’s very hard to reverse the effects. And generally, you can keep training an AI to make it better even if it’s already in use. Which I’m guessing is what deviantart is still doing with any new artworks that are added without the noai-tag

And the art on deviantart is already a neatly annotated dataset. We (the artists) have been annotating it by putting descriptions below the artworks, and giving it titles describing what the work depicts.

And oh, wouldn’t it be such a shame if someone were to accidentally add wrongly annotated art to this young, impressionable AI’s dataset.

Oh wow, if you uncheck the noai checkbox and then go back to re-edit your deviation, deviantart even let’s you directly annotate your art for their database! How convenient!

No artist alive has ever been able to draw a proper horse, I’m going to make sure this AI imitates life. Maybe in my next work I will teach it how to draw ‘the other eye’.

This is the beauty of the system. The AI can learn all it wants, but in the end it’s not a person. It doesn’t understand what it’s looking at.

Thus, we can do the most human thing possible to this mindless piece of code, lie!

This is the toddler we can learn curse words. This tool was designed to steal art en-masse, but it was left hilariously open and vulnerable. Let’s break this stupid thing!

Honestly, they called it deviantart, if anything I’m living up to the name. 

[sak-shi]

Python Libraries to Learn Before Tackling Data Analysis

To tackle data analysis effectively in Python, it's crucial to become familiar with several libraries that streamline the process of data manipulation, exploration, and visualization. Here's a breakdown of the essential libraries:

 1. NumPy

   - Purpose: Numerical computing.

   - Why Learn It: NumPy provides support for large multi-dimensional arrays and matrices, along with a collection of mathematical functions to operate on these arrays efficiently.

   - Key Features:

     - Fast array processing.

     - Mathematical operations on arrays (e.g., sum, mean, standard deviation).

     - Linear algebra operations.

 2. Pandas

   - Purpose: Data manipulation and analysis.

   - Why Learn It: Pandas offers data structures like DataFrames, making it easier to handle and analyze structured data.

   - Key Features:

     - Reading/writing data from CSV, Excel, SQL databases, and more.

     - Handling missing data.

     - Powerful group-by operations.

     - Data filtering and transformation.

 3. Matplotlib

   - Purpose: Data visualization.

   - Why Learn It: Matplotlib is one of the most widely used plotting libraries in Python, allowing for a wide range of static, animated, and interactive plots.

   - Key Features:

     - Line plots, bar charts, histograms, scatter plots.

     - Customizable charts (labels, colors, legends).

     - Integration with Pandas for quick plotting.

 4. Seaborn

   - Purpose: Statistical data visualization.

   - Why Learn It: Built on top of Matplotlib, Seaborn simplifies the creation of attractive and informative statistical graphics.

   - Key Features:

     - High-level interface for drawing attractive statistical graphics.

     - Easier to use for complex visualizations like heatmaps, pair plots, etc.

     - Visualizations based on categorical data.

 5. SciPy

   - Purpose: Scientific and technical computing.

   - Why Learn It: SciPy builds on NumPy and provides additional functionality for complex mathematical operations and scientific computing.

   - Key Features:

     - Optimized algorithms for numerical integration, optimization, and more.

     - Statistics, signal processing, and linear algebra modules.

 6. Scikit-learn

   - Purpose: Machine learning and statistical modeling.

   - Why Learn It: Scikit-learn provides simple and efficient tools for data mining, analysis, and machine learning.

   - Key Features:

     - Classification, regression, and clustering algorithms.

     - Dimensionality reduction, model selection, and preprocessing utilities.

 7. Statsmodels

   - Purpose: Statistical analysis.

   - Why Learn It: Statsmodels allows users to explore data, estimate statistical models, and perform tests.

   - Key Features:

     - Linear regression, logistic regression, time series analysis.

     - Statistical tests and models for descriptive statistics.

 8. Plotly

   - Purpose: Interactive data visualization.

   - Why Learn It: Plotly allows for the creation of interactive and web-based visualizations, making it ideal for dashboards and presentations.

   - Key Features:

     - Interactive plots like scatter, line, bar, and 3D plots.

     - Easy integration with web frameworks.

     - Dashboards and web applications with Dash.

 9. TensorFlow/PyTorch (Optional)

   - Purpose: Machine learning and deep learning.

   - Why Learn It: If your data analysis involves machine learning, these libraries will help in building, training, and deploying deep learning models.

   - Key Features:

     - Tensor processing and automatic differentiation.

     - Building neural networks.

 10. Dask (Optional)

   - Purpose: Parallel computing for data analysis.

   - Why Learn It: Dask enables scalable data manipulation by parallelizing Pandas operations, making it ideal for big datasets.

   - Key Features:

     - Works with NumPy, Pandas, and Scikit-learn.

     - Handles large data and parallel computations easily.

Focusing on NumPy, Pandas, Matplotlib, and Seaborn will set a strong foundation for basic data analysis.

[cerulity]

The two problems i have with calling images generated by machine learning "art".

Art is expression through a medium, and by that definition, machine learning cannot generate art. There is no cognitive, psychological or emotional difference between Midjourney and Control Panel. It's all manipulation of numbers, incapable of emotion or expression.

Machine learning will reference existing art. If those artists haven't consented to their art being used, that's not right. Collaging is a form of art because it uses human creativity to combine many different pieces into one. However, machine learning is basically just another form of tracing. It has no creativity.

So everyone that makes digital art and hasn't heard, there are two tools developed by the University of Chicago called Glaze and Nightshade.

Glaze is a tool to defend against neural networks mimicking art styles. It causes the neural networks to output weird and mangled versions of what would be generated otherwise: https://glaze.cs.uchicago.edu/downloads.html

Nightshade is a tool to 'poison' datasets and confuse neural networks into generating incorrect images. It turns hats into cakes: https://nightshade.cs.uchicago.edu/downloads.html

Use these to defend against models that steal art. Remember, they are taking your art without asking for permission, and it's their fault if it screws up their models. It's like Mark Rober's glitter bomb. It's not his fault the package was trapped, why should it be your fault when your Nightshaded art poisons a dataset?