social.sokoll.com

Search

Items tagged with: AI

"Scientists create online games to show risks of AI emotion recognition." "A team of researchers have created a website -- emojify.info -- where the public can try out emotion recognition systems through their own computer cameras. One game focuses on pulling faces to trick the technology, while another explores how such systems can struggle to read facial expressions in context."

"Alexa Hagerty, project lead and researcher at the University of Cambridge Leverhulme Centre for the Future of Intelligence and the Centre for the Study of Existential Risk, said many people were not aware how common emotion recognition systems were, noting they were employed in situations ranging from job hiring, to customer insight work, airport security, and even education to see if students are engaged or doing their homework."

Scientists create online games to show risks of AI emotion recognition

#solidstatelife #ai #computervision #emotionrecognition
 
"Scientists create online games to show risks of AI emotion recognition." "A team of researchers have created a website -- emojify.info -- where the public can try out emotion recognition systems through their own computer cameras. One game focuses on pulling faces to trick the technology, while another explores how such systems can struggle to read facial expressions in context."

"Alexa Hagerty, project lead and researcher at the University of Cambridge Leverhulme Centre for the Future of Intelligence and the Centre for the Study of Existential Risk, said many people were not aware how common emotion recognition systems were, noting they were employed in situations ranging from job hiring, to customer insight work, airport security, and even education to see if students are engaged or doing their homework."

Scientists create online games to show risks of AI emotion recognition

#solidstatelife #ai #computervision #emotionrecognition
 
AI photo editing: 3 new things an AI can do with your photos. 1. Change car geometry, 2. Repaint paintings, 3. Change facial expressions. He introduces these with a preamble about how you can change generated images if you can change the "latent space" inputs that the generative neural network uses, but the dimensions don't have any "interpretable" meaning -- or rather, they didn't used to, but now they do, and that's what makes all these new capabilities possible.



#solidstatelife #ai #computervision #generativenetworks #gans
 
AI photo editing: 3 new things an AI can do with your photos. 1. Change car geometry, 2. Repaint paintings, 3. Change facial expressions. He introduces these with a preamble about how you can change generated images if you can change the "latent space" inputs that the generative neural network uses, but the dimensions don't have any "interpretable" meaning -- or rather, they didn't used to, but now they do, and that's what makes all these new capabilities possible.



#solidstatelife #ai #computervision #generativenetworks #gans
 
"Furious AI researcher creates a list of non-reproducible machine learning papers." "I just spent a week implementing a paper as a baseline and failed to reproduce the results. I realized today after googling for a bit that a few others were also unable to reproduce the results. Is there a list of such papers? It will save people a lot of time and effort."

"'Easier to compile a list of reproducible ones...,' one user responded."

"The next day, ContributionSecure14 created 'Papers Without Code,' a website that aims to create a centralized list of machine learning papers that are not implementable."

"This will give the authors a chance to either release their code, provide pointers or rescind the paper."

Furious AI researcher creates a list of non-reproducible machine learning papers

#solidstatelife #ai #reproducibility
 
"Furious AI researcher creates a list of non-reproducible machine learning papers." "I just spent a week implementing a paper as a baseline and failed to reproduce the results. I realized today after googling for a bit that a few others were also unable to reproduce the results. Is there a list of such papers? It will save people a lot of time and effort."

"'Easier to compile a list of reproducible ones...,' one user responded."

"The next day, ContributionSecure14 created 'Papers Without Code,' a website that aims to create a centralized list of machine learning papers that are not implementable."

"This will give the authors a chance to either release their code, provide pointers or rescind the paper."

Furious AI researcher creates a list of non-reproducible machine learning papers

#solidstatelife #ai #reproducibility
 
"Furious AI researcher creates a list of non-reproducible machine learning papers." "I just spent a week implementing a paper as a baseline and failed to reproduce the results. I realized today after googling for a bit that a few others were also unable to reproduce the results. Is there a list of such papers? It will save people a lot of time and effort."

"'Easier to compile a list of reproducible ones...,' one user responded."

"The next day, ContributionSecure14 created 'Papers Without Code,' a website that aims to create a centralized list of machine learning papers that are not implementable."

"This will give the authors a chance to either release their code, provide pointers or rescind the paper."

Furious AI researcher creates a list of non-reproducible machine learning papers

#solidstatelife #ai #reproducibility
 
Maia: the human-like neural chess engine. Instead of trying to choose the optimal move, Maia tries to choose the move a human will play. "We trained 9 versions of Maia, one for each rating milestone between 1100 and 1900, on over a total of 100 million Lichess games between human players. Each Maia captures human style in chess at its targeted level: Maia 1100 is most predictive of human play around the 1100 level, and Maia 1900 is most predictive of human play around the 1900 level."

"Note that each Maia plays at a level above the rating range it was trained on, for an interesting reason: Maia's goal is to make the average move that players at its target level would make. Playing Maia 1100, for example, is more like playing a committee of 1100-rated players than playing a single 1100-rated player -- they tend to average out their idiosyncratic mistakes (but Maia still makes a lot of human-like blunders!)."

"Because we trained 9 different versions of Maia, each at a targeted skill level, we can begin to algorithmically capture what kinds of mistakes players at specific skill levels make -- and when people stop making them." "Maia could look at your games and tell you which blunders were predictable, and which were more random mistakes. Guidance like this could be valuable for players trying to improve: If you repeatedly make a predictable kind of mistake, you know what you need to work on to hit the next level."

The article doesn't say how the chess engine works, but it's using the code from Leela Chess, which in turn is an open-source implementation of DeepMind's AlphaZero system, the code of which has never been released to the public. Leela Chess is normally trained by self-play, and what they're doing here is training it on human games instead.

Introducing Maia, a human-like neural network chess engine

#solidstatelife #ai #chess
 
Maia: the human-like neural chess engine. Instead of trying to choose the optimal move, Maia tries to choose the move a human will play. "We trained 9 versions of Maia, one for each rating milestone between 1100 and 1900, on over a total of 100 million Lichess games between human players. Each Maia captures human style in chess at its targeted level: Maia 1100 is most predictive of human play around the 1100 level, and Maia 1900 is most predictive of human play around the 1900 level."

"Note that each Maia plays at a level above the rating range it was trained on, for an interesting reason: Maia's goal is to make the average move that players at its target level would make. Playing Maia 1100, for example, is more like playing a committee of 1100-rated players than playing a single 1100-rated player -- they tend to average out their idiosyncratic mistakes (but Maia still makes a lot of human-like blunders!)."

"Because we trained 9 different versions of Maia, each at a targeted skill level, we can begin to algorithmically capture what kinds of mistakes players at specific skill levels make -- and when people stop making them." "Maia could look at your games and tell you which blunders were predictable, and which were more random mistakes. Guidance like this could be valuable for players trying to improve: If you repeatedly make a predictable kind of mistake, you know what you need to work on to hit the next level."

The article doesn't say how the chess engine works, but it's using the code from Leela Chess, which in turn is an open-source implementation of DeepMind's AlphaZero system, the code of which has never been released to the public. Leela Chess is normally trained by self-play, and what they're doing here is training it on human games instead.

Introducing Maia, a human-like neural network chess engine

#solidstatelife #ai #chess
 
"Can a fruit fly learn word embeddings?" Don't worry, this time no fruit flies were hurt in the process of doing this research. In fact no fruit flies were used at all. What they did here was make a 'fruit fly brain-inspired' neural network.

Apparently in the fruit fly brain, there's a pair of parts called "mushroom bodies" that has about 2,000 Kenyon cells in each one. This is a key area of the fruit fly brain's sensory input processing, input from not just sight and hearing but smell (which is actually the biggest source of input), temperature, and humidity. Mushroom bodies are something that fly brains have that human brains don't have -- only insects, other arthropods, and some annelids have them. Kenyon cells, named after F. C. Kenyon, are cells specific to the mushroom bodies that creatures such as us don't have.

The researchers replaced the usual recurrent neural networks used to make word embeddings with their own inspired by the structure of Kenyon cells, which results in a much sparser network than usual. They chose the word embedding task because it is an "unsupervised" task -- all they have to do is shovel huge amounts of text into it, but they don't have to create by hand any list of matching sets of input with the "correct" output the neural network has to learn.

The output of the neural network is sparse binary hash codes. Which is to say, instead of a list of real numbers (vector of floating-point values), it outputs a list of true/false values, and because it is sparse, most of them are false.

They tested the network on 4 tasks: static word embeddings, word clustering, context-dependent word embeddings, and document classification. The "static" test just compares the semantic similarity of word embeddings with human-generated scores. The "clustering" task uses a clustering algorithm to group words into clusters, and compares this with clustering word embeddings from traditional algorithms (like Word2Vec and GloVe) the same way. The "context-dependent" test challenges the neural network to distinguish between different meanings of the same word using a dataset designed for that purpose, for example distinguish between a bank account and a river bank, and an Apple iPhone vs an apple pie. The "document classification" task involves taking news articles and putting each one in one of 35 categories. The "fruit fly network" performed comparably to regular word embedding systems on all these tasks.

In addition to performing as well with sparse binary hash codes, the system is much more computationally efficient. Training takes only a few hours, vs 24 hours for GloVe, and even more for newer models like BERTBASE.

Can a fruit fly learn word embeddings?

#solidstatelife #ai #biomimicry
 
"Can a fruit fly learn word embeddings?" Don't worry, this time no fruit flies were hurt in the process of doing this research. In fact no fruit flies were used at all. What they did here was make a 'fruit fly brain-inspired' neural network.

Apparently in the fruit fly brain, there's a pair of parts called "mushroom bodies" that has about 2,000 Kenyon cells in each one. This is a key area of the fruit fly brain's sensory input processing, input from not just sight and hearing but smell (which is actually the biggest source of input), temperature, and humidity. Mushroom bodies are something that fly brains have that human brains don't have -- only insects, other arthropods, and some annelids have them. Kenyon cells, named after F. C. Kenyon, are cells specific to the mushroom bodies that creatures such as us don't have.

The researchers replaced the usual recurrent neural networks used to make word embeddings with their own inspired by the structure of Kenyon cells, which results in a much sparser network than usual. They chose the word embedding task because it is an "unsupervised" task -- all they have to do is shovel huge amounts of text into it, but they don't have to create by hand any list of matching sets of input with the "correct" output the neural network has to learn.

The output of the neural network is sparse binary hash codes. Which is to say, instead of a list of real numbers (vector of floating-point values), it outputs a list of true/false values, and because it is sparse, most of them are false.

They tested the network on 4 tasks: static word embeddings, word clustering, context-dependent word embeddings, and document classification. The "static" test just compares the semantic similarity of word embeddings with human-generated scores. The "clustering" task uses a clustering algorithm to group words into clusters, and compares this with clustering word embeddings from traditional algorithms (like Word2Vec and GloVe) the same way. The "context-dependent" test challenges the neural network to distinguish between different meanings of the same word using a dataset designed for that purpose, for example distinguish between a bank account and a river bank, and an Apple iPhone vs an apple pie. The "document classification" task involves taking news articles and putting each one in one of 35 categories. The "fruit fly network" performed comparably to regular word embedding systems on all these tasks.

In addition to performing as well with sparse binary hash codes, the system is much more computationally efficient. Training takes only a few hours, vs 24 hours for GloVe, and even more for newer models like BERTBASE.

Can a fruit fly learn word embeddings?

#solidstatelife #ai #biomimicry
 
"Towards fully automated manga translation." "Due to the high cost of translation, most comics have not been translated and are only available in their domestic markets. What if all comics could be immediately translated into any language?"

"What makes the translation of comics difficult?" Utterances by a character are divided up into multiple bubbles, interlaced with bubbles from other characters. They are not necessarily aligned in a straightforward way such as from left to right or right to left or top to bottom. It is necessary to figure out the visual structure just to figure out what the correct order is to arrange the bubbles. Not only that, but the meaning of the pictures has to be deciphered to do a correct translation, for example, some Japanese words can be translated into both "he", "him", "she", or "her", and without understanding the pictures, these ambiguities can't be resolved.

So what these researchers did was make a manga translation dataset which is a massive collection of "parallel corpus", that is, manga that have already been translated between languages, with the translated versions paired up with the original, that an AI system can train from, as well as a system for extracting pictures and text from the corpus. From this they built a "multimodal context-aware translation system" that was able to fully automate the manga translation process. By "multimodal", they mean the same system takes input in the form of text and pictures (so two "modes"), and by "context-aware", they mean the language translation takes into account the context provided by the pictures.

The translation system receives a manga image and a set of texts. The image represents a whole page, not just a single frame, and the system will figure out whether all the frames represent the same scene, or whether there are multiple scenes. Each of the texts has a "bounding box" that indicates where it came from in the image. A convolutional neural network called Faster R-CNN is used to determine what scene or scenes are in the frames.

Next the reading order is determined. This is first estimated by a simple algorithm that assumes frames are in rows, and within each row, frames are in columns, with a right-to-left (for Japanese), then top-to-bottom reading order. This gets the reading order about 92% correct. The remaining 8% irregular cases are handled by a supervised learning algorithm.

Next a system called illustration2vec is used to create a vector (this is the "embedding" idea though they don't use the word) describing the scene. The output of this is tags like "boy", "girl".

To do the actual translation, the input text is all concatenated together, translated all at once, and then broken back apart into the proper speech bubbles on the other end. To make this work, though, tags from illustration2vec are inserted into the text, which provide context to the translation system.

All of this glosses over additional complex issues that the researchers had to solve to make this work, such as identifying where text is in an image, figure out what style the text is written in, since text in manga is written in various styles, and extract the text.

Towards Fully Automated Manga Translation

#solidstatelife #ai #computervision #nlp #translation #embeddings #manga
 
"Towards fully automated manga translation." "Due to the high cost of translation, most comics have not been translated and are only available in their domestic markets. What if all comics could be immediately translated into any language?"

"What makes the translation of comics difficult?" Utterances by a character are divided up into multiple bubbles, interlaced with bubbles from other characters. They are not necessarily aligned in a straightforward way such as from left to right or right to left or top to bottom. It is necessary to figure out the visual structure just to figure out what the correct order is to arrange the bubbles. Not only that, but the meaning of the pictures has to be deciphered to do a correct translation, for example, some Japanese words can be translated into both "he", "him", "she", or "her", and without understanding the pictures, these ambiguities can't be resolved.

So what these researchers did was make a manga translation dataset which is a massive collection of "parallel corpus", that is, manga that have already been translated between languages, with the translated versions paired up with the original, that an AI system can train from, as well as a system for extracting pictures and text from the corpus. From this they built a "multimodal context-aware translation system" that was able to fully automate the manga translation process. By "multimodal", they mean the same system takes input in the form of text and pictures (so two "modes"), and by "context-aware", they mean the language translation takes into account the context provided by the pictures.

The translation system receives a manga image and a set of texts. The image represents a whole page, not just a single frame, and the system will figure out whether all the frames represent the same scene, or whether there are multiple scenes. Each of the texts has a "bounding box" that indicates where it came from in the image. A convolutional neural network called Faster R-CNN is used to determine what scene or scenes are in the frames.

Next the reading order is determined. This is first estimated by a simple algorithm that assumes frames are in rows, and within each row, frames are in columns, with a right-to-left (for Japanese), then top-to-bottom reading order. This gets the reading order about 92% correct. The remaining 8% irregular cases are handled by a supervised learning algorithm.

Next a system called illustration2vec is used to create a vector (this is the "embedding" idea though they don't use the word) describing the scene. The output of this is tags like "boy", "girl".

To do the actual translation, the input text is all concatenated together, translated all at once, and then broken back apart into the proper speech bubbles on the other end. To make this work, though, tags from illustration2vec are inserted into the text, which provide context to the translation system.

All of this glosses over additional complex issues that the researchers had to solve to make this work, such as identifying where text is in an image, figure out what style the text is written in, since text in manga is written in various styles, and extract the text.

Towards Fully Automated Manga Translation

#solidstatelife #ai #computervision #nlp #translation #embeddings #manga
 
"AI-supported test predicts eye disease three years before symptoms." The test "involves injecting into the bloodstream (via the arm) a fluorescent dye that attaches to retinal cells, and illuminates those that are undergoing stress or in the process of apoptosis, a form of programmed cell death. The damaged cells appear bright white when viewed in eye examinations -- the more damaged cells detected, the higher the Detection of Apoptosing Retinal Cells (DARC) count."

If you're wondering where AI comes in, they're using a convolutional neural network to do the counting. The process gives highly consistent results, which is not necessarily the case with humans.

"Our new test was able to predict new Wet age-related macular degeneration (wet AMD) lesions up to 36 months in advance of them occurring and that is huge -- it means that DARC activity can guide a clinician into treating more intensively those patients who are at high risk of new lesions of wet AMD and also be used as a screening tool."

AI-supported test predicts eye disease three years before symptoms

#solidstatelife #ai #medicalai
 
"AI-supported test predicts eye disease three years before symptoms." The test "involves injecting into the bloodstream (via the arm) a fluorescent dye that attaches to retinal cells, and illuminates those that are undergoing stress or in the process of apoptosis, a form of programmed cell death. The damaged cells appear bright white when viewed in eye examinations -- the more damaged cells detected, the higher the Detection of Apoptosing Retinal Cells (DARC) count."

If you're wondering where AI comes in, they're using a convolutional neural network to do the counting. The process gives highly consistent results, which is not necessarily the case with humans.

"Our new test was able to predict new Wet age-related macular degeneration (wet AMD) lesions up to 36 months in advance of them occurring and that is huge -- it means that DARC activity can guide a clinician into treating more intensively those patients who are at high risk of new lesions of wet AMD and also be used as a screening tool."

AI-supported test predicts eye disease three years before symptoms

#solidstatelife #ai #medicalai
 
The conclusions of a lot of EEG and fMRI experiments may be invalid. Ok, so the key to understanding this is the concepts of "temporal autocorrelation" and "block-level effects". Correlation is when two things change in the same way at the same time, more or less. Autocorrelation is when you compare something at one time point to the same thing at another time point. If you think something might be random and run an autocorrelation calculation and it shows there's a correlation, it means your variable isn't random. EEG and fMRI time series show temporal autocorrelations in both the short and long term regardless of what's going on in the experiment.

As for "block-level effects", when doing the statistical analysis on experiments, blocking is the grouping of data into blocks, for example male and female blocks.

With that preamble in mind, they pick on a recent research experiment that claims to make a neural network that learns to classify EEG data according to what ImageNet image the person is looking at. "The Purdue researchers originally began questioning the dataset when they could not obtain similar outcomes from their own tests. That's when they started analyzing the previous results and determined that a lack of randomization contaminated the dataset."

"The classifier employed makes extensive, if not sole, use of long-term static brain activity that persists much longer than the duration of individual stimuli. Since the paper employs a block design, where all stimuli of a given class are presented to a subject in succession, the classifiers employed tend to classify the brain activity during that block, which appears to be largely uncorrelated with stimulus class. This is exacerbated by the reliance of the classifier on DC and very-low frequency components in the EEG signal that reflect arbitrary long-term static mental states during a block rather than dynamic brain activity. Since each trial in the test sets employed comes from the same block as many trials in the corresponding training sets, the reported high classification accuracy results from classifying arbitrary temporal artifacts of the data instead of stimulus-related activity. When the experiment is repeated with a rapid-event design, where stimuli of different classes are randomly intermixed, classification accuracy drops to chance."

"As a result, this renders suspect all of the results and claims advanced in multiple published papers. Our experiments suggest that the underlying tasks are far more difficult than they appear on the surface and are far beyond the current state of the art."

Purdue researchers uncover blind spots at the intersection of AI and neuroscience

#solidstatelife #ai #discoveries #neuroscience #eeg #fmri #statistics
 
The conclusions of a lot of EEG and fMRI experiments may be invalid. Ok, so the key to understanding this is the concepts of "temporal autocorrelation" and "block-level effects". Correlation is when two things change in the same way at the same time, more or less. Autocorrelation is when you compare something at one time point to the same thing at another time point. If you think something might be random and run an autocorrelation calculation and it shows there's a correlation, it means your variable isn't random. EEG and fMRI time series show temporal autocorrelations in both the short and long term regardless of what's going on in the experiment.

As for "block-level effects", when doing the statistical analysis on experiments, blocking is the grouping of data into blocks, for example male and female blocks.

With that preamble in mind, they pick on a recent research experiment that claims to make a neural network that learns to classify EEG data according to what ImageNet image the person is looking at. "The Purdue researchers originally began questioning the dataset when they could not obtain similar outcomes from their own tests. That's when they started analyzing the previous results and determined that a lack of randomization contaminated the dataset."

"The classifier employed makes extensive, if not sole, use of long-term static brain activity that persists much longer than the duration of individual stimuli. Since the paper employs a block design, where all stimuli of a given class are presented to a subject in succession, the classifiers employed tend to classify the brain activity during that block, which appears to be largely uncorrelated with stimulus class. This is exacerbated by the reliance of the classifier on DC and very-low frequency components in the EEG signal that reflect arbitrary long-term static mental states during a block rather than dynamic brain activity. Since each trial in the test sets employed comes from the same block as many trials in the corresponding training sets, the reported high classification accuracy results from classifying arbitrary temporal artifacts of the data instead of stimulus-related activity. When the experiment is repeated with a rapid-event design, where stimuli of different classes are randomly intermixed, classification accuracy drops to chance."

"As a result, this renders suspect all of the results and claims advanced in multiple published papers. Our experiments suggest that the underlying tasks are far more difficult than they appear on the surface and are far beyond the current state of the art."

Purdue researchers uncover blind spots at the intersection of AI and neuroscience

#solidstatelife #ai #discoveries #neuroscience #eeg #fmri #statistics
 
A clever trick to defeat face recognition systems by in essence using adversarial examples in reverse. The idea is, before you upload photos to social media (for example), you run them through this software. You can actually download and run the software yourself (Windows, Mac, & Linux).

The software makes tiny, invisible-to-the-human-eye perturbations to the image that, when fed into a facial recognition system, will learn the wrong thing so when your face is seen in a face recognition system in a CCTV camera (for example), it will fail to recognize you are you.

This takes advantage of the fact that computer vision doesn't work the same as human vision. When you look at someone, the way your brain determines who the person is is completely different from the way computers make that determination. Your brain relies much more on the overall shape of the person's face and how the parts fit together while computers rely more on tiny details and are less able to use large-scale structure. When a facial recognition system is trained on images, it extracts and stores an array of numbers that represent "features" of your face (though what exactly the numbers mean, no one knows because it is all learned automatically by deep learning systems, rather than programmed by humans). Future pictures of you can be run through the same process and if the "features" match the previously stored "features", the face recognition system will say it's you.

The idea here is to take advantage of the difference in how machines and humans recognize faces to trick the system to store the wrong "features" in the first place. All while making your photo look the same to other actual humans.

The way the system works is it takes your photo and someone else's photo as a "target" photo. Then what the system tries to do is optimize for the combination of a) getting a "feature" set as similar as possible to the target photo, while at the same time b) minimizing the changes to the pixels in your photo.

The researchers tested the system against three state-of-the-art facial recognition systems: Microsoft Azure Face API), Amazon Rekognition, and Face++. Amazon Rekognition is used by the US Immigration and Customs Enforcement Agency (ICE), law enforcement in Florida and Oregon, as well as various large corporations such as CBS. Face++ is a Chinese system produced by Megvii (supposed to make you think "megavision" in English -- the company is called Kuàngshì in China) and is widely deployed in China. These systems fail 100% of the time to recognize the right person when trained on "cloaked" photos and then tested on "uncloaked" photos.

If you don't have the ability to cloak all the images of you that could be scraped off social networking sites and other internet sites, the testing showed that even if only half of the training images are "cloaked", the state-of-the-art commercial systems still failed to recognize the right person 80% of the time.

The researchers anticipated the face recognition companies will try to develop countermeasures. A simple technique is to simply apply a Gaussian blur or add noise to an image before feeding it to the training system, but in tests this reduced the failure rate from 100% to 98%, so it actually didn't work. Applying extra image compression was a bit more effective, with failure rate going below 90% when then compression was turned down to 5% -- but turning the compression quality down that low causes the normal classification accuracy to drop precipitously (below 70%).

Techniques designed to counter data "poisoning" attacks only work if outliers can be detected in the training data. If you succeed in getting all or nearly all the training images to be "cloaked", you will move the "baseline" and these systems will be unable to detect the cloaking as an "outlier".

If the training system has access to both "cloaked" and "uncloaked" images of you, then there is a possibility to detect what is going on by using a clustering algorithm to discover that your input images have two distinct feature clusters. The only way to mitigate is to choose a target with smaller "feature" separation.

Fawkes: Image "Cloaking" for Personal Privacy

#solidstatelife #ai #computervision #adversarialexamples
 
A clever trick to defeat face recognition systems by in essence using adversarial examples in reverse. The idea is, before you upload photos to social media (for example), you run them through this software. You can actually download and run the software yourself (Windows, Mac, & Linux).

The software makes tiny, invisible-to-the-human-eye perturbations to the image that, when fed into a facial recognition system, will learn the wrong thing so when your face is seen in a face recognition system in a CCTV camera (for example), it will fail to recognize you are you.

This takes advantage of the fact that computer vision doesn't work the same as human vision. When you look at someone, the way your brain determines who the person is is completely different from the way computers make that determination. Your brain relies much more on the overall shape of the person's face and how the parts fit together while computers rely more on tiny details and are less able to use large-scale structure. When a facial recognition system is trained on images, it extracts and stores an array of numbers that represent "features" of your face (though what exactly the numbers mean, no one knows because it is all learned automatically by deep learning systems, rather than programmed by humans). Future pictures of you can be run through the same process and if the "features" match the previously stored "features", the face recognition system will say it's you.

The idea here is to take advantage of the difference in how machines and humans recognize faces to trick the system to store the wrong "features" in the first place. All while making your photo look the same to other actual humans.

The way the system works is it takes your photo and someone else's photo as a "target" photo. Then what the system tries to do is optimize for the combination of a) getting a "feature" set as similar as possible to the target photo, while at the same time b) minimizing the changes to the pixels in your photo.

The researchers tested the system against three state-of-the-art facial recognition systems: Microsoft Azure Face API), Amazon Rekognition, and Face++. Amazon Rekognition is used by the US Immigration and Customs Enforcement Agency (ICE), law enforcement in Florida and Oregon, as well as various large corporations such as CBS. Face++ is a Chinese system produced by Megvii (supposed to make you think "megavision" in English -- the company is called Kuàngshì in China) and is widely deployed in China. These systems fail 100% of the time to recognize the right person when trained on "cloaked" photos and then tested on "uncloaked" photos.

If you don't have the ability to cloak all the images of you that could be scraped off social networking sites and other internet sites, the testing showed that even if only half of the training images are "cloaked", the state-of-the-art commercial systems still failed to recognize the right person 80% of the time.

The researchers anticipated the face recognition companies will try to develop countermeasures. A simple technique is to simply apply a Gaussian blur or add noise to an image before feeding it to the training system, but in tests this reduced the failure rate from 100% to 98%, so it actually didn't work. Applying extra image compression was a bit more effective, with failure rate going below 90% when then compression was turned down to 5% -- but turning the compression quality down that low causes the normal classification accuracy to drop precipitously (below 70%).

Techniques designed to counter data "poisoning" attacks only work if outliers can be detected in the training data. If you succeed in getting all or nearly all the training images to be "cloaked", you will move the "baseline" and these systems will be unable to detect the cloaking as an "outlier".

If the training system has access to both "cloaked" and "uncloaked" images of you, then there is a possibility to detect what is going on by using a clustering algorithm to discover that your input images have two distinct feature clusters. The only way to mitigate is to choose a target with smaller "feature" separation.

Fawkes: Image "Cloaking" for Personal Privacy

#solidstatelife #ai #computervision #adversarialexamples
 
The polysemy problem for word embeddings. Word embeddings, you'll recall, are those high-dimensional vectors that represent words that are called "embeddings" for some reason I was told once but can never remember. (Obviously it's not intuitive at all). Anyway, the classic example of word embeddings is if you take the difference in vector representations between "king" and "queen" you get the same vector as you get with "man" and "woman". So you see the idea is that the word embeddings contain a representation of the "meaning" that the word has and how that meaning relates to other words.

The problem is words don't have only one meaning. This is called "polysemy". If you're not familiar with the word "polysemy" but are familiar with "synonyms", "polysemy" is the opposite of "synonyms": with synonyms, multiple words have the same meaning, but with polysemy, multiple meanings have the same word.

Consider, for example, the word "mole", which this machine learning system identified as having 4 meanings: one is associated with words like "counterspy", "spy", "espionage", etc, one associated with "beautymark", "birthmark", "nevus", "pigment", "skin", etc, one associated with "mol", "unit", "gram", "molecule", etc, and one associated with "talpidae", "nocturnal", "digging", "mammal", etc.

Ok, so obviously they have a machine learning system to identify this -- how does it work? Believe it or not it comes from the mathematical field of topology. Basically they hypothesize that meanings lie on a manifold in "meaning space", which can't be directly observed, and words lie on a manifold in "word space", which can be observed. A polysemous word represents taking the manifold in "meaning space" and "pinching" it. Thus, they conjured up an algorithm to detect these "pinchings" in the manifold in word space.

The algorithm relies on something called topological data analysis, which is a technique whereby you take a bunch of vectors, which are points in a high-dimensional space, and figure out the most likely topological structure that would encompass those points.

To do this, they took a popular word embedding dataset called fastText, and did this topological data analysis on it. To detect the pinchings, they used something called Wasserstein distance. Wasserstein distance provides a notion of the distance between the topological structures outputted by the topological data analysis. They look at the topology in the neighborhood of a word, but excluding the word itself, and compare it with a non-pinched topology. The difference reveals how many polysemous meanings the word has.

To verify the system, they used a test devised in 2010 to test machine learning systems' ability to handle polysemy called the SemEval-2010 task on Word Sense Induction & Disambiguation. The task involves 8,915 sentences from various news sources like CNN and ABC of 100 different polysemous target words, 50 nouns and 50 verbs, and the goal is to cluster them based on context -- so all the instances with the same meaning get put in the same cluster but different meanings get put in a different cluster. The task comes with a training set with 65 million occurrences of 127,151 different words, and they used this corpus to produce their own embeddings using the fastText algorithm. When using the 10 nearest words to determine topology, their system correlated tightly with the "gold standard" for the test, the correct answers created by humans. When using more than the 10 nearest words, their system tended to find more meanings, and the correlation went down.

#solidstatelife #ai #nlp #embeddings #topology
 
The polysemy problem for word embeddings. Word embeddings, you'll recall, are those high-dimensional vectors that represent words that are called "embeddings" for some reason I was told once but can never remember. (Obviously it's not intuitive at all). Anyway, the classic example of word embeddings is if you take the difference in vector representations between "king" and "queen" you get the same vector as you get with "man" and "woman". So you see the idea is that the word embeddings contain a representation of the "meaning" that the word has and how that meaning relates to other words.

The problem is words don't have only one meaning. This is called "polysemy". If you're not familiar with the word "polysemy" but are familiar with "synonyms", "polysemy" is the opposite of "synonyms": with synonyms, multiple words have the same meaning, but with polysemy, multiple meanings have the same word.

Consider, for example, the word "mole", which this machine learning system identified as having 4 meanings: one is associated with words like "counterspy", "spy", "espionage", etc, one associated with "beautymark", "birthmark", "nevus", "pigment", "skin", etc, one associated with "mol", "unit", "gram", "molecule", etc, and one associated with "talpidae", "nocturnal", "digging", "mammal", etc.

Ok, so obviously they have a machine learning system to identify this -- how does it work? Believe it or not it comes from the mathematical field of topology. Basically they hypothesize that meanings lie on a manifold in "meaning space", which can't be directly observed, and words lie on a manifold in "word space", which can be observed. A polysemous word represents taking the manifold in "meaning space" and "pinching" it. Thus, they conjured up an algorithm to detect these "pinchings" in the manifold in word space.

The algorithm relies on something called topological data analysis, which is a technique whereby you take a bunch of vectors, which are points in a high-dimensional space, and figure out the most likely topological structure that would encompass those points.

To do this, they took a popular word embedding dataset called fastText, and did this topological data analysis on it. To detect the pinchings, they used something called Wasserstein distance. Wasserstein distance provides a notion of the distance between the topological structures outputted by the topological data analysis. They look at the topology in the neighborhood of a word, but excluding the word itself, and compare it with a non-pinched topology. The difference reveals how many polysemous meanings the word has.

To verify the system, they used a test devised in 2010 to test machine learning systems' ability to handle polysemy called the SemEval-2010 task on Word Sense Induction & Disambiguation. The task involves 8,915 sentences from various news sources like CNN and ABC of 100 different polysemous target words, 50 nouns and 50 verbs, and the goal is to cluster them based on context -- so all the instances with the same meaning get put in the same cluster but different meanings get put in a different cluster. The task comes with a training set with 65 million occurrences of 127,151 different words, and they used this corpus to produce their own embeddings using the fastText algorithm. When using the 10 nearest words to determine topology, their system correlated tightly with the "gold standard" for the test, the correct answers created by humans. When using more than the 10 nearest words, their system tended to find more meanings, and the correlation went down.

#solidstatelife #ai #nlp #embeddings #topology
 
Waymo report "shows an incredible, superhuman safety record and suggests it is past time for them to deploy a service at scale, at least in simpler low-urban/suburban zones like Chandler, Arizona," says Brad Templeton.

"In 6.1 million miles they report 30 'dings' with no injury expected, 9 with 10% chance of injury and 8 with airbag deployment but still 10% chance of injury, suggesting less than 2 modest injuries. Human drivers would have had about 6."

"Nationally, 6.1 million miles of driving by a good driver should result in about 40-60 events, most of which are small dings, 22-27 or which would involve an insurance claim, 12 which would get reported to police and 6 injury crashes. With no at-fault events in 8 lifetimes of human driving, Waymo's performance is significantly superior to a human, even in an easy place like Chandler."

Waymo Data Shows Superhuman Safety Record. They Should Deploy Today

#solidstatelife #ai #autonomousvehicles #waymo
 
Waymo report "shows an incredible, superhuman safety record and suggests it is past time for them to deploy a service at scale, at least in simpler low-urban/suburban zones like Chandler, Arizona," says Brad Templeton.

"In 6.1 million miles they report 30 'dings' with no injury expected, 9 with 10% chance of injury and 8 with airbag deployment but still 10% chance of injury, suggesting less than 2 modest injuries. Human drivers would have had about 6."

"Nationally, 6.1 million miles of driving by a good driver should result in about 40-60 events, most of which are small dings, 22-27 or which would involve an insurance claim, 12 which would get reported to police and 6 injury crashes. With no at-fault events in 8 lifetimes of human driving, Waymo's performance is significantly superior to a human, even in an easy place like Chandler."

Waymo Data Shows Superhuman Safety Record. They Should Deploy Today

#solidstatelife #ai #autonomousvehicles #waymo
 
"With neural filters, Photoshop can adjust a subject's age and facial expression, amplifying or reducing feelings like 'joy,' 'surprise,' or 'anger' with simple sliders. You can remove someone's glasses or smooth out their spots. One of the weirder filters even lets you transfer makeup from one person to another. And it's all done in just a few clicks, with the output easily tweaked or reversed entirely."

"Adobe is harnessing the power of generative adversarial networks -- or GANs -- a type of machine learning technique that's proved particularly adept at generating visual imagery. Some of the processing is done locally and some in the cloud, depending on the computational demands of each individual tool."

Photoshop's AI neural filters can tweak age and expression with a few clicks

#solidstatelife #ai #adobe #photoshop
 
"With neural filters, Photoshop can adjust a subject's age and facial expression, amplifying or reducing feelings like 'joy,' 'surprise,' or 'anger' with simple sliders. You can remove someone's glasses or smooth out their spots. One of the weirder filters even lets you transfer makeup from one person to another. And it's all done in just a few clicks, with the output easily tweaked or reversed entirely."

"Adobe is harnessing the power of generative adversarial networks -- or GANs -- a type of machine learning technique that's proved particularly adept at generating visual imagery. Some of the processing is done locally and some in the cloud, depending on the computational demands of each individual tool."

Photoshop's AI neural filters can tweak age and expression with a few clicks

#solidstatelife #ai #adobe #photoshop
 
The Guardian’s GPT-3-written article misleads readers about AI. Here’s why. – TechTalks https://bdtechtalks.com/2020/09/14/guardian-gpt-3-article-ai-fake-news/

The usage or claiming something is AI is usually a huge indicator for bullshit.
It's just an advanced text generator which stitches words together.

#gpt3 #ai
The Guardian’s GPT-3-written article misleads readers about AI. Here’s why.
#gpt3 #ai
 
Mmmm. AI "You Keep Using That Word, I Do Not Think It Means What You Think It Means"

These students figured out their tests were graded by AI — and the easy way to cheat

On Monday, Dana Simmons came downstairs to find her 12-year-old son, Lazare, in tears. He’d completed the first assignment for his seventh-grade history class on Edgenuity, an online platform for virtual learning. He’d received a 50 out of 100. That wasn’t on a practice test — it was his real grade.
“He was like, I’m gonna have to get a 100 on all the rest of this to make up for this,” said Simmons in a phone interview with The Verge. “He was totally dejected.”
At first, Simmons tried to console her son. “I was like well, you know, some teachers grade really harshly at the beginning,” said Simmons, who is a history professor herself. Then, Lazare clarified that he’d received his grade less than a second after submitting his answers. A teacher couldn’t have read his response in that time, Simmons knew — her son was being graded by an algorithm.
Simmons watched Lazare complete more assignments. She looked at the correct answers, which Edgenuity revealed at the end. She surmised that Edgenuity’s AI was scanning for specific keywords that it expected to see in students’ answers. And she decided to game it.
Now, for every short-answer question, Lazare writes two long sentences followed by a disjointed list of keywords — anything that seems relevant to the question. “The questions are things like... ‘What was the advantage of Constantinople’s location for the power of the Byzantine empire,’” Simmons says. “So you go through, okay, what are the possible keywords that are associated with this? Wealth, caravan, ship, India, China, Middle East, he just threw all of those words in.”
“I wanted to game it because I felt like it was an easy way to get a good grade,” Lazare told The Verge. He usually digs the keywords out of the article or video the question is based on.
Apparently, that “word salad” is enough to get a perfect grade on any short-answer question in an Edgenuity test.
#AI #education #hanginthere
 
Mmmm. AI "You Keep Using That Word, I Do Not Think It Means What You Think It Means"

These students figured out their tests were graded by AI — and the easy way to cheat

On Monday, Dana Simmons came downstairs to find her 12-year-old son, Lazare, in tears. He’d completed the first assignment for his seventh-grade history class on Edgenuity, an online platform for virtual learning. He’d received a 50 out of 100. That wasn’t on a practice test — it was his real grade.
“He was like, I’m gonna have to get a 100 on all the rest of this to make up for this,” said Simmons in a phone interview with The Verge. “He was totally dejected.”
At first, Simmons tried to console her son. “I was like well, you know, some teachers grade really harshly at the beginning,” said Simmons, who is a history professor herself. Then, Lazare clarified that he’d received his grade less than a second after submitting his answers. A teacher couldn’t have read his response in that time, Simmons knew — her son was being graded by an algorithm.
Simmons watched Lazare complete more assignments. She looked at the correct answers, which Edgenuity revealed at the end. She surmised that Edgenuity’s AI was scanning for specific keywords that it expected to see in students’ answers. And she decided to game it.
Now, for every short-answer question, Lazare writes two long sentences followed by a disjointed list of keywords — anything that seems relevant to the question. “The questions are things like... ‘What was the advantage of Constantinople’s location for the power of the Byzantine empire,’” Simmons says. “So you go through, okay, what are the possible keywords that are associated with this? Wealth, caravan, ship, India, China, Middle East, he just threw all of those words in.”
“I wanted to game it because I felt like it was an easy way to get a good grade,” Lazare told The Verge. He usually digs the keywords out of the article or video the question is based on.
Apparently, that “word salad” is enough to get a perfect grade on any short-answer question in an Edgenuity test.
#AI #education #hanginthere
 

This ‘Cloaking’ Algorithm Breaks Facial Recognition by Making Tiny Edits








A team of researchers at the University of Chicago have developed an algorithm that makes tiny, imperceptible edits to your images in order to mask you from facial recognition technology. Their invention is called Fawkes, and anybody can use it on their own images for free.

The algorithm was created by researchers in the SAND Lab at the University of Chicago, and the open-source software tool that they built is free to download and use on your computer at home.

The program works by making "tiny, pixel-level changes that are invisible to the human eye," but that nevertheless prevent facial recognition algorithms from categorizing you correctly. It's not so much that it makes you impossible to categorize; it's that the algorithm will categorize you as a different person entirely. The team calls the result "cloaked" photos, and they can be used like any other:
You can then use these "cloaked" photos as you normally would, sharing them on social media, sending them to friends, printing them or displaying them on digital devices, the same way you would any other photo.
The only difference is that a company like the infamous startup Clearview AI can't use them to build an accurate database that will make you trackable.

Here's a before-and-after that the team created to show the cloaking at work. On the left is the original image, on the right a "cloaked" version. The differences are noticeable if you look closely, but they look like the result of dodging and burning rather than actual alterations that might change the way you look:




You can watch an explanation and demonstration of Fawkes by co-lead authors Emily Wenger and Shawn Shan below:

According to the team, Fawkes has proven 100% effective against state-of-the-art facial recognition models. Of course, this won't make facial recognition models obsolete overnight, byt if technology like this caught on as "standard" when, say, uploading an image to social media, it would make maintaining accurate models much more cumbersome and expensive.

"Fawkes is designed to significantly raise the costs of building and maintaining accurate models for large-scale facial recognition," explains the team. "If we can reduce the accuracy of these models to make them untrustable, or force the model's owners to pay significant per-person costs to maintain accuracy, then we would have largely succeeded."

To learn more about this technology, or if you want to download Version 0.3 and try it on your own photos, head over to the Fawkes webpage. The team will be (virtually) presenting their technical paper at the upcoming USENIX Security Symposium running from August 12th to the 14th.

(via Fstoppers via Gizmodo)

Bild/Foto Bild/Foto Bild/Foto Bild/Foto Bild/Foto Bild/Foto

Bild/Foto

#finds #software #technology #ai #algorithm #artificialintelligence #clearview #clearviewai #cloaking #face #facialrecognition #fawkes #photoediting #privacy #security
posted by pod_feeder_v2
This ‘Cloaking’ Algorithm Breaks Facial Recognition by Making Tiny Edits

PetaPixel: This 'Cloaking' Algorithm Breaks Facial Recognition by Making Tiny Edits (DL Cade)

 

This ‘Cloaking’ Algorithm Breaks Facial Recognition by Making Tiny Edits








A team of researchers at the University of Chicago have developed an algorithm that makes tiny, imperceptible edits to your images in order to mask you from facial recognition technology. Their invention is called Fawkes, and anybody can use it on their own images for free.

The algorithm was created by researchers in the SAND Lab at the University of Chicago, and the open-source software tool that they built is free to download and use on your computer at home.

The program works by making "tiny, pixel-level changes that are invisible to the human eye," but that nevertheless prevent facial recognition algorithms from categorizing you correctly. It's not so much that it makes you impossible to categorize; it's that the algorithm will categorize you as a different person entirely. The team calls the result "cloaked" photos, and they can be used like any other:
You can then use these "cloaked" photos as you normally would, sharing them on social media, sending them to friends, printing them or displaying them on digital devices, the same way you would any other photo.
The only difference is that a company like the infamous startup Clearview AI can't use them to build an accurate database that will make you trackable.

Here's a before-and-after that the team created to show the cloaking at work. On the left is the original image, on the right a "cloaked" version. The differences are noticeable if you look closely, but they look like the result of dodging and burning rather than actual alterations that might change the way you look:




You can watch an explanation and demonstration of Fawkes by co-lead authors Emily Wenger and Shawn Shan below:

According to the team, Fawkes has proven 100% effective against state-of-the-art facial recognition models. Of course, this won't make facial recognition models obsolete overnight, byt if technology like this caught on as "standard" when, say, uploading an image to social media, it would make maintaining accurate models much more cumbersome and expensive.

"Fawkes is designed to significantly raise the costs of building and maintaining accurate models for large-scale facial recognition," explains the team. "If we can reduce the accuracy of these models to make them untrustable, or force the model's owners to pay significant per-person costs to maintain accuracy, then we would have largely succeeded."

To learn more about this technology, or if you want to download Version 0.3 and try it on your own photos, head over to the Fawkes webpage. The team will be (virtually) presenting their technical paper at the upcoming USENIX Security Symposium running from August 12th to the 14th.

(via Fstoppers via Gizmodo)

Bild/Foto Bild/Foto Bild/Foto Bild/Foto Bild/Foto Bild/Foto

Bild/Foto

#finds #software #technology #ai #algorithm #artificialintelligence #clearview #clearviewai #cloaking #face #facialrecognition #fawkes #photoediting #privacy #security
posted by pod_feeder_v2
This ‘Cloaking’ Algorithm Breaks Facial Recognition by Making Tiny Edits

PetaPixel: This 'Cloaking' Algorithm Breaks Facial Recognition by Making Tiny Edits (DL Cade)

 
Schau dir "ORDEN OGAN - In The Dawn Of The AI (2020) // Official Music Video // AFM Records" auf YouTube an https://youtu.be/cAYvwbUUhD0

Really cool and nice topic covered!
#music #metal #ai
 
Machine Learning Summer School starts June 28th. All virtual. From the Max Planck Institute for Intelligent Systems, Tübingen, Germany. Topics covered include symbolic AI, statistical AI, causality and learning theory, AI fairness, computational neuroscience in AI, Bayesian AI, game theory in AI, kernel methods, AI in healthcare, deep learning, geometric deep learning, deep reinforcement learning, and quantum machine learning, whatever that is.

The Machine Learning Summer School

#solidstatelife #ai #aieducation
 
Machine Learning Summer School starts June 28th. All virtual. From the Max Planck Institute for Intelligent Systems, Tübingen, Germany. Topics covered include symbolic AI, statistical AI, causality and learning theory, AI fairness, computational neuroscience in AI, Bayesian AI, game theory in AI, kernel methods, AI in healthcare, deep learning, geometric deep learning, deep reinforcement learning, and quantum machine learning, whatever that is.

The Machine Learning Summer School

#solidstatelife #ai #aieducation
 
High resolution neural face swapping. Deepfakes taken to the next level. There's one encoder for any input face, but every output face has its own output decoder trained for that face. The decoding has a "progressive" system where it adds resolution is steps, rather than going end-to-end in high resolution. Face alignment by detecting facial landmarks combined with an "ablation" system eliminates any jitter. The background is composited using a separate compositing process.



#solidstatelife #ai #computervision #generativenetworks #deepfakes
 
High resolution neural face swapping. Deepfakes taken to the next level. There's one encoder for any input face, but every output face has its own output decoder trained for that face. The decoding has a "progressive" system where it adds resolution is steps, rather than going end-to-end in high resolution. Face alignment by detecting facial landmarks combined with an "ablation" system eliminates any jitter. The background is composited using a separate compositing process.



#solidstatelife #ai #computervision #generativenetworks #deepfakes
 
"When to assume neural networks can solve a problem. A pragmatic guide to the powers and limits of neural networks" from Skynet Today. "A neural network can almost certainly solve a problem if another ML algorithm has already been used to solve it." "A neural network can almost certainly solve a problem very similar to ones already solved by neural nets." "A neural network can solve problems that a human can solve if these problems are 'small' in data and require little-to-no context." (Yeah but we like big data, right?) "A neural network might be able to solve a problem when we are reasonably sure that a) it's deterministic, b) we provide any relevant context as part of the input data, and c) the data is reasonably small."

When to Assume Neural Networks Can Solve a Problem

#solidstatelife #ai
 
"When to assume neural networks can solve a problem. A pragmatic guide to the powers and limits of neural networks" from Skynet Today. "A neural network can almost certainly solve a problem if another ML algorithm has already been used to solve it." "A neural network can almost certainly solve a problem very similar to ones already solved by neural nets." "A neural network can solve problems that a human can solve if these problems are 'small' in data and require little-to-no context." (Yeah but we like big data, right?) "A neural network might be able to solve a problem when we are reasonably sure that a) it's deterministic, b) we provide any relevant context as part of the input data, and c) the data is reasonably small."

When to Assume Neural Networks Can Solve a Problem

#solidstatelife #ai
 
Why is artificial intelligence so useless for business? Ponders Matthew Eric Bassett. "Today's work in artificial intelligence is amazing. We've taught computers to beat the most advanced players in the most complex games. We've taught them to drive cars and create photo-realistic videos and images of people. They can re-create works of fine-art and emulate the best writers. Yet I know that many businesses still need people to, e.g., read PDF documents about an office building and write down the sizes of the leasable units contained therein. If artificial intelligence can do all that, why can't it read a PDF document and transform it into a machine-readable format? Today's artificial intelligence algorithms can recreate playable versions of Pacman just from playing games against itself. So why can't I get a computer to translate my colleague's financial spreadsheet into the format my SAP software wants? Despite two decades of advancements in artificial intelligence, it feels that the majority of office work consists of menial mental tasks."

Why is Artificial Intelligence So Useless for Business?

#solidstatelife #ai
 
Why is artificial intelligence so useless for business? Ponders Matthew Eric Bassett. "Today's work in artificial intelligence is amazing. We've taught computers to beat the most advanced players in the most complex games. We've taught them to drive cars and create photo-realistic videos and images of people. They can re-create works of fine-art and emulate the best writers. Yet I know that many businesses still need people to, e.g., read PDF documents about an office building and write down the sizes of the leasable units contained therein. If artificial intelligence can do all that, why can't it read a PDF document and transform it into a machine-readable format? Today's artificial intelligence algorithms can recreate playable versions of Pacman just from playing games against itself. So why can't I get a computer to translate my colleague's financial spreadsheet into the format my SAP software wants? Despite two decades of advancements in artificial intelligence, it feels that the majority of office work consists of menial mental tasks."

Why is Artificial Intelligence So Useless for Business?

#solidstatelife #ai
 
"AI techniques in medical imaging may lead to incorrect diagnoses." The researchers tested 6 medical AI systems on MRI and CT images. They made tiny perturbations to the images to see if those destabilized the AI algorithms. This was done by adding small bits of random noise or small samples from a Fourier transform. They also tested making "structural" changes to the images, in this case adding characters to them. They also tested upsampling the images.

Tiny perturbations lead to a myriad of different artifacts. Not only that, but different AI systems have different artifacts and instabilities. There is no common denominator.

Likewise, there are a variety of failures in trying to recover from structural changes to images. Failures range from complete removal of details to more subtle distortions and blurring of features.

AI systems have to be retrained from scratch on any subsampling pattern. Even increasing the number of samples can cause the quality of reconstruction to deteriorate.

These instabilities are not necessarily rare events. A key question regarding instabilities with respect to tiny perturbations is how much they occur in practice. There can be noise in the images, machines can malfunction, patients can move while images are being made, there can be subtle anatomic differences between patients, and so on.

Current deep learning methods lack any easy way to make the instability problem go away.

AI techniques in medical imaging may lead to incorrect diagnoses

#solidstatelife #ai #medicalai
 
"AI techniques in medical imaging may lead to incorrect diagnoses." The researchers tested 6 medical AI systems on MRI and CT images. They made tiny perturbations to the images to see if those destabilized the AI algorithms. This was done by adding small bits of random noise or small samples from a Fourier transform. They also tested making "structural" changes to the images, in this case adding characters to them. They also tested upsampling the images.

Tiny perturbations lead to a myriad of different artifacts. Not only that, but different AI systems have different artifacts and instabilities. There is no common denominator.

Likewise, there are a variety of failures in trying to recover from structural changes to images. Failures range from complete removal of details to more subtle distortions and blurring of features.

AI systems have to be retrained from scratch on any subsampling pattern. Even increasing the number of samples can cause the quality of reconstruction to deteriorate.

These instabilities are not necessarily rare events. A key question regarding instabilities with respect to tiny perturbations is how much they occur in practice. There can be noise in the images, machines can malfunction, patients can move while images are being made, there can be subtle anatomic differences between patients, and so on.

Current deep learning methods lack any easy way to make the instability problem go away.

AI techniques in medical imaging may lead to incorrect diagnoses

#solidstatelife #ai #medicalai
 
DeepDesigns.ai. AI-designed face masks and other fashion. You pick an initial design, and it generates mutations, and then you pick one of those, and keep iterating as long as you like.

DeepDesigns.ai

#solidstatelife #ai #fashion
 
Later posts Earlier posts