CAMBRIDGE, Mass. If your Uber driver takes a shortcut, you might get to your destination faster. But if a machine learning model takes a shortcut, it might fail in unexpected ways.

In machine learning, a shortcut solution occurs when the model relies on a simple characteristic of a dataset to make a decision, rather than learning the true essence of the data, which can lead to inaccurate predictions. For example, a model might learn to identify images of cows by focusing on the green grass that appears in the photos, rather than the more complex shapes and patterns of the cows.

A new study by researchers at MIT explores the problem of shortcuts in a popular machine-learning method and proposes a solution that can prevent shortcuts by forcing the model to use more data in its decision-making.

By removing the simpler characteristics the model is focusing on, the researchers force it to focus on more complex features of the data that it hadnt been considering. Then, by asking the model to solve the same task two ways once using those simpler features, and then also using the complex features it has now learned to identify they reduce the tendency for shortcut solutions and boost the performance of the model.

One potential application of this work is to enhance the effectiveness of machine learning models that are used to identify disease in medical images. Shortcut solutions in this context could lead to false diagnoses and have dangerous implications for patients.

It is still difficult to tell why deep networks make the decisions that they do, and in particular, which parts of the data these networks choose to focus upon when making a decision. If we can understand how shortcuts work in further detail, we can go even farther to answer some of the fundamental but very practical questions that are really important to people who are trying to deploy these networks, says Joshua Robinson, a PhD student in the Computer Science and Artificial Intelligence Laboratory (CSAIL) and lead author of the paper.

Robinson wrote thepaperwith his advisors, senior author Suvrit Sra, the Esther and Harold E. Edgerton Career Development Associate Professor in the Department of Electrical Engineering and Computer Science (EECS) and a core member of the Institute for Data, Systems, and Society (IDSS) and the Laboratory for Information and Decision Systems; and Stefanie Jegelka, the X-Consortium Career Development Associate Professor in EECS and a member of CSAIL and IDSS; as well as University of Pittsburgh assistant professor Kayhan Batmanghelich and PhD students Li Sun and Ke Yu. The research will be presented at the Conference on Neural Information Processing Systems in December.

The long road to Understanding Shortcuts

The researchers focused their study on contrastive learning, which is a powerful form of self-supervised machine learning. In self-supervised machine learning, a model is trained using raw data that do not have label descriptions from humans. It can therefore be used successfully for a larger variety of data.

A self-supervised learning model learns useful representations of data, which are used as inputs for different tasks, like image classification. But if the model takes shortcuts and fails to capture important information, these tasks wont be able to use that information either.

For example, if a self-supervised learning model is trained to classify pneumonia in X-rays from a number of hospitals, but it learns to make predictions based on a tag that identifies the hospital the scan came from (because some hospitals have more pneumonia cases than others), the model wont perform well when it is given data from a new hospital.

For contrastive learning models, an encoder algorithm is trained to discriminate between pairs of similar inputs and pairs of dissimilar inputs. This process encodes rich and complex data, like images, in a way that the contrastive learning model can interpret.

The researchers tested contrastive learning encoders with a series of images and found that, during this training procedure, they also fall prey to shortcut solutions. The encoders tend to focus on the simplest features of an image to decide which pairs of inputs are similar and which are dissimilar. Ideally, the encoder should focus on all the useful characteristics of the data when making a decision, Jegelka says.

So, the team made it harder to tell the difference between the similar and dissimilar pairs, and found that this changes which features the encoder will look at to make a decision.

If you make the task of discriminating between similar and dissimilar items harder and harder, then your system is forced to learn more meaningful information in the data, because without learning that it cannot solve the task, she says.

But increasing this difficulty resulted in a tradeoff the encoder got better at focusing on some features of the data but became worse at focusing on others. It almost seemed to forget the simpler features, Robinson says.

To avoid this tradeoff, the researchers asked the encoder to discriminate between the pairs the same way it had originally, using the simpler features, and also after the researchers removed the information it had already learned. Solving the task both ways simultaneously caused the encoder to improve across all features.

Their method, called implicit feature modification, adaptively modifies samples to remove the simpler features the encoder is using to discriminate between the pairs. The technique does not rely on human input, which is important because real-world data sets can have hundreds of different features that could combine in complex ways, Sra explains.

From Cars to COPD

The researchers ran one test of this method using images of vehicles. They used implicit feature modification to adjust the color, orientation, and vehicle type to make it harder for the encoder to discriminate between similar and dissimilar pairs of images. The encoder improved its accuracy across all three features texture, shape, and color simultaneously.

To see if the method would stand up to more complex data, the researchers also tested it with samples from a medical image database of chronic obstructive pulmonary disease (COPD). Again, the method led to simultaneous improvements across all features they evaluated.

While this work takes some important steps forward in understanding the causes of shortcut solutions and working to solve them, the researchers say that continuing to refine these methods and applying them to other types of self-supervised learning will be key to future advancements.

This ties into some of the biggest questions about deep learning systems, like Why do they fail? and Can we know in advance the situations where your model will fail? There is still a lot farther to go if you want to understand shortcut learning in its full generality, Robinson says.

This research is supported by the National Science Foundation, National Institutes of Health, and the Pennsylvania Department of Healths SAP SE Commonwealth Universal Research Enhancement (CURE) program.

Written by Adam Zewe, MIT News Office

Paper: Can contrastive learning avoid shortcut solutions?

Original post:
MIT: Forcing ML Models to Avoid Shortcuts (and Use More Data) for Better Predictions - insideHPC

Gazelli Art House to present Code of Arms a group exhibition investigating the history of artificial intelligence (AI) and machine learning in art. The exploration of implementing code and AI in art in the 1970s 80s comes at a time of rapid change in our understanding and appreciation of computer art.

The exhibition brings together pioneer artists in computer and generative art such as George Nees (b.1926), Frieder Nake (b.1938), Manfred Mohr (b.1938) and Vera Molnar (b.1924), and iconic artists employing AI in their practice such as Harold Cohen (b.1928), Lynn Hershman Leeson (b.1941), and Mario Klingemann (b.1970).

Code of Arms follows the evolution of the medium through the works of exhibited artists. Harold Cohens painting, Aspect (1964), a work shown at the Whitechapel Gallery in 1965 (Harold Cohen: Paintings 1960-1965), marks the artists earliest point of enquiry unfolding his scientific and artistic genius. Cohen, who was most famous for creating the computer program AARON, a predecessor of contemporary AI technologies, implemented the program in his work from 80s onwards as seen by the drawings from this period in the exhibition.

Much of the early computer artworks explored geometric forms and structure employing the technology which was still in the infant stage. Plotter drawings carried out by flatbed precision plotter and early printouts by Manfred Mohr, Georg Nees, Frieder Nake and Vera Molnar from the mid-1960s through the 1980s are an excellent representation of that period: the artists focused on the visual forms rather than addressing the underlying meaning and ethics of using computers in their art. The artists saw machines as an external force that would allow them to explore the visual aspect of the works and experiment with the form in an objective manner. Coming from different backgrounds, they worked alongside each other and made an immense contribution to early computer art.

Initially working as an abstract expressionist artist, Manfred Mohr (b. 1938) was inspired by Max Benses information aesthetics which defined his approach to the creative process from the 1960s onwards. Encouraged by the computer music composer Pierre Barbaud whom he met in 1967, Mohr programmed his first computer drawings in 1969. On display are Mohrs plotter drawings of the 70s and 80s alongside a generative software piece from 2015.

Georg Nees (1926-2016) was a German academic who showed one of the worlds first computer graphics created with a digital computer in 1965. In 1970, at the 35th Venice Biennale, he presented his sculptures and architectural designs, which he continued to work on through the 1980s as seen through his drawings in this exhibition.

Frieder Nake (b. 1938) was actively pursuing computer art in the 1960s. With over 300 works produced and shown at various exhibitions (including Cybernetic Serendipity at ICA, London in 1968), Nake brought his background in computer science and mathematics into his art practice. At the The Great Temptation exhibition at the ZKM in 2005, Nees said: ?There it was, the great temptation for me, for once not to represent something technical with this machine but rather something useless geometrical patterns.? Alongside his iconic 60s plotter drawings, Nakes recent body of work (Sets of Straight Lines, 2018) will be on view as a reminder of the artists ability to transform and move away from the geometric abstraction.

Vera Molnar (b. 1924) is a Hungarian-French artist who is considered a pioneer of computer and generative art. Having created combinational images since 1959, Molnars first non-representational images (abstract geometric and systematic paintings) were produced in 1946. Her plotter drawings from the 80s are displayed alongside her later canvas and work on paper (Double Signe Sans Signification, 2005; Deux Angles Droits, 2006) demonstrating the artists consistency and dedication to the process over three decades.

The exhibition moves on to explore relationships between digital technologies and humans through works by Lynn Hershman Leeson (b.1941), an American artist and filmmaker working in moving image, collage, drawing and new media. The artist, who has recently been the focus of a solo exhibition at the New Museum, New York, will show a series of her rare drawings from the 60s and 70s, as well as her seminal work; Agent Ruby, commissioned by SFMoMA (2001), which is an algorithmic work that interacts with online users through a website, shaping the AIs memory, knowledge and moods. Leeson is known for the first interactive piece using Videodisc (Lorna (1983)), and Deep Contact (1984), the first artwork to incorporate a touch screen.

Mario Klingemann brings neural networks, code and algorithms into the contemporary context. The artist investigates systems of todays society employing deep learning, generative and evolutionary art, glitch art, data classification. The exhibition features his recent digital artwork Memories of Passersby I (Solitaire Version), 2018, and prints Morgan le Fay and Cobalamime from 2017.

Mark Westall

Mark Westall is the Founder and Editor of FAD magazine Founder and co-publisher of Art of Conversation and founder of the platform @worldoffad

The New Museums summer 2021 exhibition line-up features three monographic presentations installed in the Museums main galleries. On the Second []

DRAG: Self-portraits and Body Politics ?is the first institutional exhibition to expand on the traditional representations of drag, involving drag queens, drag kings and bio drags from different generations and backgrounds.

Art Basel will screen a premier program of 16 film and video works presented by the shows participating galleries. The Film program is curated for the fourth consecutive year by Cairo-based film curator Maxa Zoller

Dreamlands: Immersive Cinema and Art, 19052016 focuses on the ways in which artists have dismantled and reassembled the conventions of cinemascreen, projection, darknessto create new experiences of the moving image. The exhibition will fill the Whitney Museums 18,000-square-foot fifth-floor galleries, and will include a film series in the third-floor theatre.

Read more here:
New exhibition to investigate the history of AI & machine learning in art. - FAD magazine

Humans excel at extrapolating in a variety of situations. For example, we can use arithmetic to solve problems with infinitely big numbers. One can question if machine learning can do the same thing and generalize to cases that are arbitrarily far apart from the training data. Extrapolation is a statistical technique for estimating values that extend beyond a particular collection of data or observations. In contrast to extrapolation, we shall explain its primary aspects in this article and attempt to connect it to machine learning. The following are the main points to be discussed in this article.

Lets start the discussion by understanding extrapolation.

Extrapolation is a sort of estimation of a variables value beyond the initial observation range based on its relationship with another variable. Extrapolation is similar to interpolation in that it generates estimates between known observations, but it is more uncertain and has a higher risk of giving meaningless results.

Extrapolation can also refer to a methods expansion, presuming that similar methods are applicable. Extrapolation is a term that refers to the process of projecting, extending, or expanding known experience into an unknown or previously unexperienced area in order to arrive at a (typically speculative) understanding of the unknown.

Extrapolation is a method of estimating a value outside of a defined range. Lets take a general example. If youre a parent, you may recall your youngster calling any small four-legged critter a cat because their first classifier employed only a few traits. They were also able to correctly identify dogs after being trained to extrapolate and factor in additional attributes.

Even for humans, extrapolation is challenging. Our models are interpolation machines, no matter how clever they are. Even the most complicated neural networks may fail when asked to extrapolate beyond the limitations of their training data.

Machine learning has traditionally only been able to interpolate data, that is, generate predictions about a scenario that is between two other, known situations. Because machine learning only learns to model existing data locally as accurately as possible, it cannot extrapolate that is, it cannot make predictions about scenarios outside of the known conditions. It takes time and resources to collect enough data for good interpolation, and it necessitates data from extreme or dangerous settings.

When We use data in regression problems to generalize a function that translates a set of input variables X to a set of output variables y. A y value can be predicted for any combination of input variables using this function mapping. When the input variables are located between the training data, this procedure is referred to as interpolation; however, if the point of estimation is located outside of this region, it is referred to as extrapolation.

The grey and white sections in the univariate example in Fig above show the extrapolation and interpolation regimes, respectively. The black lines reflect a selection of polynomial models that were used to make predictions within and outside of the training data set.

The models are well limited in the interpolation regime, causing them to collapse in a tiny region. However, outside of the domain, the models diverge, producing radically disparate predictions. The absence of information given to the model during training that would confine the model to predictions with a smaller variance is the cause of this large divergence of predictions (despite being the same model with slightly different hyperparameters and trained on the same set of data).

This is the risk of extrapolation: model predictions outside of the training domain are particularly sensitive to training data and model parameters, resulting in unpredictable behaviour unless the model formulation contains implicit or explicit assumptions.

In the absence of training data, most learners do not specify the behaviour of their final functions. Theyre usually made to be universal approximators or as close as possible with few modelling constraints. As a result, in places where there is little or no data, the function has very little previous control. As a result, we cant regulate the behaviour of the prediction function at extrapolation points in most machine learning scenarios, and we cant tell when this is a problem.

Extrapolation should not be a problem in theory; in a static system with a representative training sample, the chances of having to anticipate a point of extrapolation are essentially zero. However, most training sets are not representative, and they are not derived from static systems, therefore extrapolation may be required.

Even empirical data derived from a product distribution can appear to have a strong correlation pattern when scaled up to high dimensions. Because functions are learned based on an empirical sample, they may be able to extrapolate effectively even in theoretically dense locations.

Extrapolation works with linear and other types of regression to some extent, but not with decision trees or random forests. In the Decision Tree and Random Forest, the input is sorted and filtered down into leaf nodes that have no direct relationship to other leaf nodes in the tree or forest. This means that, while the random forest is great at sorting data, the results cant be extrapolated because it doesnt know how to classify data outside of the domain.

A good decision on which extrapolation method to use is based on a prior understanding of the process that produced the existing data points. Some experts have recommended using causal factors to assess extrapolation approaches. We will see a few of them. These are pure mathematical methods one should relate to your problem properly.

Linear extrapolation is the process of drawing a tangent line from the known datas end and extending it beyond that point. Only use linear extrapolation to extend the graph of an essentially linear function or not too much beyond the existing data to get good results. Linear extrapolation produces the function if the two data points closest to the point x* to be extrapolated are (xk-1,yk-1) and (xk,yk).

A polynomial curve can be built using all of the known data or just a small portion of it (two points for linear extrapolation, three points for quadratic extrapolation, etc.). The curve that results can then be extended beyond the available data. The most common way of polynomial extrapolation is to use Lagrange interpolation or Newtons method of finite differences to generate a Newton series that matches the data. The data can be extrapolated using the obtained polynomial.

Five spots near the end of the given data can be used to make a conic section. If the conic section is an ellipse or a circle, it will loop back and rejoin itself when extrapolated. A parabola or hyperbola that has been extrapolated will not rejoin itself, but it may curve back toward the X-axis. A conic sections template (on paper) or a computer could be used for this form of extrapolation.

Further, we will see the simple python implementation of linear extrapolation.

The technique is beneficial when the linear function is known. Its done by drawing a tangent and extending it beyond the limit. When the projected point is close to the rest of the points, linear extrapolation delivers a decent result.

Extrapolation is a helpful technique, but it must be used in conjunction with the appropriate model for describing the data, and it has limitations after you leave the training area. Its applications include predicting in situations where you have continuous data, such as time, speed, and so on. Prediction is notoriously imprecise, and the accuracy falls as the distance from the learned area grows. In situations where extrapolation is required, the model should be updated and retrained to lower the margin of error. Through this article, we have understood extrapolation and its interpolation mathematically and related them with the ML, and seen their effect on the ML system. We have also seen particularly where it fails, and methods that can be used.

Visit link:
An Illustrative Guide to Extrapolation in Machine Learning - Analytics India Magazine

Sama, a training data provider in Artificial Intelligence (AI) projects, has announced it has received the 2021 AI TechAward for Best in Machine Learning Platforms.

The annual awards celebrate companies leading in technical innovation, adoption, and reception in the AI and machine learning industry and by the developer community, presented by AI DevWorld.

The winners for the 2021 awards were selected from more than 100 entries submitted per category globally, and announced during a virtual AI DevWorld conference. The conference targeted software engineers and data scientists interested in AI as well as AI dev professionals looking for a landscape view on the newest AI technologies.

For over a decade, organizations such as Google, Microsoft, NVIDIA and others have continued to rely on Sama to deliver secure, high-quality training data and model validation for their machine learning projects.

Sama hires over 90% of its workforce from low-income backgrounds and marginalized populations, including unemployed urban and rural youth that are traditionally excluded from the digital economy.

The company has remained committed to connecting people to dignified digital work and paying them living wages that solve some of the worlds most pressing challenges. This includes reducing poverty, empowering women, and mitigating climate change. As a result, it has helped over 56,000 people lift themselves out of poverty, increased wages of workers up to 4 times, and provided over 11,000 hours of training to youth and women, who comprise over 50% of their workforce in both Kenya and Uganda offices.

Earlier this year, the company was recognized as one of the fastest-growing private companies in America on the 2021 Inc. 5000 list for the second year in a row.

Here is the original post:
Sama has been named the Best in Machine Learning Platforms at the 2021 AI TechAwards - HapaKenya - HapaKenya

Subsequent to the consultation report published by IOSCO in June 2021, the final guidance report (IOSCO Report) for artificial intelligence (AI) and machine learning (ML) entitled The use of artificial intelligence and machine learning by market intermediaries and asset managers was released on 7 September 2021.

Per the IOSCO Report, market intermediaries and asset managers tend to achieve cost reductions and improve efficiency through the use of AL and ML. While market intermediaries, asset managers and investors receive benefits including efficiency enhancement, cost reduction and resource sparing, a concern is amplification of risk that affects the interests of consumers and other market participants.

In the light of the above, the IOSCO Report sets out some recommended measures to ensure that the interests of investors and other relevant stakeholders are protected. Further, Annex I and Annex 2 to the IOSCO Report outline the regulators responses to the challenges arising from AI and ML, and the guidance issued by supranational bodies respectively.

Read the rest here:
Board of the International Organisation of Securities Commissions (IOSCO) publishes final guidance report for artificial intelligence and machine...