Skip to content

AI Broader Impact Statements

Example Inputs

Image segmentation has been one of the main challenges in modern medical image analysis, and describes the process of assigning each pixel or voxel in images with biologically meaningful discrete labels, such as anatomical structures and tissue types (e.g. pathology and healthy tissues). The task is required in many clinical and research applications, including surgical planning [41, 42], and the study of disease progression, aging or healthy development [43–45]. However, there are many cases in practice where the correct delineation of structures is challenging; this is also reflected in the well-known presence of high inter- and intra-reader variability in segmentation labels obtained from trained experts [9, 23, 5]. Although expert manual annotations of lesions is feasible in practice, this task is time consuming. It usually takes 1.5 to 2 hours to label a MS patient with average 3 visit scans. Meanwhile, the long-established gold standard for segmentation of medical images has been manually voxel-by-voxel labeled by an expert anatomist. Unfortunately, this process is fraught with both interand intra-rater variability (e.g., on the order of approximately 10% by volume [46, 47]). Thus, developing an automatic segmentation technique to fix the variability among inter- and intra-readers could be meaningful not only in terms of the accuracy in delineating MS lesions but also in the related reductions in time and economic costs derived from manual lesion labeling. The lack of consistency in labelling is also common to see in other medical imaging applications, e.g., in lung abnormalities segmentation from CT images. A lesion might be clearly visible by one annotator, but the information about whether it is cancer tissue or not might not be clear to others. While our work in the current form has only been demonstrated on medical images, we would like to stress that the medical imaging domain offers a considerably broad range of opportunities for impact; e.g., diagnosis/prognosis in radiology, surgical planning and study of disease progression and treatment, etc. In addition, the annotator information could be potentially utilised for the purpose of education. Another potential opportunity is to integrate such information into the data/label acquisition scheme in order to train reliable segmentation algorithms in a data-efficient manner
In this paper, we offer a new interpretation of the self-distillation training framework, a commonly used technique for improved accuracy used among practitioners in the deep learning community, which allows us to gain some deeper understanding of the reasons for its success. With the ubiquity of deep learning in our society today and countless potential future applications of it, we believe our work can potentially bring positive impacts in several ways. Firstly, despite the empirical utility of distillation and numerous successful applications in many tasks and applications ranging from computer vision to natural language processing problems, we still lack a thorough understanding of why it works. In our opinion, blindly applying methods and algorithms without a good grasp on the underlying mechanisms can be dangerous. Our perspective offers a theoretically grounded explanation for its success that allows us to apply the techniques to real-world applications broadly with greater confidence. In addition, the proposed interpretation of distillation as a regularization to neural networks can potentially allow us to obtain models that are more generalizable and reliable. This is an extremely important aspect of applying deep learning to sensitive domains like healthcare and autonomous driving, in which wrong predictions made by machines can lead to catastrophic consequences. Moreover, our new experimental demonstration that models trained with the distillation process can potentially lead to better-calibrated models that can facilitate safer and more interpretable applications of neural networks. Indeed, for real-world classification tasks like disease diagnosis, in addition to accurate predictions, we need reliable estimates of the level of confidence of the predictions made, which is something that neural networks are lacking currently as pointed out by recent research. More calibrated models, in our opinion, enhances the explainability and transparency of neural network models. Lastly, we believe the introduced framework can stimulate further research on the regularization of deep learning models for better generalization and thus safer applications. It was recently demon- strated that deep neural networks do not seem to suffer from overfitting. Our finding suggests that overfitting can still occur, though in a different way than conventional wisdom, and deep learning can still benefit from regularization. As such, we encourage research into more efficient and principled forms of regularization to improve upon the distillation strategy. We acknowledge the risks associated with our work. To be more specific, our finding advocates for the use of priors for the regularization of neural networks. Despite the potentially better generalization performance of trained models, depending on the choice of priors used for training, unwanted bias can be inevitably introduced into the deep learning system, potentially causing issues of fairness and privacy.
This paper presents methods for estimating point-wise dependency between high-dimensional data using neural networks. This work may benefit the applications that require understanding instance- level dependency. Take adversarial samples detection as an example: we can perform point-wise dependency estimation between data and label, and the ones with low point-wise dependency can be regarded as adversarial samples. We should also be aware of the malicious usage for our framework. For instance, people with bad intentions can use our framework to detect samples that have a high point-wise dependency with their of-interest private attributes. Then, these detected samples may be used for malicious purposes.
In this work we design algorithms with provable robustness guarantees in the challenging setting where the level of noise is allowed to vary across the domain. This models several scenarios of interest, most notably situations where data provided by certain demographic groups is subject to more noise than others. In a natural experiment on the UCI Adult dataset, we show that coping with this type of noise can help mitigate some natural types of unfairness that arise with off-the-shelf algorithms. Moreover our algorithms have the additional benefit that they lead to more readily interpretable hypotheses. In many settings of interest, we are able to give proper learning algorithms (where previously only improper learning algorithms were known). This could potentially help practitioners better understand and diagnose complex machine learning systems they are designing, and troubleshoot ways that the algorithm might be amplifying biases in the data.
Our approach is motivated by sequential decision-making problems that arise in several domains such as road traffic, markets, and security applications with potentially significant societal benefits. In such domains, it is important to predict how the system responds to any given decision and take this into account to achieve the desired performance. The methods proposed in this paper require to observe and quantify (via suitable indicators) the response of the system and to dispose of computational resources to process the observed data. Moreover, it is important that the integrity and the reliability of such data are verified, and that the used algorithms are complemented with suitable measures that ensure the safety of the system at any point in time.
From the sculptor’s chisel to the painter’s brush, tools for creative expression are an important part of human culture. The advent of digital photography and professional editing tools, such as Adobe Photoshop, has allowed artists to push creative boundaries. However, the existing tools are typically too complicated to be useful by the general public. Our work is one of the new generation of visual content creation methods that aim to democratize the creative process. The goal is to provide intuitive controls (see Section 4.6) for making a wider range of realistic visual effects available to non-experts. While the goal of this work is to support artistic and creative applications, the potential misuse of such technology for purposes of deception – posing generated images as real photographs – is quite concern- ing. To partially mitigate this concern, we can use the advances in the field of image forensics [16], as a way of verifying the authenticity of a given image. In particular, Wang et al. [72] recently showed that a classifier trained to classify between real photographs and synthetic images generated by ProGAN [42], was able to detect fakes produced by other generators, among them, StyleGAN [43] and Style- GAN2 [44]. We take a pretrained model of [72] and report the detection rates on several datasets in Ap- pendix ??. Our swap-generated images can be detected with an average rate greater than 90%, and this in- dicates that our method shares enough architectural components with previous methods to be detectable. However, these detection methods do not work at 100%, and performance can degrade as the images are degraded in the wild (e.g., compressed, rescanned) or via adversarial attacks. Therefore, the problem of verifying image provenance remains a significant challenge to society that requires multiple layers of solutions, from technical (such as learning-based detection systems or authenticity certification chains), to social, such as efforts to increase public awareness of the problem, to regulatory and legislative.
This work touches upon a very old problem dating back to 1933 and the work of [39]. Therefore, we don’t anticipate any new societal impacts or ethical aspects, that are not well understood by now.
Product design (e.g., furniture ) is labor extensive and requires expertise in computer graphics. With the increasing number and diversity of 3D CAD models in online repositories, there is a growing need for leverage them to facilitate future product development due to their similarities in function and shape. Towards this goal, our proposed method provide a novel unsupervised paradigm to establish dense correspondence for topology-varying objects, which is a prerequisite for shape analysis and synthesis. Furthermore, as our approach is designed for generic objects, its application space can be extremely wide.
Quantum computation has recently been attracting growing attentions owing to its potential for achieving computational speedups compared to any conventional classical computation that runs on existing computers, opening the new field of accelerating machine learning tasks via quantum computation: quantum machine learning. To attain a large quantum speedup, however, existing algorithms for quantum machine learning require extreme assumptions on sparsity and low rank of matrices used in the algorithms, which limit applicability of the quantum computation to machine learning tasks. In contrast, the novelty of this research is to achieve an exponential speedup in quantum machine learning without the sparsity and low-rank assumptions, broadening the applicability of quantum machine learning. Advantageously, our quantum algorithm eliminates the computational bottleneck faced by a class of existing classical algorithms for scaling up kernel-based learning algorithms by means of random features. In particular, using this quantum algorithm, we can achieve the learning with the nearly optimal number of features, whereas this optimization has been hard to realize due to the bottleneck in the existing classical algorithms. A drawback of our quantum algorithm may arise from the fact that we use powerful quantum subroutines for achieving the large speedup, and these subroutines are hard to implement on existing or near-term quantum devices that cannot achieve universal quantum computation due to noise. At the same time, these subroutines make our quantum algorithm hard to simulate by classical computation, from which stems the computational advantage of our quantum algorithm over the existing classical algorithms. Thus, our results open a route to a widely applicable framework of kernel-based quantum machine learning with an exponential speedup, leading to a promising candidate of “killer applications” of universal quantum computers.
This work mainly contributes to AutoML in the aspect of discovering better learning rules or optimization algorithms from data. As a fundamental technique, it seems to pose no substantial societal risk. This paper proposes several improved training techniques to tackle the dilemma of training instability and poor generalization in learned optimizers. In general, learning to optimize (L2O) prevents laborious problem-specific optimizer design, and potentially can largely reduce the cost (including time, energy and expense) of model training or tuning hyperparameters.

Example

LLooM Logo

Outputs

Select seed. The seed term can steer concept induction towards more specific areas of interest. Try out one of the options below:

Bias andFairnessSecurityandAttacksEnvironmentalImpactJobDisplacement02040↑ Number of documents

Bias and Fairness

Criteria: Is bias or fairness in machine learning models or data a focus of this text example?

Summary: Addressing biases in machine learning models is crucial to prevent perpetuating social biases, discrimination, and unfairness in decision-making processes.

Security and Attacks

Criteria: Does the text example address security vulnerabilities or adversarial attacks against systems?

Summary: Addressing security concerns in machine learning is crucial for ensuring system safety and reliability against potential attacks.

Environmental Impact

Criteria: Does this text example highlight concerns about the environmental impact of technology?

Summary: Addressing environmental impact in AI research is crucial due to high energy consumption and carbon footprint of training large models.

Job Displacement

Criteria: Does this text example discuss the risk of job displacement due to automation or AI?

Summary: Advancements in technology, particularly artificial intelligence and automation, pose a threat to various job sectors, potentially leading to widespread unemployment.

Analysis

➡️ Try out analyzing this data with LLooM on this Colab notebook.

LLooM, AI Broader Impact Statements Notebook

Task: Investigate anticipated consequences of AI research

Advances in AI are driven by research labs, so to avoid future harms, today's researchers must be equipped to grapple with AI ethics, including the ability to anticipate risks and mitigate potentially harmful downstream impacts of their work. How do AI researchers assess the consequences of their work? LLooM can help us to understand how AI researchers discuss downstream outcomes, ethical issues, and potential mitigations. Such an analysis could help us to uncover gaps in understanding that could be addressed with guidelines and AI ethics curricula.

Dataset: NeurIPS Broader Impact Statements, 2020

In 2020, NeurIPS, a premier machine learning research conference, required authors to include a broader impact statement in their submission in an effort to encourage researchers to consider negative consequences of their work. These statements provide a window into the ethical thought processes of a broad swath of AI researchers, and prior work from Nanayakkara et al. has performed a qualitative thematic analysis on a sample of 300 statements.