class6

02/26/20 – Fanglan Chen – Will You Accept an Imperfect AI? Exploring Designs For Adjusting End-user Expectations of AI Systems

Fanglan Chen 4 Comments

Summary

Kocielnik et al.’s paper “Will You Accept an Imperfect AI?” explores approaches for shaping expectations of end-users before their initial working with an AI system and studies how appropriate expectations impact users’ acceptance of the system. Prior study has presented that end-user expectations of AI-powered technologies are influenced by various factors, such as external information, knowledge and understanding, and first hand experience. The researchers indicate that expectations vary among users and users perception/acceptance of AI systems may be negatively impacted when their expectations are set too high. To fill in the gap of understanding how end-user expectations can be directly and explicitly impacted, the researchers use a Scheduling Assistant – an AI system for automatic meeting schedule detection in email – to study the impact of several methods of expectation shaping. Specifically, they explore two system versions with the same accuracy level of the classifier but each is intended to focus on mitigating different types of errors(False Positives and False Negatives). Based on their study, error types highly relate to users’ subjective perceptions of accuracy and acceptance. Expectation adjustment techniques are proposed to make users fully aware of AI imperfections and enhance their acceptance of AI systems.

Reflection

We need to be aware that AI-based technologies cannot be perfect, just like nobody is perfect. Hence, there is no point setting a goal that involves AI systems making no mistake. Realistically defining what success and failure look like associated with working with AI-powered technologies is of great importance in adopting AI to improve the imperfection of nowadays solutions. That calls for an accurate positioning of where AI sits in the bigger picture. I feel the paper mainly focuses on how to set appropriate expectations but lacks a discussion on different scenarios associated with the users expectations to AI. For example, users expectation greatly vary to the same AI system in different decision making frameworks: in human-centric decision making process, the expectation of AI component is comparatively low as AI’s role is more like a counselor who is allowed to make some mistakes; in machine-centric system, all the decisions are made by algorithms which render users’ low tolerance of errors, simply put, some AIs will require more attention than others, because the impact of errors or cost of failures will be higher. Expectations of AI systems vary not only among different users but also under various usage scenarios.

To generate positive user experiences, AI needs to exceed expectations. One simple way to achieve this is to not over-promise the performance of AI in the beginning. That relates with the intention of the researchers on designing the Accuracy Indicator component in the Scheduling Assistant. In the study case, they set the accuracy to 50%. This accuracy is actually very low in AI-based applications. I’m interested in whether the evaluation results would change with AI systems of higher performance (e.g. 70% or 90% in accuracy). I think it is worthwhile to conduct a survey about users’ general expectations of AI-based systems. 

Interpretability of AI is another key component that shapes user experiences. If people cannot understand how AI works or how it comes up with its solutions, and in turn do not trust it, they would probably not choose to use it. As people accumulate more positive experiences, they build trust with AI. In this way, easy-to-interpret models seem to be more promising to deliver success compared with complex black-box models. 

To sum up, by being fully aware of AI’s potential but also its limitations, and developing strategies to set appropriate expectations, users can create positive AI experiences and build trust in an algorithmic approach in decision making processes.

Discussion

I think the following questions are worthy of further discussion.

  • What is your expectation of AI systems in general? 
  • How would users expectations of the same AI system vary in different usage scenarios?
  • What are the negative impacts brought by the inflated expectations? Please give some examples. 
  • How can we determine which type of errors is more severe in an AI system?

Read More

02/26/20 – Lulwah AlKulaib- Explaining Models

Lulwah AlKulaib 3 Comments

Summary

The authors believe that in order to ensure fairness in machine learning systems, it is mandatory to have a human in the loop process. In order to identify fairness problems and make improvements, they suppose relying on developers, users, and the general public is an effective way to follow that process. The paper conducts an empirical study with four types of programmatically generated explanations to understand how they impact people’s fairness judgments of ML systems. They try to answer three research questions:

  • RQ1 How do different styles of explanation impact fairness judgment of a ML system?
  • RQ2 How do individual factors in cognitive style and prior position on algorithmic fairness impact the fairness judgment with regard to different explanations?
  • RQ3 What are the benefits and drawbacks of different explanations in supporting fairness judgment of ML systems?

The authors focus on a racial discrimination case study in terms of model unfairness and Case-specific disparate impact. They performed an experiment with 160 Mechanical Turk workers. Their hypothesis proposed that given local explanations focus on justifying a particular case, they should more effectively surface fairness discrepancies between cases. 

 The authors show that: 

  • Certain explanations are considered inherently less fair, while others can enhance people’s confidence in the fairness of the algorithm
  • Different fairness problems-such as model-wide fairness issues versus case-specific fairness discrepancies-may be more effectively exposed through different styles of explanation
  • Individual differences, including prior positions and judgment criteria of algorithmic fairness, impact how people react to different styles of explanation.

Reflection

This is a really informative paper. I like that it had a straightforward hypothesis and chose one existing case study that they evaluated. But I would have loved to see this addressed with judges instead of crowdworkers. They mentioned it in their limitations and I hope that they find enough judges willing to work on a follow-up paper. I believe that they would have insightful knowledge to contribute especially since they practice it. It would give a more meaningful analysis to the case study itself from professionals in the field.

I also wonder how this might scale to different machine learning systems that cover similar racial biases. Having a specific case study makes it harder to generalize even for something in the same domain. But definitely worth investigating since there are so many existing case studies! I also wonder if changing the case study analyzed, we’d notice a difference in the local vs. global explanations patterns in fairness judgement. And how would a mix of both affect the judgement, too. 

Discussion

  • What are other ways you would approach this case study?
  • What are some explanations that weren’t covered in this study?
  • How would you facilitate this study to be performed with judges?
  • What are other case studies that you could generalize this to with small changes to the hypothesis?

Read More

02/26/20 – Sukrit Venkatagiri – Will You Accept an Imperfect AI?

Sukrit Venkatagiri 2 Comments

Paper: Rafal Kocielnik, Saleema Amershi, and Paul N. Bennett. 2019. Will You Accept an Imperfect AI? Exploring Designs for Adjusting End-user Expectations of AI Systems. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (CHI ’19), 1–14.

Summary: 

This paper explores people’s perceptions and expectations of an intelligent scheduling assistant. The paper specifically considers three broad research questions: the impact of AI’s focus on error avoidance versus user perception, ways to set appropriate expectations, and impact of expectation setting on user satisfaction and acceptance. The paper explores this through an experimental setup, whose design process is explored in detail. 

The authors find that expectation adjustment designs significantly affected the desired aspects of expectations, similar to what was hypothesized. They also find that high recall resulted in significantly higher perceptions of accuracy and acceptance compared to high precision, and that expectation adjustment worked by intelligible explanations and tweaking model evaluation metrics to emphasize one over the other. The paper concludes with a discussion of the findings.

Reflection:

This paper presents some interesting findings using a relatively simple, yet powerful “technology probe.” I appreciate the thorough exploration of the design space, taking into consideration design principles and how they were modified to meet the required goals. I also appreciate the varied and nuanced research questions. However, I feel like the setup may have been too simple to explore in more depth. Certainly, this is valuable as a formative study, but more work needs to be done. 

It was interesting that people valued high recall over high precision. I wonder if the results would differ among people with varied expertise, from different countries, and from different socioeconomic backgrounds. I also wonder how this might differ based on the application scenario, e.g. AI scheduling assistant versus a movie recommendation system. In the latter, a user would not be aware of what movies they were not recommended but that they would actually like, while with an email scheduling assistant, it is easy to see false negatives.

I wonder how these techniques, such as expectation setting, might apply not only to users’ expectations of AI systems, but also to exploring the interpretability or explainability of more complex ML models.

At what point do explanations tend to result in the opposite effect? I.e. reduced user acceptance and preference? It may be interesting to experimentally study how different levels of explanations and expectation settings affect user perceptions versus a binary value. I also wonder how it might change with people of different backgrounds.

In addition, this experiment was relatively short in duration. I wonder how the findings would change over time. Perhaps users would form inaccurate expectations, or their mental models might be better steered through expectation-setting. More work is needed in this regard. 

Questions:

  1. Will you accept an imperfect AI?
  2. How do you determine how much explanation is enough? How would this work for more complex models?
  3. What other evaluation metrics can be used?
  4. When is high precision valued over high recall, and vice versa?

Read More

02/26/2020 – Palakh Mignonne Jude – Interpreting Interpretability: Understanding Data Scientists’ Use Of Interpretability Tools For Machine Learning

Palakh Mignonne Jude 2 Comments

SUMMARY

In this paper, the authors attempt to study two interpretability tools – the InterpretML implementation of GAMs and the SHAP Python package. They conducted a contextual inquiry and survey of data scientists in order to analyze the ability of these tools to aid in uncovering common issues that arise when evaluating ML models. The results obtained during the course of these studies indicate that data scientists tend to over-trust these tools. The authors conducted pilot interviews with 6 participants to identify common issues faced by data scientists. The contextual inquiry performed included 11 participants who were allowed to explore the dataset and an ML model in a hands-on manner via the use of a Jupyter notebook whereas the survey comprised of 197 participants and was conducted through Qualtrics. For the survey, the participants were given access to a description of the dataset and a tutorial on the interpretability tool they were to use. The authors found that the visualizations provided by the interpretability tools considered in the study as well as the fact that these tools were popular and publicly available caused the data scientists to over-trust these tools.

REFLECTION

I think it is good that the authors performed a study to observe the usage of interpretability tools by data scientists. I was surprised to learn that a large number of these data scientists over-trusted the tools and that visualizations impacted their ability to judge the tools as well. However, considering that the authors state ‘participants relied too heavily on the interpretability tools because they has not encountered such visualizations before’ makes me wonder if the authors should have created a separate pool of data scientists who had better experience with such tools and visualizations and then presented a separate set of results for that set of individuals. I also found it interesting to learn that some participants used the tools to rationalize suspicious observations.

As indicated by the limitations section of this paper, I think a follow-up study that includes a richer dataset as well as interpretability techniques for deep learning would be very interesting to learn about and I wonder how data scientists would use such tools versus the ones studied in this paper.

QUESTIONS

  1. Considering that the complexity of ML systems and the time taken for researchers to truly understand how to interpret ML, both the contextual inquiry as well as the survey was conducted with people who had as little as 2 months of experience with ML. Would a study with experts in the field of ML (all with over 4 years of experience) have yielded different results? Perhaps these data scientists would have been able to better identify issues and would not have over-trusted the interpretable tools?
  2. Would a more extensive study comprise of a number of different (commonly used as well as not-so-commonly used) interpretability tools have changed the results? If the tools were not available so easily would it truly impact the amount of trust the users had for the tools?
  3. Does a correlation exist between the amount of experience a data scientist has and the amount of trust for a given interpretability tool? Would the replacement of visualizations with other representations of interpretations of the models impact the amount of trust the human had towards the tool?

Read More

02/26/2020 – Ziyao Wang – Interpreting Interpretability: Understanding Data Scientists’ Use of Interpretability Tools for Machine Learning

Ziyao Wang 2 Comments

As machine learning models are deployed in variety domains of industry, it is important to design some interpretability to help model users, such as data scientists and machine learning practitioners, better understand how these models work. However, there have been little researches focused on the evaluation of the performance of these tools. The authors in this paper did experiments and surveys to fill this gap. They interviewed 6 data scientists from a large technology company to find out the most common issues faced by data scientists. Then they conducted a contextual inquiry towards 11 participants based on the common issues using the InterpretML implementation of the Gams and the SHAP python package. Finally, they made a survey of 197 data scientists. With the experiments and surveys, the authors highlighted the misuse and over-trust problem and the need for the communication between members of HCI and ML communities.

Reflection:

Before reading this paper, I hold the view that the interpretability tools should be able to cover most of the data scientists’ need. However, now I have the view that the tools for interpretation are not designed by the ML community, which will result in the lack of accuracy of the tools. When data scientists or machine learning practitioners want to use these tools to learn how the models operate, they may face problems like misuse or over-trust. I don not think this is the users’ fault. Tools are designed for make users feel more convenient when doing tasks. If the tools will make users confuse, the developers should make change to the tools to give users better user experiences. In this case, the authors suggested that the members of HCI and ML communities should work together when developing the tools. This need the members to leverage their strength so that the designed tools can let users understand the models easily while the tools are user-friendly. Meanwhile, comprehensive instructions should be written to explain how the users can use the tools to understand the models accurately and easily. Finally, both the efficiency and accuracy of both the tools and the implementation of models will be improved.

From data scientists and machine learning practitioners’ point of view, they should try to avoid to over-trust the tools. The tools cannot fully explain the models and there may be mistakes. The users should always be critic to the tools instead of fully trusting them. They should read the instructions carefully, understand how to use the tools and what the tools are used for, what is the models being used for and how to use the models. If they can consider thoughtfully when using these tools and models, instead of guessing the meaning of the results from the tools, the number of misuse and over-trust cases will be decreased sharply.

Questions:

  1. How to design a proposed interactive interpretability tool? What kinds of interactions should be included?
  2. How to design a tool that can make users to dig the models conveniently instead of letting them use the models without knowing how the models work?
  3. How to design tools which can leverage the strength of mental models mostly

Read More

2/26/20 – Jooyoung Whang – Explaining Models: An Empirical Study of How Explanations Impact Fairness Judgment

Jooyoung Whang 2 Comments

The paper provides research on Fairness, Explainable Artificial Intelligence (XAI), and people’s judgment change. The authors introduce a preprocessing method to reduce the bias of a dataset for known bias-inducing attributes. They also show four explanation methods of the classification results: Sensitivity, Input-Influence, Case, and Demographic. Using different combination of the above configurations, AI classifications of the COMPAS data was presented to MTurk workers for feedback. As a result, the paper reports that case-based explanations were often seen as less fair than other explanation methods. The authors also found that sensitivity explanations are the most effective at addressing unfairness. Finally, the paper shows that the evaluator’s position on machine learning heavily impacts his or her reaction to a classifier output and explanations.

When I looked at the paper’s sample sensitivity explanation, it gave me a strong impression that the system was racist. I think many others would have had a similar thought, especially if they do not have enough knowledge about machine learning and regression. Because of this, it concerned me that some people may be lured more towards making the opposite decision than the one that the AI made as a repulsive reaction. This is clearly adding another bias in the opposite direction. I believe an explanatory model should only give helpful information about the model instead of giving bias. Thinking of a possible solution, the authors could have rephrased the same information in a different way. For example, instead of bluntly saying that the classifier would have made a different decision, the system could have reported the probability for each label. This provides the same information but adds less obvious bias. Another solution would be preprocessing the data to not have the bias in the first place like the authors suggested.

I liked the idea of comparing the subject’s prior position to using ML with their judgment of the classifier. This relates to a reflection I made last week, where I stated the possibility that people may make decisions by putting more weight when the model makes a wrong decision. As I have expected, the paper reported that prior positions do in fact make a huge difference in a user’s judgment. Either building more trust with the users or building the software to effectively address both kinds of users would be needed to address this issue.

The followings are the questions I had while reading the paper:

1. Would there be a possibility where preprocessing the data would add bias to the data instead of removing it? What if the attribute that was thought to be unneeded for the classification was actually crucial to the judgment?

2. The authors state that one of the limitations of their study is conducting it with MTurk workers and not the actual users of the software. Do you think this was really a limitation? The attributes used for the classifier and explanations in their experiment seemed general enough for non-professionals to make a meaningful judgment.

3. If you were to design a classifier with an explanation model, which explanation method would you pick? (Out of Sensitivity, Input-Influence, Case, and Demographic) What do you like about the chosen method?

Read More

02/26/2020 – Vikram Mohanty – Will you accept an imperfect AI? Exploring Designs for Adjusting End-user Expectations of AI Systems

Vikram Mohanty 2 Comments

Authors: Rafal Kocielnik, Saleema Amershi, Paul Bennett

Summary


This paper discusses the impact of end-user expectations on the subjective perception of AI-based systems. The authors conduct studies to better understand how different types of errors (i.e. False Positives and False Negatives) are perceived differently by users, even though accuracy remains the same. The paper uses the context of an AI-based scheduling assistant (in an email client) to demonstrate 3 different design interventions for helping end-users adjust their expectations of the system. The studies in this paper showed that these 3 techniques were effective in preserving user satisfaction and acceptance of an imperfect AI-based system. 

Reflection

This paper is basically an evaluation of the first 2 guidelines from the “Guidelines of Human-AI Interaction” paper i.e. making clear what the system can do, and how well it can do what it does. 

Even though the task in the study was artificial (i.e. using workers from an internal crowdsourcing platform instead of real users of a system and subjecting to an artificial task instead of a real one), the study design, the research questions and the inference from the data initiates the conversation on giving special attention to the user experience in AI-infused systems. Because the tasks were artificial, we could not assess scenarios where users actually have a dog in the fight e.g. they miss an important event by over-relying on the AI assistant and start to depend less on the AI suggestions. 

The task here was scheduling events from emails, which is somewhat simple in the sense that users can almost immediately assess how good or bad the system is at. Furthermore, the authors manipulated the dataset for preparing the High Precision and High Recall versions of the system. For conducting this study in a real-world scenario, this would require a better understanding of user mental models with respect to AI imperfections. It becomes slightly trickier when these AI imperfections can not be accurately assessed in a real-world context e.g. search engines may retrieve pages containing the keywords, but may not account context into the results, and thus may not always give users what they want.  

The paper makes an excellent case of digging deeper into error recovery costs and correlating that with why participants in this study preferred a system with high false positive rates. This is critical for system designers to keep in mind while dealing with uncertain agents like an AI core. This gets further escalated when it’s a high-stakes scenario. 

Questions

  1. The paper starts off with the hypothesis that avoiding false positives is considered better for user experience, and therefore systems are optimized for high precision. The findings however contradicted it. Can you think about scenarios where you’d prefer a system with a higher likelihood of false positives? Can you think about scenarios where you’d prefer a system with a higher likelihood of false negatives?
  2. Did you think the design interventions were exhaustive? How would you have added on to the ones suggested in the paper? If you were to adopt something for your own research, what would it be? 
  3. The paper discusses factoring in other aspects, such as workload, both mental and physical, and the criticality of consequences. How would you leverage these aspects in design interventions? 
  4. If you used an AI-infused system every day (to the extent it’s subconsciously a part of your life)
    1. Would you be able to assess the AI imperfections purely on the basis of usage? How long would it take for you to assess the nature of the AI? 
    2. Would you be aware if the AI model suddenly changed underneath? How long would it take for you to notice the changes? Would your behavior (within the context of the system) be affected in the long term? 

Read More

02/26/2020 – Sushmethaa Muhundan – Explaining Models: An Empirical Study of How Explanations Impact Fairness Judgment

Sushmethaa Muhundan 1 Comment

The paper explores how people make fairness judgments of ML systems and the impact that different explanations can have on these fairness judgments. The paper also explores how providing personalized and adaptive explanations can support such fairness judgments of ML systems. It is extremely important to ensure algorithm fairness and there is a need to consciously work towards avoiding the risk of amplifying existing biases. In this context, providing explanations can be beneficial in two aspects, not only do they help in providing implementation details which would otherwise be a “black box” to a user, but they also facilitate better human-in-the-loop experiences by enabling people to identify fairness issues. The COMPAS recidivism data was utilized for the study and four different explanations styles were examined: input-influence based, demographic-based, sensitivity-based, and case-based. Through the study, it is highlighted that there is no one-size-fits-all solution for an effective explanation. The dataset, context, kinds of fairness issues, and user profiles vary and need to be addressed individually. The paper proposes providing hybrid explanations as a solution to address this problem thereby providing both an overview of the ML model and information about specific cases to help aid accurate fairness judgment.

While there has been a lot of research focus on developing non-discriminatory ML algorithms, this paper specifically deals with the human aspect which is necessary to identify and remedy fairness issues. I feel that this is equally important and is often overlooked. It was interesting to note that they auto-generated the explanations, unlike previous studies. 

With respect to the different explanation styles used, I found the sensitivity-based explanation particularly interesting since it clearly shows the difference in the prediction result if certain attributes were modified. According to me, this form of explanation, out of the four proposed, is extremely effective in bringing out any bias that may be present in the ML system.

I felt that the input-influence based explanation was also effective since it had the +/- markers corresponding to features that match the particular case and this gives the users a clearer picture of which attributes specifically influenced the result thereby providing the implementation details to a certain extent.

The study results documents various insights from participants, and I found some of them to be extremely fascinating. While some believed that certain predictions were biased, others found it normal for that verdict to be predicted. It truly captured the diversity in opinions and perspectives of the same ML system based on the different explanations provided.

  1. Through this study, it is revealed that the perception of bias is not uniform and is extremely subjective. Given this lack of agreement on the definition of moral concepts, how can a truly unbiased ML system be achieved?
  2. What are some practices that can be followed by ML model developers to ensure that the bias in the input dataset is identified and removed?
  3. Apart from gender-bias and ethnic-bias, what are some other prevalent biases in existing ML systems that need to be eradicated?

Read More

02/26/20 – Vikram Mohanty – Explaining Models: An Empirical Study of How Explanations Impact Fairness Judgment

Vikram Mohanty 1 Comment

Authors: Jonathan Dodge, Q. Vera Liao, Yunfeng Zhang, Rachel K. E. Bellamy, Casey Dugan

Summary

This paper discusses how different types of programmatically generated explanations can impact people’s fairness judgments of ML systems. The authors conduct studies with Mechanical Turk workers by showing them profiles from a recidivism dataset and the explanations for a classifier’s decision. Findings from the paper show that certain explanations can enhance people’s confidence in the fairness of the algorithm, and individual differences, including prior positions and judgment criteria of algorithmic fairness, impact how people react to different styles of explanation.

Reflection

For the sake of the study, the participants were shown only one type of explanation. While that worked for the purpose of this study, there is value in seeing the global and local explanations together. For e.g. the input-influence explanations can highlight the features that is more/less likely to re-offend, and allowing the user to dig deeper into the features by showing a local explanation can help in forming more clarity. There is some scope of building interactive platforms with the “overview first, details on demand” philosophy. It is, therefore, interesting to see the paper discuss about the potentials of a human-in-the-loop workflow.

I agree with the paper that a focus on data oriented explanation has the unintended consequence of shifting blame from the algorithms, which can slow down the “healing process” from the biases we interact with when we use these systems. Re-assessing the “how” explanations i.e. how the decisions were made is the right approach. The Effect of Population and “Structural” Biases on Social Media-based Algorithms – A Case Study in Geolocation Inference Across the Urban-Rural Spectrum by Johnson et al. illustrates how bias can be attributed to the design of algorithms themselves rather than population biases in the underlying data sources.

The paper makes an interesting contribution regarding the participants’ prior beliefs and positions and how that impacts the way they perceive these judgments. In my opinion, as a system developer, it seems like a good option to take a position (obviously, being informed and depends on the task) and advocate for normative explanations, rather than appeasing everyone and reinforcing meaningless biases which could have been avoided otherwise.

Questions

  1. Based on Figure 1, what other explanations would you suggest? If you were to pick 2 explanations, which 2 would you pick and why?
  2. If you were to design a human-in-the-loop workflow, what sort of input would you seek from the user? Can you outline some high-level feedback data points for a dummy case?
  3. Would normative explanations frustrate you if your beliefs didn’t align with the explanations (even though the explanations make perfect sense)? Would you adapt to the explanations? (PS Read about the backfire offer here: https://youarenotsosmart.com/2011/06/10/the-backfire-effect/)

Read More

02/26/2020 – Sukrit Venkatagiri – Interpreting Interpretability

Sukrit Venkatagiri 1 Comment

Paper: Harmanpreet Kaur, Harsha Nori, Samuel Jenkins, Rich Caruana, Hanna Wallach, and Jennifer Wortman Vaughan. 2020. Interpreting Interpretability: Understanding Data Scientists’ Use of Interpretability Tools for Machine Learning. In CHI 2020, 13.

Summary: There have been a number of tools developed to aid in increasing the interpretability of ML models, which are used in a wide variety of applications today. However, very few of these tools have been studied with a consideration of the context of use and evaluated by actual users. This paper presents a user-centered evaluation of two ML interpretability tools using a combination of interviews, contextual inquiry, and a large-scale survey with data scientists.

From the interviews, they found six key themes: missing values, temporal changes in the data, duplicate data masked as unique, correlated features, adhoc categorization, and difficulty of trying to debug or identify potential improvements. From the contextual inquiry with a glass-box model (GAM) and a post-hoc explanation technique (SHAP), they found a misalignment between data scientists’ understanding of the tools and their intended use. And finally, from the surveys, they found that participants’ mental models differed greatly, and that their interpretations of these interpretability tools also varied on multiple axes. The paper concludes with a discussion on bridging HCI and ML communities and designing more interactive interpretability tools.

Reflection:

Overall, I really liked the paper and it provided a nuanced as well as broad overview of data scientists’ expectations and interpretations of interpretability tools. I especially appreciate the multi-stage, mixed-methods approach that is used in the paper. In addition, I commend the authors for providing access to their semi-structured interview guide, as well as other study materials, and that they had pre-registered their study. I believe other researchers should strive to be this transparent in their research as well.

More specifically, it is interesting that the paper first leveraged a small pilot study to inform the design of a more in-depth “contextual inquiry” and a large-scale study. However, I do not believe the methods that are used for the “contextual inquiry” to be a true contextual inquiry, rather, it is more like a user study involving semi-structured interview. This is especially true since many of the participants were not familiar with the interpretability tools used in the study, which means that it was not their actual context of use/work.

I am also unsure how realistic the survey is, in terms of mimicking what someone would actually do, and appreciate that the authors acknowledge the same in the limitations section. A minor concern is also the 7-point scale that is used in the survey that ranges from “not at all” to “extremely,” which does not follow standard survey science practices.

I wonder what would happen if the participants were a) nudged to not take the visualizations at face value or to employ “system 2”-type thinking, and/or b) asked to use the tool for a longer. Indeed, they do notice some emergent behavior in the findings, such as a participant questioning whether the tool was actually an interpretability tool. I also wonder what would have happened if two people had used the tool side-by-side, as a “pair programming” exercise. 

It’s also interesting how varied participants’ backgrounds, skills, baseline expectations, and interpretations were. Certainly, this problem has been studied elsewhere, and I wonder whether the findings in this paper are a result of not only the failure of these tools to be designed in a user-centered manner, but also the broad range in technical skills of the users themselves. What would it mean to develop a tool for users with such a range in skillsets, especially statistical and mathematical skills? This certainly calls for increased certification—at the behest of increased demand for data scientists—within the ML software industry.

I appreciate the point surrounding Kahneman’s system 1 and system 2 work in the discussion, but I believe this section is possibly too short. I acknowledge that there are page restrictions, which meant that the results could not have been discussed in as much depth as is warranted for such a formative study.

Overall, this was a very valuable study that was conducted in a methodical manner and I believe the findings to be interesting to present and future developers of ML interpretability tools, as well as the HCI community that is increasingly interested in improving the process of designing such tools.

Questions:

  1. Is interpretability only something to be checked off a list, and not inspected at depth?
  2. How do you inspect the interpretability of your models, if at all? When do you know you’ve succeeded?
  3. Why is there a disconnect between the way these tools are intended to be used and how they are actually used? How can this be fixed?
  4. Do you think there needs to be greater requirements in terms of certification/baseline understanding and skills for ML engineers?

Read More