This paper describes how machine learning systems can be explained and demystified for non-technical audiences through the use of an online simulation. This research is the result of a European Union funded project, SageCare, which focuses on developing machine learning systems to classify clinical and genomic data. In disseminating the use of machine learning to non-specialists we often encounter resistance or suspicion on the veracity of approach. Hence, we present artificial intelligence/machine learning for non-specialists and present a case study and an interactive simulation on how machine learning can be used in cancer diagnostics. The simulation system serves as a basis for both informing clinical practitioners how machine learning can be used to build diagnostic models and describes how feedback from users will be gathered and analyzed to assess how machine learning is viewed in such an application.
The SageCare Project [
Cancer is one of the diseases which has a huge impact on patients and their families, so understanding how artificial intelligence can be leveraged to aid diagnosis is important in order to help find ways to alleviate the prevalence of this disease. This paper outlines how machine learning driven artificial intelligence (AI) can be used to aid diagnosis of cancer by building a model that assesses visual input features of cell nuclei. It also serves as a useful example to non-specialists interested in AI to help them understand the dynamics of machine learning algorithms and to understand how to assess their performance.
Breast cancer is one of the most commonly occurring cancers, with over 2 million
new cases diagnosed globally every year [
The mean, max and standard error of each feature are computed for each image to give a total of 30 input features per sample, which are suitable for machine learning. 3
Machine learning is a computational approach to AI that uses algorithms that iterate
over datasets to build statistical models [
The SVM is a supervised learning algorithm that has been shown to have good
performance as a classifier [
One approach to gauging the performance of the classifier is to compute the F1 score,
which is a useful measure of the level of precision and recall in a machine learning
system [
An algorithm with high recall will classify most of the relevant data correctly and can be thought of as the ratio of correctly classified true positives tp, over the sum of true positives tp and false negatives fn (the number of instances falsely classified as negative instances):
There is a trade-off between precision and recall as it is possible to have an algorithm with high precision but low recall and vice versa. For example, the algorithm may be precise by correctly classifying a subset of malignant breast cancer cases, however it could achieve this by being stringent in its classification and could exclude many other malignant cases, which would give it a low recall.
The balance between precision and recall can be captured using an F1 score which
is the harmonic mean of the precision and recall scores, where a score of 1 indicates
perfect precision and recall [
1 =
1
The machine learning model should be trained in such a way that the algorithm does not overfit, which occurs when the algorithm fits a decision boundary tightly to all of the data, including any noise in the training data, so that it generalises poorly to any unseen input. To avoid over-estimation of model performance, a test data set is held back and is used as the final unbiased measure of the algorithm’s performance on the training data. A model that produces a high score on the training set but a low score on the test set will overfit, while a model that produces a high score on the training set and a high score on the test set should provide good classifications. A model that underfits, by failing to find any useful decision boundary will perform poorly on both data sets.
The simulation in Figure 3 shows the F1 scores for the algorithm on the training set and the test set, thereby allowing users to gauge the performance of the algorithm. This also challenges the user to investigate how tuning the hyper-parameters for a machine learning algorithm affects its performance and so will enhance their understanding of the dynamics of the problem.
In the example demonstrated to the non-specialist audience a two-dimensional feature space was presented. The simulator developed by Karpathy was enhanced to allow users to select which feature they would like to evaluate. Figure 3 shows use of a linear kernel when trying to find the ideal separation, and the F1 score for both the training set and test set is shown. The choice of using a kernel is an important machine learning hyperparameter; practitioners needs to consider if the data set is linearly separable or not. This simulation is presented as a game to the users, where the goal is to reach a perfect F1 score of 1 on the test data set.
The use of a nonlinear kernel is shown in Figure 3. Choosing a non-linear kernel for
a linear data set will tend to cause the model to over fit the data, which will reduce its
ability to generalize as indicated by a poor performance on the test data set F1 score.
SVMs use a technique known as the kernel trick which maps data points to a higher
dimensional space where a linear separation may be found [
For non-linear kernel the Karpathy SVM JavaScript framework uses a Gaussian radial basis function, which allows the SVM algorithm to fit the maximum margin separating hyperplane in a transformed input feature space. The radial basis function is controlled by the parameter sigma (σ), which determines the influ ence that feature vectors have on the kernel mapping. Intuitively, low values of sigma narrow the region of influence of the kernel for vectors in the feature space, which can cause the SVM to overfit the data. High values of sigma widen the region of influence, making the algorithm better at generalizing at the expense of losing precision. Users can experiment with features, kernels and hyper-parameters as shown in the adaption of Karpathy’s software, see Figure 3 (b). Communicating the effect of σ and other parameters, to nontechnical audiences in a visual manner supports the objective of this study; to investigate how an interactive tool enhances their understanding and user of machine learning tools. 4
In the last decade, we have witnessed a growing interest on AI applications. Numbers of commercials on a variety of everyday objects (e.g. mobile phones, vacuum cleaners, thermostats etc.), present AI as an important added value. Indeed, it is. With AI powered cameras we can get better photographs, with an autonomous vacuum cleaner we can gain more spare time and with a smart thermostat we can save a lot of money on heating bills.
However, despite this growing interest, very often AI is perceived, by the general public, like a sort of magic, or, to put it in the words used by Arthur C. Clarke, an advanced technology indistinguishable from magic. But such perception has a problem in that it presents a technology’s inner mechanisms as being incomprehensible. Rather, AI is a powerful tool that can be harnessed for personal or professional purposes, even by lay people thanks to off-the-shelf software packages in which users are requested only to add their data. Adding data alone, however, could simplify the process too much, preventing users from grasping the inner mechanics of what they are using, leading to potential mistakes or misuse.
AI, undoubtedly, represents an effective tool to solve or to simplify many relevant problems in our daily lives. For this reason, we should disclose as much as possible about how certain algorithms work. Such an operation can be beneficial to improve a basic understanding of a particular algorithm, whether a user is just willing to better grasp a topic or where a user is interested in using that kind of technology with and increased awareness for his own purposes. This can provide the following benefits of explaining the technology to potential users as they can: 1) make better use of the tools, 2) better understand the problems it can and cannot solve and 3) make a more informed assessment and evaluation of the produced solution 5
In order to understand which kind of information should be conveyed about an
intelligent technology in general and in particular to a classifier system such as that presented
in this paper, we organized a focus group, held in Rome in December 2018, with a small
number of participants. Focus groups have a long tradition in behavioral sciences where
have been used to understand how an issue or a product is perceived by a group of
people [
Seven participants, 1 woman and 6 men, with an average age of 37.57 (SD= 6.23) with a background in AI research were identified within an Italian research center. For the aim of this exploratory work, participants were chosen for their unfamiliarity with support vector machines although well-versed in other AI techniques such as neural networks, genetic algorithms etc. The focus group was organized with the following structure: 1) Short introduction on the topic of the focus group 2) Brief presentation of the participants 3) Discussion of the topic: • Question 1: How do you think people perceive AI? • Question 2: Do you think it is necessary to explain how AI works? • Question 3: Which features of a classifier system should be stressed for educational purposes? 4) Presentation of SVMs through the software we developed 5) Request of feedback on the software as an educational tool
The focus group lasted three hours and stimulated a very interesting discussion on the general importance of AI in our lives and the features that should be shared in order to increase people’s awareness on specific algorithms.
The first question raised a dualistic view on AI. To the participants, people seem to see AI either as the ultimate evil or magic. Both polarized views, however, lack a realistic perspective and all the participants agreed with the fact that AI at the moment is an inflated word due to marketing purposes.
The second question divided the participants. Three of them underlined that in order to understand many AI algorithms a strong background in maths is needed, hence it is impossible to provide such kinds of concepts to a general public, who likely lack specific hard skills. The rest of the group in different ways highlighted the need to explain how algorithms work. In particular, two proposals emerged on how best to explain how AI works to members of the general public: a) using a very simple language without referring to mathematical jargon; and b) demonstrating algorithm with micro-educational software in which users can manipulate data and parameters.
The final question firstly collected a series of answers related to the fact that the outcome of classifiers system, regardless of the algorithm that is being used, is strictly connected to the data we put in. Secondly, although sometimes very complicated math is required to understand the specific aspects of an algorithm, an extremely simple formula can often be used to evaluate the outcome of a classifier (see for example precision and recall).
After the discussion raised by the first three questions, we presented our software (by explaining the objectives and the algorithm behind it) to the participants and asked them if it was, in their opinion, viable for educational purposes.
Participants appreciated two aspects of the software: 1) that is web-based and it is able to seamlessly run on mobile phones without issues, and 2) the two windows easily allow for seeing what happens to the outcome when different parameters are applied to the underlying algorithm. Less appreciated was the graphical aspect. Participants suggested to improve the graphics in order to make the training and the testing sets more visible.
An overall positive consideration emerged about the possible use of the software as an educational tool. Although not experts in SVMs, they understood how this type of classifier works in a relatively simplified way (here we remember that the participants shared an AI background) 7
AI and its applied arm, machine learning, are becoming an important part of our daily life. Our mobile phone recognizes our vocal commands and the AI powered cameras can take pictures with a professional quality. Applications, however, are not limited to our spare time. AI can also be added to the toolkit of our professions: a biologist or a psychologist, for example, could exploit a machine learning solution for their own purposes. In particular, a biologist could use a classifier such as the one described in this paper to classify his/her own data points or re-run the algorithm with new collected data. In order to do that it is not required to be a data scientist but just to be able to use an off-the-shelf solution with a proper awareness.
Qualitative data collected in an exploratory focus group seems to suggest that our approach goes in this direction, however, in order to evaluate its effectiveness we need quantitative data. Gathering this quantitative data will be the objective of the next step of this research.
PW is supported by funding through Science Foundation Ireland (SFI) Multi-Gene Assay Cloud Computing Platform - (16/IFA/4342), BK, JL, CP, OG and RDF are funded under H2020 RISE project SemAntically integrating Genomics with Electronic health records for Cancer CARE (SageCare), grant number 644186. Available: