Jun 26, 2019 · newsletter

The trouble with unexplainable algorithms

Thanks to a mix of technology-driven disruption and savvy competitors, the business environment is an increasingly challenging one. Staying competitive requires a better understanding of customers’ behaviour and preferences. It also requires the ability to optimise internal processes to more efficiently support both these things. This is a big reason that so many organisations are investing in machine learning.

And yet the use of machine learning in organisations has not been universally welcomed by customers and service users. In fact, in some areas, there’s been a growing push back – or ‘techlash,’ as it has been dubbed. People have expressed concern about a range of things: the breadth of data being collected and the terms of its use, as well as the types of decisions being recommended or made by the algorithms trained on this data. The sources of concern vary, but a lack of trust is the common denominator.

The trust deficit

One corollary of this trust deficit is an increasingly adversarial relationship between organisations and their service users and customers. Lately, I have been thinking about this a lot. It was definitely on my mind when I read about the legal action that Ed Bridges has brought against the South Wales police for violating his privacy through their facial recognition trial. Bridges, who is a private citizen, is doing this with the support of Liberty HQ . This is the first case of its kind in the UK, and it comes hard on the heels of San Francisco city legislature’s decision to ban the use of facial recognition by police and city agencies. Oakland and Massachusetts are also considering similar legislation:

_“Big Sister is watching us,” [says Senate Majority Leader, Cynthia Creem], “and yet we don’t even know how those pictures are being used … The system that they’re using now raises issues of due process and significant issues with regards to civil liberties.” _(

These are widely reported examples of techlash, but they’re on the same spectrum as smaller scale adversarial action against algorithmically-driven decision-making. This latter behaviour was a topic of discussion in a recent episode of the tech podcast, Sleepwalkers. Lisa Talia Moretti, a guest on the show, spoke about students who figured out that a certain selection algorithm prioritised applications from students from universities such as Harvard, Oxford, or Cambridge. They started writing the names of these universities in white text (so human recruiters couldn’t pick up on it) on their CVs, in order to dupe the algorithm.

Image Source: Helloquence on Unsplash

Screening out applications from certain candidates may well have been the intent of the organisation(s) using this algorithm. If so, it’s surprising that the human recruiters didn’t immediately pick up on the fact that applications from less desirable candidates were getting through - in which case, they could have simply screened them out at a subsequent stage or been explicit about their preference for candidates from certain universities. My suspicion is that this wasn’t their intention, and that the emergent selection bias or preference was simply a quirk of the algorithm that was left unexplored and unaddressed.

However this state of affairs came about, the low-level adversarial behaviour it has triggered is corrosive - slowly transforming a potentially positive engagement into a grudge interaction.

Building trust, making friends

So how do organisations go about (re)building trust and - by extension - resetting the dynamic between themselves and their users/customers? There are some really obvious choices namely, clarity and honesty about the way any collected data will be used, and taking steps to protect that data. In this context, being honest is often a matter of being upfront about business models (in the case of commercial entities) as well as governance and decision-making policies (especially in the case of government), and not trying to sneakily change models and policies through privacy policy updates.

However, achieving the requirement for clarity is a tougher challenge. There are a number of reasons for this, but I will simply touch on two:

  • Explaining how complex systems interact with the various features of a large dataset - in a way that’s easy for time-poor, non-experts to grasp - is hard.
  • Sometimes,(especially in the case of so-called ‘black box’ models) organisations don’t know quite how their algorithms arrive at the results that they do.

These problems are related: explaining the workings of a system you don’t really understand is a tall order. That’s why attaining this understanding should be a critical area of focus for organisations that use machine learning. Thankfully, approaches for doing this are becoming more mature; this was the focus of our Interpretability research report. Understanding the way systems work also paves the way for better, and more considered, trust-building conversations with end-users. (If you’re curious about what this looks like in practice, I recommend reading this blog post about “designing to expose the seams of [automated decision making].” It’s based on a collaboration between the design agencies Comuzi and Projects by If.)

With the advent of the General Data Protection Regulation (GDPR), there has been a lot of focus on explainability as a regulatory requirement. There is great benefit for organisations in recognising the value of interpretability and explainability as tools for building trust with customers.

Read more

Jul 8, 2019 · post
Jun 26, 2019 · newsletter

Latest posts

Sep 22, 2021 · post

Automatic Summarization from TextRank to Transformers

by Melanie Beck · Automatic summarization is a task in which a machine distills a large amount of data into a subset (the summary) that retains the most relevant and important information from the whole. While traditionally applied to text, automatic summarization can include other formats such as images or audio. In this article we’ll cover the main approaches to automatic text summarization, talk about what makes for a good summary, and introduce Summarize. – a summarization prototype we built that showcases several automatic summarization techniques. more
Sep 21, 2021 · post

Extractive Summarization with Sentence-BERT

by Victor Dibia · In extractive summarization, the task is to identify a subset of text (e.g., sentences) from a document that can then be assembled into a summary. Overall, we can treat extractive summarization as a recommendation problem. That is, given a query, recommend a set of sentences that are relevant. The query here is the document, relevance is a measure of whether a given sentence belongs in the document summary. How we go about obtaining this measure of relevance varies (a common dilemma for any recommendation system). more
Sep 20, 2021 · post

How (and when) to enable early stopping for Gensim's Word2Vec

by Melanie Beck · The Gensim library is a staple of the NLP stack. While it primarily focuses on topic modeling and similarity for documents, it also supports several word embedding algorithms, including what is likely the best-known implementation of Word2Vec. Word embedding models like Word2Vec use unlabeled data to learn vector representations for each token in a corpus. These embeddings can then be used as features in myriad downstream tasks such as classification, clustering, or recommendation systems. more
Jul 7, 2021 · post

Exploring Multi-Objective Hyperparameter Optimization

By Chris and Melanie. The machine learning life cycle is more than data + model = API. We know there is a wealth of subtlety and finesse involved in data cleaning and feature engineering. In the same vein, there is more to model-building than feeding data in and reading off a prediction. ML model building requires thoughtfulness both in terms of which metric to optimize for a given problem, and how best to optimize your model for that metric! more
Jun 9, 2021 ·

Deep Metric Learning for Signature Verification

By Victor and Andrew. TLDR; This post provides an overview of metric learning loss functions (constrastive, triplet, quadruplet, and group loss), and results from applying contrastive and triplet loss to the task of signature verification. A complete list of the posts in this series is outlined below: Pretrained Models as Baselines for Signature Verification -- Part 1: Deep Learning for Automatic Offline Signature Verification: An Introduction Part 2: Pretrained Models as Baselines for Signature Verification Part 3: Deep Metric Learning for Signature Verification In our previous blog post, we discussed how pretrained models can serve as strong baselines for the task of signature verification. more
May 27, 2021 · post

Pretrained Models as a Strong Baseline for Automatic Signature Verification

By Victor and Andrew. Figure 1. Baseline approach for automatic signature verification using pretrained models TLDR; This post describes how pretrained image classification models can be used as strong baselines for the task of signature verification. The full list of posts in the series is outlined below: Pretrained Models as Baselines for Signature Verification -- Part 1: Deep Learning for Automatic Offline Signature Verification: An Introduction Part 2: Pretrained Models as Baselines for Signature Verification Part 3: Deep Metric Learning for Signature Verification As discussed in our introductory blog post, offline signature verification is a biometric verification task that aims to discriminate between genuine and forged samples of handwritten signatures. more

Popular posts

Oct 30, 2019 · newsletter
Exciting Applications of Graph Neural Networks
Nov 14, 2018 · post
Federated learning: distributed machine learning with data locality and privacy
Apr 10, 2018 · post
PyTorch for Recommenders 101
Oct 4, 2017 · post
First Look: Using Three.js for 2D Data Visualization
Aug 22, 2016 · whitepaper
Under the Hood of the Variational Autoencoder (in Prose and Code)
Feb 24, 2016 · post
"Hello world" in Keras (or, Scikit-learn versus Keras)


In-depth guides to specific machine learning capabilities


Machine learning prototypes and interactive notebooks


A usable library for question answering on large datasets.

Explain BERT for Question Answering Models

Tensorflow 2.0 notebook to explain and visualize a HuggingFace BERT for Question Answering model.

NLP for Question Answering

Ongoing posts and code documenting the process of building a question answering model.

Interpretability Revisited: SHAP and LIME

Explore how to use LIME and SHAP for interpretability.


Cloudera Fast Forward is an applied machine learning reseach group.
Cloudera   Blog   Twitter