Elod P Csirmaz’s Blog: 2020

28 June 2020

Interpreting Neural Networks by Reducing Nonlinearities during Training

Machine learning and neural networks are used more and more widely, despite the fact that they mostly function as black boxes, and we have little insight into how they generate their output. This makes it difficult to scrutinize and account for decisions made using these systems.

In this note I describe a simple solution that makes it possible to extract rules from a neural network that employs Parametric Rectified Linear Units (PReLUs). This is done by introducing a force, applied in parallel to backpropagation, that aims to reduce PReLUs into the identity function, which then causes the neural network to collapse into a smaller system of linear functions and inequalities suitable for review or use by human decision makers.

As this force reduces the capacity of neural networks, it is expected to help avoid overfitting as well.

Introduction

Machine learning solutions and, more specifically, neural networks are used in more and more areas of our lives. At the same time, since it is difficult to grasp at any level how they function and so they can be seen as black boxes, there is a reluctance in their adoption in the case of companies who need to remain accountable for business and/or legal reasons. This fact can also impact the trust the public places in such systems.

Because of this, providing tools that can help us understand how these models arrive at a certain decision is an area of active research. The tools range from visualization tools, tools that identify the most significant input, tools that allow experimenting with the input data to understand the relative importance the model attributes to different signals, and, the focus of this note, significantly reducing the complexity of a neural network to the point where “rules” for decisions could be extracted from it, which can then be reviewed or even used by human decision-makers.

One form these rules can take is a closed, relatively simple algebraic expression (for example, a linear function) which has a behavior more predictable for humans than the full function implemented by a neural network involving parameters and nonlinear functions in the range of millions.

In this note I describe a method of reducing the complexity of a neural network during its training, which then allows describing its output as a simple system of linear functions limited to regions defined by linear inequalities. In particular, feed-forward neural networks are considered where all nonlinearities are Parametric Rectified Linear Units (PReLUs) defined as

f(x) = max(0, x) + a min(0, x),

where 0 ≤ a ≤ 1.

Let the input of the network be the vector (or tensor, which we flatten) I = (i1 , i2 , . . . , in) and the output be the vector (or tensor) O = (o1 , o2, . . . , om).

The Neural Network as a Combination of Linear Functions

It is easy to see that the output O of such a network can be expressed as a set of linear combinations of ix, each restricted to a particular region of input values by a set of linear inequalities on ix (the conditions).

Intuitively, this is true because the output is a linear function of the input as long as (that is, inside regions of the possible input vectors in which) none of the inputs of PReLUs cross from negative to positive or vice versa; and because the borders of these regions are linear as well.

More formally, we can show that this statement is true by constructing a suitable system of linear combinations and inequalities for any given neural network. For all PReLUs in the network, let us decide whether their inputs are negative, or nonnegative. If the number of PReLUs is δ, this means 2δ possible scenarios. In each scenario, the neural network becomes completely linear, with the output being a linear function of the input, and, in fact, with all partial results inside the network being linear functions of the input as well. This entails that the inputs of PReLUs are also linear functions of the input I, yielding the δ linear inequalities (the conditions) that mark out the region of the input space where this scenario applies.

Naturally, this system of 2δ linear combinations yielding the output, each controlled by δ conditions is likely to be redundant, as the conditions may not be independent, or may be contradictory, in which case some of the 2δ regions of the input space will be empty. But this system will be an accurate description of the output of the neural network nevertheless.

While a linear function may seem easy to interpret and to predict for us humans, clearly 2δ functions will not be so if δ is in the range of thousands or millions as in most modern neural networks. We therefore need to construct an approximation of the output of our neural network which is linear in much larger regions of the input space, and has far less different regions as well. This is equivalent to trying to find another neural network that approximates our network well, but contains a small number of PReLUs only.

Reducing PReLUs into Linearities

This can be achieved this by continuing training the original neural network while applying a force on each of its PReLUs that moves its parameter a towards 1.

Note that if a = 1, then the PReLU degenerates into the identity function f(x) = x, and ceases to be a nonlinearity. In a sense, it disappears, and the neural network “collapses” around it, inasmuch as the linear mapping that precedes the degenerate PReLU and the one after it can now be combined into a single linear map.

This force therefore finds a balance between approximating the training data and reducing the number of PReLUs, thereby yielding a neural network that is feasible to express as a set of linear functions and inequalities for human consumption. By removing nonlinearities, this force also reduces the capacity of the neural network, and therefore it can be expected to help avoid overfitting.

I chose a force that is independent of a and is applied after each backpropagation step, and moves the parameter by an adjustment rate ηp but never above 1:

at+1 = max(0, min(1, at + ηp sgn(1 − at)))

In my research I considered any PReLU fully linear if

a > 0.9995

ηp was also made dependent on the current error exhibited by the network, to allow the network initially to train without interference, and then pull the PReLU parameters to 1 more and more aggressively. In particular,

ηp = η0 0.01 max(0, − log10(err) − 2)

where η0 = 0.01 is the learning rate for the whole model, and err is the training error. This means that ηp will start becoming non-zero when the training error falls below 0.01, and can grow indefinitely as the training error falls (although the a parameters are clipped anyway).

Example

Sample code demonstrating the above is available at https://github.com/csirmaz/trained-linearization. The code implements training a small neural net with the force on PReLU parameters, and contains logic to extract a system of linear inequalities and combinations from the weights and other parameters of the model.

The neural network itself has 4 input nodes, 3 output nodes, and 4 hidden layers in between with 5 nodes on each. These layers are linked by trainable fully-connected layers with PReLU activation following them, except for the last fully connected layer, which simply provides the output.

The network is expected to learn the following relationships:

out1 = in1 xor in2
out2 = in3 xor in4
out3 = in1 · in2

where all the inputs and outputs are 0 or 1 with some extra noise added to the inputs. The code generates the training data according to the expressions above, for all 24 possible combinations of 0 and 1 for the input. New noise is generated for each batch of training data.

The figure below visualizes a trained model. Two nonlinearities are enough to solve this problem, and indeed the training settled on a solution where only two PReLUs were nonlinear.

(There are 4 input nodes (top) and 3 output nodes (bottom). Red lines denote positive, blue lines negative weights. Lines leading from the right denote bias values. Green boxes mark PReLUs that are considered linear, while red ones mark nonlinear PReLUs.)

Although in the visualization the relationships between the inputs and the outputs are far from clear, once the rest of the PReLUs are considered linear, all the weights collapse into very simple expressions. See below for the output of the code listing these expressions for the 22 cases arising from the two nonlinear PReLUs. These correspond completely to the relationships in the training data.

IF +0.00*in1 +0.00*in2 -1.00*in3 -1.00*in4 +1.00 < 0
  (PReLU #3 on level 1 is neg. ln(1-weight)=-2.58)
IF +1.00*in1 +1.00*in2 -0.00*in3 -0.00*in4 -1.00 < 0
  (PReLU #1 on level 4 is neg. ln(1-weight)=-2.60)
THEN
  out1 = +1.01*in1 +1.01*in2 -0.00*in3 -0.00*in4 +0.00
  out2 = -0.00*in1 -0.00*in2 -1.03*in3 -1.03*in4 +2.04
  out3 = -0.00*in1 -0.00*in2 +0.00*in3 +0.00*in4 -0.00

IF +0.00*in1 +0.00*in2 -1.00*in3 -1.00*in4 +1.00 < 0
  (PReLU #3 on level 1 is neg. ln(1-weight)=-2.58)
IF +1.00*in1 +1.00*in2 -0.00*in3 -0.00*in4 -1.00 > 0
  (PReLU #1 on level 4 is pos. ln(1-weight)=-2.60)
THEN
  out1 = -1.01*in1 -1.01*in2 +0.00*in3 +0.00*in4 +2.02
  out2 = -0.00*in1 -0.00*in2 -1.03*in3 -1.03*in4 +2.04
  out3 = +1.00*in1 +1.00*in2 +0.00*in3 +0.00*in4 -1.01

IF +0.00*in1 +0.00*in2 -1.00*in3 -1.00*in4 +1.00 > 0
  (PReLU #3 on level 1 is pos. ln(1-weight)=-2.58)
IF +1.00*in1 +1.00*in2 +0.00*in3 +0.00*in4 -1.00 < 0
  (PReLU #1 on level 4 is neg. ln(1-weight)=-2.60)
THEN
  out1 = +1.01*in1 +1.01*in2 +0.00*in3 +0.00*in4 +0.00
  out2 = -0.00*in1 -0.00*in2 +1.00*in3 +1.00*in4 +0.01
  out3 = -0.00*in1 -0.00*in2 -0.00*in3 -0.00*in4 +0.00

IF +0.00*in1 +0.00*in2 -1.00*in3 -1.00*in4 +1.00 > 0
  (PReLU #3 on level 1 is pos. ln(1-weight)=-2.58)
IF +1.00*in1 +1.00*in2 +0.00*in3 +0.00*in4 -1.00 > 0
  (PReLU #1 on level 4 is pos. ln(1-weight)=-2.60)
THEN
  out1 = -1.01*in1 -1.01*in2 -0.00*in3 -0.00*in4 +2.02
  out2 = -0.00*in1 -0.00*in2 +1.00*in3 +1.00*in4 +0.01
  out3 = +1.00*in1 +1.00*in2 -0.00*in3 -0.00*in4 -1.01

References

[1] Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun, “Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification,” Microsoft Research, 6 February 2015. http://arxiv.org/abs/1502.01852.

6 June 2020

Buying shallow computer desks

I now have a small office, and wanted to buy some furniture including a chest of drawers that fit next to the door and under the light switch. To my surprise, this turned out to be one of the most frustrating internet searches, until I found a website that allows one to search for and buy beds, tables, sofas and so on by size and price, for example to buy shallow computer desks. Happy furniture hunting!

5 May 2020

Remote Management in 9 Simple Steps: A Quick Practical Guide to Transitioning Away from the Office

Introduction

This booklet offers practical ideas to team leaders and managers who suddenly find themselves organizing the work of and managing their team remotely and are not entirely sure where to start.

The coronavirus pandemic means that many of us are now asked to work from our homes. However, for many team members and leaders remote working is uncharted territory, and efficiency and worker satisfaction can suffer without a new kind of management tailored to this new kind of working. This book builds on years and years of experience in remote working, hiring and managing to help you avoid many pitfalls and make remote working a long-lasting success. You will not only allow your team to match the performance and satisfaction levels of its old office-based version, but also to surpass these by creating new channels of communication inside the team, and a much improved work-life balance for everyone.

1. Basic Equipment and Setup

Most work that can be done remotely will be done in front of a computer, so everyone will need a decent desktop or laptop and an internet connection with sufficient bandwidth to support video calls. The computer could be their own, but if the company is able to supply a separate computer for work purposes, security can be easier to manage - see below.

If the job involves handling (physical) post, arranging postal redirection and collection may also be needed. There are companies that offer a service where they print and post documents sent to them electronically so trips to the local post office can be avoided. Providing a scanner for the home office can also help to digitize paper documents and distribute them electronically inside the company, which reduces the carbon footprint and avoids any danger of passing on an infection.

There are two legal points to mention here. One is that any legal requirements or insurance policies need to be considered if documents are stored outside the registered address of the company. The other is that depending on where the company is registered, it may be required to ensure that all employees have an adequate work environment, including space, seating and lighting, some of which may be relevant for the home office as well.

There is no shortage of software that allows calls, video calls and chat so that team members can communicate. Depending on the pricing and security requirements, one can choose, for example, Slack, Skype, or Google Hangouts/Meet, to name a few. It is important to provide training, if necessary, so that everyone can use these tools efficiently. One should, at minimum, be able to send direct messages, create group chats, initiate voice and video calls, mute their microphone, invite others to existing calls, start screen sharing, and post images.

Having chosen one of the available services, some thought should be given to creating permanent group chats for various purposes. There can be a company-wide group chat or “room” for some social chit-chat, saying hello in the mornings, birthday wishes, and general announcements. Team-specific chats are also helpful for the same purposes as well as for more work-related communication: asking for help and assistance, sharing knowledge, announcing any changes or updates to the systems used, and so on. These permanent chats should have clear guidelines for their usage that can usually be “pinned” to the chat for future reference and for everyone who joins later.

And finally, to successfully transform any in-person meeting where pen and paper or a whiteboard is used to the remote world (for example, presentations, brainstorming, sketching out diagrams, graphs, relationships, etc.), a digitizer tablet at home can be an invaluable tool. These are usually small plastic tablets with pens with which one can control the mouse cursor and write or draw on the screen. They should “just” work when plugged in under most operating systems. Then if one is presenting in an online meeting and needs to sketch a graph or note something, they can either use a local drawing app to draw on and share their screen in the call, or use an online whiteboard app that handles the sharing itself and can even allow others to draw on the same board. There are numerous such whiteboard apps, even entirely web-based ones, too, that do not require installing any software.

2. Online Etiquette

Home working brings the office into the very space of our lives where we feel relaxed, dress comfortably, and feel free to let our guard down. To ensure that this is balanced with the usually more formal work relationships, and avoid embarrassment, it can be helpful to remind everyone in the team of some basic etiquette when communicating in a chat or video calls.

Everyone should be aware of what is visible (and audible) in a video call, from their dress (even if they need to stand up), their background (books, pictures, posters), others in the household (e.g. through an open door or in the same room), to music played by others in the family, babies crying, construction work, and so on and so forth. There should be nothing intimidating or embarrassing in the extra things seen or heard, and noise that can make the online call difficult or impossible should also be minimized. The company can also require some minimum dress code for home working, too. If one works with sensitive data, it is important to minimize visibility of this for others in their household. The reverse is also true: no one in the team would like to deal with witnessing some sensitive information from another company that someone else in the household works for.

For those who have not used online chats much, it can be difficult to find the right tone, and it may be difficult to realize the differences between spoken and written language. A spoken sentence, when written down, is more often than not ambiguous or lacks crucial information about the speaker’s intent (is it sarcastic? tongue-in-cheek? a real question or a rhetorical one?). In these cases, one may need some assistance or gentle feedback that their messages may not be entirely appropriate, or that they could be easily misread as hurtful or inconsiderate. Conversely, for those who have used chats much, but in an entirely different context (for example, among friends, gaming, etc.) may need to readjust to using a chat for professional communication.

It goes without saying that no communication in an online chat should be acceptable that is not acceptable when spoken or written at the office: language that is in any way sexist, racist, homophobic, ageist, threatening, intimidating, disrespectful or creates a feeling of exclusion should have immediate consequences. In general, it can be worth reminding team members that it is better to be safe than sorry, and there is nothing wrong with waiting a few seconds to reconsider a message and how it may be interpreted before sending it.

Other chat-related communication around which a manager may wish to set up some basic guidelines are avatars and emojis. Outside work context, in the often anonymous world of online chats and forums, avatars (profile images) are a wonderful tool to communicate something about ourselves or assume a slightly different personality. At work, however, these may not be appropriate, and it may make sense to stick to the professional photo from, for example, our LinkedIn profiles.

Emojis are fun and an efficient way of communicating, but once more one should be aware of any potential pitfalls or misinterpretations. They are so short that they almost entirely lack any kind of context. Will a poo emoji correctly convey that I agree how bad the situation is my colleague has found themself in, or will they think I’m saying it’s up to them to deal with it? A thumbs-up emoji is a quick and easy way of saying I understand or agree with something, but if there are different skin shades, can choosing the Caucasian one be seen as being proud of not being colored?

All in all, the overall tone of chats and chat rooms will evolve and will be influenced by everyone at the company. But it does reflect what the company is and its culture, so some guidelines can and should be present to ensure that it remains respectful and inclusive.

3. Security

Your company probably already has a number of security policies regarding computers, software, passwords, and so on. It is important to revisit and rethink these in the light of remote or home working. As the security requirements of each company are different, here I will only list factors that are markedly different in a home and an office setting and which should receive some consideration.

If team members are using their own computers to work, it is going to be difficult to ascertain that the operating system is kept up-to-date, that security patches are applied in a timely manner, and, if necessary, some up-to-date antivirus software is installed. Moreover, it is going to be practically impossible to audit all software that is installed on the computer for personal use, which may contain malware that may access what is displayed on the screen or what is being typed, and so may steal company credentials or sensitive data.

Some of these risks can be managed by requiring everyone to use a separate user account for work purposes. I think it is also helpful to avoid mixing up personal and work files and avoid revealing more of one’s personal life than one wished to.

The safest option, however, remains for the company to supply the computer to be used, on which appropriate software can be installed, and which, with the right tools, could be managed remotely to ensure that it stays secure and updated. It also becomes possible to limit what software can (or should only) be installed, which reduces the risk of accidentally installing some malware.

To authenticate team members before they can access company servers or data, a virtual private network (VPN) can be used. Once a user joins, they will see the same servers and resources they would see in the office on the internal network of the company, even if they are at home or in a cafe. Moreover, VPNs can be configured to send all network traffic through the company servers, which, with some extra setup, can be used to detect and avoid information leaks by, for example, malware. The IT department should be able to set up a VPN server if this looks to be an appropriate tool for your remote workers.

Companies like Microsoft or Google offer a number of services for businesses which can be helpful for remote working. Some services allow remotely controlling software on the user’s computer, or they may be cloud-based services (that is, running on Microsoft’s or Google’s servers, like online documents), which can be easily accessed from anywhere in the world, and managed in a centralized way.

Computers in a home setting are also different because they can be physically accessed by people who are not members of or related to the company. This means that team members should be extra careful when leaving their computer unattended. It can be helpful to establish policies around locking the screens, password strength, or even turning the computer off when not in use. This latter can be important when the disk (storage) of the computer is encrypted to protect sensitive business information. When the computer is powered on, it stores the key to decrypt the disk, and so disk encryption offers limited protection if the computer is accessed or stolen when it is on.

4. Social Fabric

Home or remote working can be a lonely experience. Our aim is still to build a well-functioning team, but it appears to be a hopeless effort if its members are separated by cities or even countries and time zones. To combat the feeling of isolation and loneliness, and build cohesion in the group, it is important to replicate, as much as possible, the extra levels of communication that naturally arise in an office setting.

In an office, we naturally start a conversation about where to get food for lunch, which easily leads to a discussion of yesterday’s cooking, favorite restaurants, or upcoming barbecues. It takes a bit of effort to replicate such social chit-chat virtually, but it is essential and very rewarding.

Team leaders can encourage this and also lead by example. At the beginning and end of one-to-one calls, have a little chat about the weather, pets, recent events in the colleague’s life, their family and friends. Towards the end of the day, and especially before holidays and weekends, if possible, call everyone and have a chat about their plans for the evening or upcoming days. Make sure to remember and refer to any challenges or problems colleagues have mentioned like illnesses and life events. This is only common courtesy, and most often this comes naturally because as team members, we are interested in each other. However, I find that it is easy to forget to allocate time to discuss these in an online setting, so some conscious planning and focus is needed to maintain this social contact. Team members may also be more self-conscious of time spent chatting online and so may unintentionally avoid doing so. It should be made clear that it is fine (in moderation, of course). After all, when working from home, we will not bump into each other on our way to the bathroom, the vending machine or the canteen and have a chat as one would do in the office.

I have mentioned company- and team-wide chats above, and these also play an important role in maintaining and building cohesion. Encourage posting work-related as well as interesting news, suggestions for things to do (Did you like this movie? Is there a game on?), congratulatory notes for work anniversaries and birthdays, and so on. It can be good practice to say hello in the morning on the company-wide chat, as well as good evening at the end of the day. One can also organize chat-based quizzes or participating in web-based or video games to replace the Friday evening drinks traditional at some companies.

And finally, encourage and do enable video on calls whenever possible. It is much easier to build a connection with someone if we see their face, and seeing the audience and their nonverbal reaction can be of enormous help to whoever is speaking or presenting to understand if their message is getting through; if they need to repeat, slow down, or speed up. In the absence of a video link, one needs to constantly request audible feedback (e.g. “Say OK if you’re with me”, “Does that make sense?”, etc.) which makes the presentation fragmented and much longer than necessary.

5. Managing in Absence

I can only repeat that remote working can be a lonely experience. Members of the team will not benefit from seeing or overhearing who others turn to with their problems, a manager walking past their door or being ready to be interrupted with a problem or request. Overall, they may be uncertain how to raise a problem and with whom. This is especially true of new employees who join an already remote setting, and who cannot carry over existing physical relationships to the virtual world.

To manage these problems, it can be helpful to establish, as closely as possible, unity of command, where everyone reports to just one superordinate person, making it clear where to turn to with any issues.

I have also found that managing “by walking around” works extremely well in a remote setting - just calling everyone at least once a day, one after the other, provides an excellent opportunity for a break for the team member, some socializing, and most importantly, an immediate opportunity to see, discuss and resolve any problems they may have. I find these calls work best when they are not pre-scheduled, but at the same time I have always avoided questions that imply that I am conducting a random check to see if they are working or slacking off. It should be much more a social visit and an opportunity for any feedback to me than a surprise inspection to fear.

6. Empower and Decentralize

While it is possible to monitor remote workers using software that constantly shares their screen, or webcams and/or microphones that are continuously on, I think in general there is less control over a team member when they work remotely. This should be an opportunity to rejuvenate and improve the team structure, and not something to dread.

If team members cannot be controlled, hand control to them. Micromanagement is unlikely to work, so avoid it from the get-go. Make them responsible for a particular job or task, including contacting and chasing others for information or actions. It has long been suggested that we perform better if we feel personally invested in something, if we feel ownership over something - this is what we aim to achieve and build on here, as opposed to expecting team members to slavishly follow orders and procedures.

Naturally, this kind of working may not suit everyone to the same degree, but in my experience, everyone can benefit from tending towards assuming responsibility and ownership. While this allows (acknowledges) a certain amount of freedom of how the work is carried out, it also means that one is expected to report and be responsible for the success of the work, and proactively speaking up as soon as they realize that some aspect of the requirements for the work, as originally specified, will not be achieved: a deadline will likely be missed, or some specification of the product, on further investigation, is contradictory or ambiguous. We can vary how much freedom is allowed when carrying out the work itself and how frequently reports are expected, and these allow tailoring this management style to junior and senior team members, and to different personalities.

Once a team member is responsible for most aspects of their work, they should be free to contact others (inside and outside the team) for whatever is required for the task at hand. Such horizontal connections in the business, and ad-hoc online calls, should be encouraged as much as possible. They cut out people in the middle and speed up communication, reduce the burden on line managers in an online environment where communication can be more cumbersome, build cohesion and a sense of belonging in the whole company, and offer opportunities for socializing and stepping out of working in isolation.

7. Feedback, Feedback, Feedback

Because remote and home working can be a new experience for many, and because we do not see remote team members unless we actually call them and so we may not realize that they are growing dissatisfied with their work conditions, it is important to regularly provide opportunities for feedback from everyone. This way, we can identify problems as soon as possible and then act quickly to fix them. Do not wait a year for the next scheduled appraisal to ask if team members are struggling with any area of their work - the feelings of being isolated and somehow lost can be so strong for some who try to work remotely for the first time that they will have left the company long before then.

This can take the form of regular social chats, or more formal mini-appraisals that leave time for the team member to collect their thoughts and present them either in a meeting or in writing, or, perhaps best of all, a combination of the two. In my experience, some people are reluctant to mention their problems during a quick social call. This may be because the negative emotions they have are not strong enough; the emotions are about how they feel, and not strictly speaking related to what they are working on; they may think the company cannot do anything to help anyway (which is usually not true); and/or because of the generally taboo nature of mental health issues. Overall, they may think that not feeling very well in a remote position is something they need to deal with themselves.

One will probably sense in various chats that there is a problem, but people can be difficult to “read” in a video call. This is probably because we get to know each other in a remote setting much slower, and we open up to others much slower as well. This is why mini-appraisals and the structured opportunities they provide to discover and discuss all problems are essential. Otherwise problems that are never important enough to mention in routine work calls, but can accumulate over time, can pass under the radar.

It is also important for the manager to provide feedback during these mini-appraisals. Again do not forget that the team member, because of the more limited communication, cannot see how others are doing, and so may be very uncertain as to whether their performance is adequate, whether there are any issues with what they are saying in chats and meetings or how they are saying it, and so on. Frequent positive feedback can reassure them, and frequent reminders, if needed, can help them to improve before any problem becomes too big to fix.

8. Building Respect for Time

Many worry that without constant supervision, people will simply stop working if they are at home. There’s just this neighbor to talk to, this article to read, this annoying level of a PlayStation game from last night to try again. This can be a worry, although arguably if someone does so, they are not the right person for the company anyway, either in the office or at home.

There may also be the temptation for team leaders or managers to drag on with work conversations or meetings a bit longer, well into the lunch break or the evening - after all, if we worked in the office, we would now be trying to get home, sitting on the underground or in a traffic jam, so we might as well use this time, right?

These are two sides of the same coin: lack of respect for time. As elsewhere, managers can lead by example: respect the lunch break and the end of the day. Running over a few minutes is fine, but anything regular or extremely long (without an emergency) is likely to cause resentment, which, to make matters worse, people may never mention fearing that they will sound unmotivated by their jobs. People may also have something scheduled not only for the evening, but for the lunch break: accept a delivery, pick up kids from the school, quickly move the spare bed in preparation for the in-laws coming at 5PM. It is true that none of this would be possible if they worked in an office, but the lunch break should still be their time that they can use however they wish.

Once it is clear for everyone that their time is respected, hopefully they will respect the company’s time in return, and actually do work during working hours. Admittedly, this may not happen, and someone may choose to take advantage of the reduced supervision and do other things. If their performance is measured with some tool suitable for their job, this can become very clear in extreme cases. However, in general, and in less extreme cases, this remains a question of trust - trusting that they respect the agreed hours. This respect can only be built if we offer respect first: to turn the above sentence on its head, if managers do not respect team members’ time, why would they feel it necessary to respect the company’s time?

9. Efficient Communication

And finally, as a continuation of the previous point, let me call attention to a horrible time waster: the way we ask questions online.

Let us imagine that George needed to get to the otherwise locked down office to print something out. At first sight, it may seem quite efficient for him to post a quick question to the team chat like:

[George] Does anyone know where the paper is for the printer in the office?

But what happens next? Both Alice and Bob, who work from home, remember where the paper is. Neither of them knows if the other one is free and is about to answer. Alice is in a video call and would prefer not to start typing, so she is keeping an eye on the question hoping that Bob would answer first. Bob desperately needs to go to the toilet and is hoping that Alice would deal with the question. Finally, Roger the office manager chimes in and says:

[Roger] But I filled it up with paper just before we left and no one has printed since!

[George] Yes I know there’s paper in there, but I thought it was the wrong size or something and wanted to try a fresh batch

Alice and Bob are both surprised. Alice excuses herself to deal with the problem, and reluctantly Bob sits back down to help figure out what is going on. There are now four people working on getting the printer working again.

[Alice] Why, is there an error message or smth?

[George] Yes it says out of paper

[Bob] Have you tried restarting the printer?

[George] Yes, but still says out of paper

Then Alice remembers something.

[Alice] Are you trying to print in color?

[George] Yes why

[Alice] It’s looking for photo paper in the upper tray. Just hit continue or change the default settings when you print your doc.

Problem solved, George can print, Alice can get back to the video call, and Bob can finally go to the toilet.

Why did this minor problem end up taking up around 20 minutes of four people’s time? What should have George done differently? He should have followed an entirely different approach to asking his question. Notice that his initial message was not related to the problem he encountered, but to what he concluded might be a solution. Jumping to the wrong conclusion is a phenomenon all too familiar to customer service departments, where one often needs to spend time asking the customer to take more and more steps back until the actual problem is revealed. Therefore, when asking an initial question on a chat, it is important to stick to the facts, and if one wants to venture a suggestion as to a possible cause or solution, do so separately. The pitfall of wrong conclusions can affect new team members more, who may not be aware of certain idiosyncrasies in the systems at the company, and may be keen to prove themselves to be capable problem solvers, and so invent solutions that duplicate existing solutions or may not address the underlying problem at all.

It is also crucial to include as much information as possible about the issue and what happened as otherwise others will need to engage in a lengthy conversation to pry out information they need. Enough information up front also makes it possible to quickly decide who the best person is to deal with the problem. That is, George should have mentioned all details that are likely relevant in his first message, for example, that there is actually paper in the printer, or maybe even that he is trying to print something in color. This would have jogged Alice’s memory immediately; she would have likely realized that she has seen this problem before, and would not have hesitated to respond. When phrasing a question it also helps to set the context in the first few words, so that other team members, who have never even seen the printer in question, can ignore George’s query without reading the rest.

These do mean that George is likely to spend a bit more time writing his initial message, but this time is more than saved by allowing Alice to quickly realize that this is a question for her, and reply with the actual solution in seconds. So, for example, George could have written:

[George] Hello, I’m having trouble printing in the office. These are the color charts I should be posting to Kate. The printer says it’s out of paper, but I can actually see paper in the bottom tray. Restarting didn’t help. Maybe the paper in the tray is the wrong size or someth and it doesn’t know it’s in there?

A longer message, but more efficient.

A Final Word

I hope the above list of issues specific to remote and home working will help to avoid the most serious problems as either individuals or whole teams transition from office-based work to remote working, or as more remote workers are hired. While the list is incomplete, it should provide a good starting point for the types of procedures, policies and practices that may need extra attention.

I would be thrilled to see remote and home working becoming more widespread. I think it would be beneficial to create a better work-life balance, to allow people to choose where they live and to spend more time with their families and children, to reduce their transport and childcare costs -- and save the environment by drastically reducing pollution and energy usage by overcrowded public transport systems and cars. I also strongly believe that it will prove cheaper to employ remote workers, where possible, for the businesses themselves, and that most problems and pitfalls arising from not being physically present can be solved by adjusting our management practices and finding new ways to organize work.