10. How to learn from your mistakes
Deep learning, feedback loops and human memory
With ADHD, you’re always forgetting what you’re meant to be doing. My working memory – the part where we hold information for short-term, immediate use – is constantly undermined by new thoughts, impulses or emotional responses. It feels as though everywhere you go, even just to the room next door, the working memory is always being refreshed, losing immediate context. As well as making it almost impossible to hold a grudge, it also means I often leave the house to then forget where I’m meant to be going, or why, or to only realize at work that I’ve left my keys in my gym bag at home for the third time this month. I might get home and forget to take my jacket off for hours, because I suddenly got engrossed in a book I had picked up, or decided to erect some flat-pack furniture there and then. Or in certain cases I am focusing so intently on one difficult thing, such as planning my commute or working on a project, that I completely forget about anything else, like eating meals. My thoughts are like flies buzzing around the brain, loud but disparate, as opposed to the firmly anchored tent pegs of an organized mind.
Because short-term memory is a challenge for me, I’ve thought a lot about how the brain works to process and store memories. I’ve experimented on myself to see if I can improve the functioning of my short-term memory. And, as my understanding of machine learning has developed, I’ve begun to see how the artificial intelligence systems scientists are developing can help us to think afresh about our very human struggle with remembering.
This is important because memory isn’t just about making sure you leave for work on time, with your keys, and wearing pants. Memory also represents the building blocks of who we are as people: the instincts, experiences and life events that have created the human being we are today, and will become in the future. Without understanding memory, we can’t comprehend our thought processes, our psychology, our responses to people and situations, or what we value. In fact, we can’t understand or fully know ourselves at all.
Conversely, a better understanding of how our memory works – what gets amplified and suppressed, what is near the surface or practically hidden from recall – can help us to achieve a more focused and supportive attitude to life. It allows us to escape the shackles of the bad memories that limit us, and focus on those from which we can learn or draw strength (cheesy, but true). Memory is something that can crush us if we let it: the accumulation of things we’ve done, said or thought that make us anxious or ashamed. These bad memories aren’t just painful, they can actively prevent us from moving forward in life, such as the retrospective embarrassment of that time I got ridiculed for wearing blue eyeliner for a Tuesday lunch (a choice made out of pure boredom).
Much like energy, memories cannot be destroyed, only transformed (though unlike energy they can be created – as they are being in every living moment). Memory can take us back to the people and places that formed us, providing comfort and nourishment in difficult times – grounding us for our next venture.
Memory is intrinsic to all of us, and it’s something we can take more conscious ownership over. It can be trained like a muscle – not necessarily to be stronger, but to better serve our needs, prioritizing the helpful over the harmful. We can become happier, more focused and more purposeful by developing greater awareness about how our memory works and aligning its capacity to our priorities. I learned how to do this by studying the workings of the closest scientific equivalent to the human brain – the artificial neural network. Just as these networks can be programmed to optimize how they process information to achieve certain outcomes, so we can fine-tune our brains to make more effective use of the oceans of data that our lives create.
If you’ve ever wondered how to escape the shadow of a bad experience in your life, to avoid the memories of what has gone before from limiting your future potential, then this chapter is for you. I want to show how, by applying techniques from deep learning and harnessing the power of feedback, we can leverage the power of human memory in our favour – learning from our mistakes without being constrained by our pasts. (Like that purple tank top I used to wear compulsively when I was eight.)
Memories might be made in the past, but their most important role is to inform decisions in the present and future. What we choose to remember is crucial to determining how we react to all kinds of situations in our lives. With the right, artificial intelligence-inspired tweaks, we can turn memory from a potential millstone into our most important source of power.
Deep learning and neural networks
Neural networks provide an ideal parallel to human memory for several reasons. The first, most obviously, is that they have been modelled on the brain: designed to produce the closest proxy to human intuition, perception and thought processes that artificial intelligence is currently capable of. The second is that their function is dependent on a feedback system that is crucial to understanding our own ability to retain and learn from particular memories. It is this feedback loop, and its implications for how we programme our own memories, that I want to focus on.
But let’s start at the beginning. What is a neural network and what can it teach us about ourselves? Neural networks are algorithms that are programmed to turn inputs – senses and perceptions – into outputs – decisions and judgements. They are the principal tool of deep learning, a subset of machine learning that addresses complex problems which require the machine to ‘think’, working iteratively based on the data that is inputted. In other words, the algorithm uses the information or data provided to improve its understanding of a certain problem – which might be analysing traffic flows around a city, trying to work out how much house prices might rise based on historical information, or detecting someone’s mood based on the look on their face. In all these cases, the ability to model good answers improves the more data you input into the system, and the more reference points the algorithm has to work with. Compared to traditional machine learning, a neural network is more independent and requires less input from the programmer to define what it should be searching for, since, through internal layers of logic, it is able to create its own connections.
All of the more radical examples of artificial intelligence you may have read about – from fully driverless cars to mass automation of people’s jobs – ultimately rely on deep learning, the closest we have so far got to developing a computer program that can think (within considerable limitations) like a human. Deep learning is also responsible for applications, including criminal checks, drug design, and the computer programs that rival the most competent chess players, all of which depend on an ability to simulate the connective capability of the human mind.
Modelled on the brain, a neural network is made up of neurons – in this case, the various data inputs. These come in three layers: input, output and in the middle what is called ‘hidden’, the place where the algorithm does its thinking. For example, if we take a driverless car, inputs will include the angle of the road, the speed of the car, proximity to other vehicles, the weight of passengers, and any obstacles on the road – all factors that determine the nature of the outputs, which are the decisions the algorithm makes about how to drive safely. It is the connections between these neurons, and how they fire, that really matter. The crucial aspect to a neural network is that the connections have a virtual ‘weight’ assigned to them, affecting their influence on the network and output. It’s by comparing and calculating the weights of these inputs that the program reaches its decisions, learning which inputs to trust as most indicative of a particular result. With our driverless car example, it’s likely to be speed and proximity of obstacles (which could be pedestrians or other vehicles) that have the heaviest weighting and the greatest influence on decisions. The ultimate aim of a neural network is, over time and extensive trial and error, to assign the most accurate values to these weighted connections, so it can consider new inputs with the priority (high or low) that they deserve.
So, rather than being able to tell the difference between a car and a pedestrian by feature extraction – isolating and locating wheels, legs, arms or wing mirrors – as a straightforward machine-learning program would, a neural network is able to use its weighted connections to simply detect which is which, and most importantly the combinations of data points that depict it most accurately (i.e., if it’s got legs and arms, it probably isn’t a Honda Civic). And the more images of cars and people you feed into it, the better opportunity it has, through trial and error, to optimize its weightings and combinations, and maximize the accuracy of its outputs (decisions). Much as we have different layers of memory piled up throughout our life, deepening our ability to establish connections and inform decisions, a neural network becomes more complex and sophisticated the more memories (data) it has processed. Like a child learning things for the first time, the more opportunity it has to exercise its ‘mind’, the better informed and evolved it will be.
This is thanks to its second crucial component: the feedback system. By comparing predicted and actual results, the network can calculate its estimated error, and then use our old friend gradient descent (turn to this page for a reminder) to determine which of the weighted connections are most in error, and how they should be adjusted: a process called backpropagation (aka self-reflection). In other words, the neural network does something that humans are often bad at: it learns from its mistakes. In fact, it is hardwired to do so, without the emotional baggage that humans attach to their mistakes, using feedback as an intrinsic component of its quest to improve.
By contrast, humans often need reminding that feedback is important, and any engagement with it can be reluctant. For many of us, feedback is a dirty word. In its most prevalent context, the workplace, it often acts as a neutral way to characterize a negative experience: being told that, for whatever reason, our work isn’t good enough. It carries all the connotations of the awkward conversation, shuffled feet and words that don’t quite mean what they say. But that’s only because humans are, too often, bad at both giving and taking feedback. The neural network reminds us of its vital importance. Only by comparing what we expected with what actually happened, and adjusting our assumptions or approach as a result, can we ever get better at anything. If we rely on the same old weighted connections throughout our lives or careers, we will never change or evolve, or know why we are getting bored and frustrated by doing the same things in the same way.
When it comes to memory, we can all learn from the feedback-centric approach of the neural network. Or to be more precise, we can benefit from being conscious about this process, because it is one that happens already. The brain is busy weighting the information we process every single moment, deciding what we need to remember and whether it is for immediate, short-term retention, or something that we will always need to know. The things we remember are those we do or think about often (thanks to repetition), ones that are important (because then we actually stop and focus attention on them) or events and moments that have had a particular impact on us (also related to attention). Things that fall into these boxes get remembered – which isn’t just to say that we store a memory of them, but that they themselves become part of the brain’s algorithm, affecting our biases (weighted connections) and tilting the lens through which we process new information. What the brain considers to be important one day is going to continue to condition its priorities on the next. And vice versa. These connections and associations of memory are a tinted glass through which we view our entire lives.
To be unconscious about this process – the consistent ranking of everything we encounter into memorable and not memorable – is a bit like outsourcing your dating apps, using your previous preferences to determine automatically who to swipe right on, without actually checking whether you fancy them. It also leaves us open to errors in the system. In neural networks, errors can creep in if weightings have been based on too narrow a data set – also known as overfitting your model – or one that may suggest causation where only correlation exists: false alarms. So if you’ve trained a network to tell the difference between cats and dogs based only on paw size, it might be fooled by a really large cat or a small dog.
Our own brains are no less fallible. They may prioritize things we don’t want or need to remember, or fail to log those which we do. We need the feedback loop to turn these ‘errors’ into data we can glean insight from, and use to make adjustments. As any scientist will tell you, there is no such thing as an error or a bad result, only sources of further learning. So if we want to re-programme our memory to better effect, we need to be more aware of the feedback loop that produces our core weightings, and start thinking about what we can do to optimize it. Without proper feedback, we are using only a fraction of our memory’s capability to change how we see our life and the world around us.
Re-engineering the feedback loop
So what makes this feedback loop? How can we turn the light and darkness of our pasts into a memory that functions as friend and not enemy? We know that the feedback loop works: it’s what has embedded past romantic haunts, that ugly cardigan we wore one time too many, and the most embarrassing events of our lives (you try letting your boyfriend use your iPad when you’ve forgotten to close a Google search for ‘What are the pros and cons of getting engaged?’), into our consciousness. It’s what makes us anticipate that first taste of coffee in the morning. But how do we make it work for us?
It starts with separating the mass of accumulated data from our lives into something useful. As humans, over time layers of memory have congealed in us, making it hard to separate different time frames and determine what is truly important here and now. Old bugs of the past can crawl their way into the present, clouding your judgement and ability to see clearly. Computers have the same problem – their memory getting clogged with too many programs running all at once. And they have a solution: to debug, getting rid of what is no longer useful or necessary.
To debug is hard for any human, but especially so for someone with ASD. While my working memory may be patchy, my recall for detail is the opposite – so effective that it gets in the way, as I become distracted from my present by remembering that a commuter on my train last month had a certain resemblance to an avocado. Being Aspergic means having the eyes of a hawk and the ears and nose of a bloodhound, neither of which is especially helpful at trying to be human.
Because we notice everything and store data from every detail of a situation, our memory logs soon start filling up. Letting go of this obsession with data collection is hard. Detailed memories are part of who we are: reaffirming our existence, and our connection to the people and places in our lives. And when you have a mind that is sensitive to every event and stimulus – not just when a car horn or ambulance siren is blaring, but when it isn’t and you’re waiting for the next one – it’s not an option just to flick the off switch.
This is also a part of me I would never want to lose. The obsessiveness that manifests itself in endless preparation and routine is also the sentience that allows you to see the world in a different way: noticing beauty and difference where other people would never even take a second look. My ability to observe makes me open and alive, closer to my animal spirit than technological modernity often wants to allow.
But it’s also challenging, because when you log every noise as a signal, it can be impossible to do as the neural network would, and establish a hierarchy of weighted connections. (It also makes for somewhat high-maintenance shopping trips: sorry, Mum.)
And, like everyone else, I have always wanted to fit in. I may have felt like I landed on the wrong planet, but that doesn’t mean I want to live as an alien among natives. From growing up in Wales, to school in the Cotswolds, university in Bristol and jobs in London, I have worked diligently to try to swim in the mainstream. And one thing that I could never get out of my mind was the distinctively British reserve: the way people often talk and behave reticently, not quite saying what they’re thinking, or trying to ignore something outrageous in the hope that it will go away.
I am not a reticent person. I am an enthusiastic giggler, a squealer in delight and a nostril-flaring howler in rage. My emotions have very rarely been a closed book to anyone. But I wanted to experiment. I wanted to become more reserved, to be more neutral and to enjoy the benefits of this seemingly more objective stance on life, which seemed to be the hypothetical optimum. Becoming less Millie and more English seemed like the opportunity to kill two birds with one scientifically aimed stone: a means to fit in, and a method to prune the tendrils of my overgrown memory bank. I wanted to be more like Siri or Alexa: all the knowledge, none of the emotional baggage. And people actually listen to them.
So I set out on an experiment. I would re-engineer my neural feedback loop, blocking off the impulses that make me the ‘emotional weirdo’ I am, and gaining the calm perspective of a completely neutral, very British being. This wasn’t just shutting down my mental computer, but restoring it to factory settings: unlearning all of the associations that tangled around in my head, often constricting the clear thinking and the connection that I craved. In my mind, I could turn myself from an ADHD-wired, emotionally driven tempest into a gentle breeze of logical thinking and measured behaviour, who wouldn’t forget her keys every other day, or have her views discounted for being too emotional. Through doing this, I would be able to disregard the unhelpful accumulation of old memories, and make logical judgements based only on the new inputs at that point in time. I would erase my biases, reset my neural weightings and start completely from scratch. A holiday in my own head.
But in trying to make myself remember better, I ended up forgetting things that were far more important. Over the course of this experiment, on a first date, a boy asked me what I was passionate about. And I realized I had absolutely nothing to say. I had made such a conscious effort to erase my biases and preferences, to get rid of those messy vulnerabilities, that I no longer knew what I cared about. I felt as though my spirit had fossilized, and I hadn’t so much forgotten a few things as completely lost myself in the mental mist. I felt instantly, crushingly sad. And then scared. What had I done to myself? Ironically, by this stage, I couldn’t even remember why I had set out on this course in the first place, since I forgot to make a note on my whiteboard. (Great, nice one, Millie. Once again, full marks for OCD consistency.)
As many experiments must be, this one was an almost total failure. It was a dangerous brush with trying to erase and deny my natural biases and true self. But, as failed experiments also tend to do, it taught me some important things. First, we have one spirit and personality, which is totally our own and should never be a source of shame or regret. We have to nurture that person, not deny or reject them. But, at the same time, we are not their hostage. I have learned to love myself as the unashamedly ASD, ADHD, GAD person I am: the full Millie. It has been the work of my life to balance out these competing parts of me, and capitalize on them where they are most useful. It’s a full-time job in itself, a science and an art.
But that doesn’t mean I don’t find a lot of her behaviours a complete pain in the arse. The forgetfulness. The fear. The struggle to cope with big feelings. You can love a person while simultaneously hating being them. But, even better, you can chip away at the behaviour you find problematic. I forget things because my attention is constantly being pulled this way and that, eroding my instantaneous recall. I am afraid of smoke and loud noises because, deep in my neural network, the connections are irreversibly weighted by twenty-six years of having responded to these things in those ways. So my mental computer spits out a response to be afraid, and run away. These are responses conditioned by memory – its accumulation and its inconsistency. Which means they are problems that can also be addressed by training the memory and targeting the feedback loop.
I can’t magically undo my forgetfulness or wipe away my fears. But I can find ways of managing them better, preparing myself for situations I know I will find difficult, and re-plumbing those neural connections to counterbalance the existing weights. It’s a painful but rewarding and enlightening process: the human luxury of being able to fine-tune our own mental computers.
Some of these adjustments are very practical. Although my room might look disordered, it is actually full of clues to guide me through the day – starting with my dressing gown and toothbrush sitting by the right side of my bed, to remind me to get up, go to the bathroom and brush my teeth first thing in the morning. Others will strike you as plain odd. To remind myself to take any medications, I have to make an event out of it – shouting, ‘Hagrid!’ and doing a dance to myself. This routine might sound deranged, but at least it’s memorable – adding weight to the likelihood that I will remember to do something really important, but too easily forgotten. And it’s supported by a litany of Post-it Notes reminding me to pick up my socks, call my mum (twice) and not to wash the jeans that have £5 in the pocket.
Remembering to remember things is largely a question of finding the right mechanisms to remind yourself. Forgetting to be afraid is more complex. But this is about the feedback loop and backpropagation as well. Because I know that smoke or bad smells won’t actually do me any harm, I can use that proven outcome to counterbalance the weighted connection that tells me to be afraid. I can try to update the inputs that condition how I respond to particular situations, by reassuring myself about a track record of outputs. This is never going to magically turn a negative feeling into a positive one, but it can reduce the intensity of the feeling, shift the dial slightly on that connection and enable me to retreat more often than not from the precipice of a panic attack.
You probably have quirks and kinks in your own memory and feedback loop: the past experiences that loom larger than they should (like a bad break-up), or the positive affirmations that we can over-interpret (just because you ultimately lived to tell the tale doesn’t mean that last drink was a good idea). The important thing is to take some conscious ownership over a process that otherwise hums away unconsciously, robbing us of complete ownership of how we think about life situations and make decisions. If you are struggling to commit to a relationship because your last one was difficult, you need to remember that your previous one does not define you; the weighting of that may be too heavy in your feedback loop, inhibiting your ability to judge the new relationship on its merits. We need to think about why we feel a certain way, whether that is uncertain or over-confident, and try to locate the root of that emotion in the previous experiences that have filled our memory banks and conditioned our feedback loops. Once we have done that, it becomes easier to put both good and bad memories in their proper context and adjust the weightings accordingly: learning from our mistakes, getting over our hang-ups and looking forward to the future with something as close to objectivity as a human can ever achieve.
If we want to change how we feel about things, or approach particular situations in life, then the feedback loop is the place to start. We should recognize that our instinctive responses have been conditioned by a lifetime of memory and experience, creating the weighted connections that determine how our brain makes calculations. The things we value in life and feel strongly about haven’t emerged by accident. They are rooted in our living memories, and the only way to change is by gradual adjustment via the feedback loop.
There are two kinds of feedback loop, positive and negative, and both have important roles in training systems. A positive feedback loop is an incitement to do more of something, giving it greater weighting and prominence in the overall calculation. These are for when we want to encourage a system (aka ourselves) to be bolder about something. Its negative counterpart is designed for the opposite effect, to restrict or limit a certain factor. Both have their benefits, and drawbacks. A positive feedback loop is stimulating, but can allow the inspiration and joy of living to spiral out of control, especially when it comes to drugs and alcohol, as we seek to revisit the same high our memory associates with them. Whereas a negative one, while acting as a stabilizing force, can also leave you tunnelling into a rut of introspection and futility: my experiences of depression have been when bad memories and experiences suppressed my positive energy to such an extent that I felt totally futile and unable to function for days at a time. This is the ultimate manifestation of a negative loop – where you experience an effective eclipse of every good memory and feeling that you have ever experienced.
If we are trying to create a positive feedback loop, then small doses of experiencing what we fear can build confidence, chipping away at the weight that tells us to hang back and be afraid. And we really can do things that make us afraid. I even once forced myself to go with my friends to a music festival (allegedly ‘the best thing in the world’), which is basically the Asperger’s equivalent of Mordor: excessive noise, endless mess, dubious smells and unpredictable crowds. It was there that I broke my personal record, of having five fully blown panic attacks in thirteen hours, not to mention the one when I accidentally got caught at the front, sardined in a mosh pit. After fainting from shock, I was crowd-surfed over to the medical tent for attention where they rang my parents. I then had to be rescued by my dad – who laughingly reminded me that us Pangs have never been happy campers.
I have used that experience to draw comfort from my ability to experiment, test my boundaries and remind myself that the unfamiliar (and even the unpleasant) isn’t necessarily going to be fatal. You’ll never catch me going back to Beach Break festival, or probably ever sleeping in a tent again, but I don’t regret my abbreviated, abridged adventure for a second. I lived it up, however briefly, and I’d be willing to try something new like that again.
At other times we want to create a negative feedback loop, to stop ourselves from doing something. To achieve this, we might focus harder on the problematic outcome of a certain behaviour, reminding the brain of the asymmetry between the reason we do a certain thing and the end point that it invariably leads us to – be that a hangover, a sugar headache or being sick because you pushed yourself too hard at the gym.
These positive and negative feedback loops are constantly pinging across our brains, whether we pay any heed to them or not. My experience is that, the more conscious I am of their existence, and the harder I work to re-engineer them (reminding myself of those good or bad outputs, contra to expectations), the more in control of my state of mind I become. It’s also worth remembering that a well-functioning system, whether human or algorithmic, relies on the right balance of positive and negative feedback. We need enough positive to allow us to experience new things and learn, and sufficient negative to limit us from making silly decisions or putting ourselves in danger. We can’t overdose on either positive or negative feedback if we want to maintain a form of equilibrium; just as the neural network powering a driverless car mustn’t interpret information either too aggressively or too cautiously to be a safe driver. But neither can we do without them – both need to be deployed, and tweaked, to match the different situations we will encounter.
Through my experiment I learned that it’s not possible to simply abandon a lifetime’s accumulation of memory and mental preconditioning. Like it or not, these are the things that make us human, allowing us to feel and giving us an identity or personality that provides our anchor. These biases might sometimes feel like an enemy, but in reality they are just us – the purest expression of ourselves. But accepting the existence of these biases is not the same thing as surrendering to them. We can remain in control by being cognizant of them, working gradually to use real experiences to prime the feedback loop, and adjust those all-important weightings of experience. We have to bring our subconscious biases into the realm of the conscious so we know what we are dealing with. Like going through old photos, this is a process that can be both scary and hilarious.
Experimenting with my memory has taught me that wiping it entirely clean – and having to re-learn how to hold my favourite mug – is not the way to go. We can learn from neural networks, but we’re not like computers that can make their memory work more effectively by getting rid of everything they have accumulated. In the place of a total memory format, I have settled on a continuous process of mental upgrading. Every one or two years I will look at the different layers of memory that are foremost in my mind – putting aside those which used to be useful but have now served their purpose, and trying to knit together those that give me inspiration, focus and happiness. It’s a way to have fewer regrets about what is past, and to sharpen your mind for the challenges that lie ahead. At any one time, our memory represents the tapestry of our life. We shouldn’t forget that we get to choose what it features.
We can’t control everything that happens in our lives, but we can condition how the memory both stores and uses those experiences. Something entirely in our control is what we give weight to, how we remember it and for what reasons. What things in your life give you strength and remind you of the person you really are and what you are capable of? Conversely, what low points can act as speed bumps on future behaviour and decisions, reminding us of what we are likely to regret later? We get to choose and prioritize: the good versus the bad, the logical versus the emotional, our feelings versus those of others. All these elements are fizzing across the feedback loop on a constant basis: like any recipe, it is the proportions that matter, and we choose those by deciding what to really focus on, and how to process and store those memories. While learning from your mistakes might sound trite and simplistic, it is actually an important part of how we condition the mind and memory to work in our favour, creating a healthy, balanced feedback loop that better equips us to tackle future challenges – and not being afraid to fine-tune it over time.
The feedback loop happens instinctively, but there is a great power in actually owning it: making ourselves think about how it works and what adjustments we can make. Focus and attention are such important parts of how we form memory. Because my dad’s lasagne is my favourite food in the whole world, when I was young I would always let it sit in front of me, taking in the smell and sight of it for ten seconds without blinking, logging the memory as one to come back to whenever I needed to feel the comfort of home. In small ways like this, we can all train our memory to prioritize the useful over the unhelpful, making the most of its capacity to provide strength, reassurance and comfort.
Memory can be a place of anxiety, shame and regret. Left to grow and evolve unconsciously, it encourages us to dwell often negatively on our experiences and past decisions, reliving them in the seconds, months and even years afterwards. Our challenge with memory is not getting stuck in its maze of history, regrets or people and places we can never return to. If we’re being honest, most of us probably live more in the negative feedback loop than the positive: selectively accumulating the bad experiences and memories that chip away at our confidence and influence the trajectory of future judgements.
But memory is also something we can’t live without. Something which, as I discovered, is too intrinsic an element of our humanity simply to strip out like a defective part in an engine. To format our memory, as we would that of a computer, is to wipe away too much that can never be properly replaced. So our best option is to fine-tune, making the adjustments over time that allow us to get the most out of this powerful, sometimes dangerous, source of ourselves.