Just a recap, in my previous post, I had primarily focused on the tools/techniques/frameworks and hardware that are prominent in the implementation of an AI/ML framework for any organization. In this post, we are going to dwell a little deeper into the Algorithms that are at the core of the ML/DL space and how humans are at the helm of this impact of machines.
Algorithms are becoming an integral part of our daily lives. About 80% of the viewing hours on Netflix and 35% of the retail on Amazon are due to automated recommendations by the so called AI/ML engines. Designers in companies like Facebook know how to make use of notifications and gamifications to increase user engagement and exploit & amplify various human vulnerabilities such as social approval and instant gratification. In short, we are nudged and sometimes even tricked into making choices by algorithms that need to learn. In case of the products and services we buy, the information we consume and who we mingle with online, algorithms are playing an important role in practically every aspect of it.
In the world of AI, a new challenge that is being increasingly discussed is the biases that creep into algorithms. Due to these biases, when we leave it to algorithms to take decisions, there can be unintended consequences. More so, as algorithms deployed by tech companies are used by billions of people, its damage because of biases can be significant. Moreover, we have a tendency to believe that algorithms are predictable and rational. So we tend to overlook many of their side effects.
How today’s algorithms differ?
In the past, developing an algorithm involved writing a series of steps that a machine could implement repeatedly without getting tired or making a mistake. In comparison, today’s algorithms, based on machine learning, do not follow a programmed sequence of instructions but ingest data and figure out for themselves the most logical sequence of steps and then keep working on improvement as they consume more and more data.
Machine learning itself is more sophisticated as traditional (supervised) ML, a programmer usually specifies what patterns to look for. The performance of these methods improves as they are exposed to greater data but this is limited. In deep learning, programmers do not specify what patterns to look for but the algorithm evaluates the training data in different ways to identify patterns that truly matter. Human beings may not identify some of these patterns.
Deep learning models contain an input layer of data and an output layer of the desired prediction and the multiple hidden layers in between that combine patterns from previous layers to identify abstract and complex patterns in the data. For instance
Unlike traditional algorithms, the performance of deep learning algorithms keeps improving as more data is fed.
Decision making and avoiding unintended consequence
AI involves enabling computers to do the tasks that human beings can handle. This means computers must be able to reason, understand language, navigate the visual world and manipulate objects. Machine learning enhances this by learning from experience. As algorithms become more and more sophisticated and develop newer capabilities, they are going beyond their original role of decision support to decision making. The flip side is that as algorithms become more powerful, there are growing concerns about their opaqueness and unknown biases. The benefits of algorithms seem to far outweigh the small chance of an algorithm going rogue now and then. It is important to recognize that while algorithms do an exceptionally good job of achieving what they are designed to achieve, they are not completely predictable. They do have side effects like some medicines. These consequences are of three types
Perverse results affect precisely what is measured and have a better chance of being detected. Unexpected drawbacks do not affect the exact performance metrics that are being tracked. and difficult to avoid them. Facebook’s Trending Topics algorithm is a good example. The algorithm ensured that the highlighted stories were genuinely popular. But it failed to question the credibility of sources and inadvertently promoted fake news. So inaccurate and fabricated stories were widely circulated in the months leading up to the 2016 US Presidential elections. The top 20 false stories in that period received greater engagement on Facebook than the top 20 legitimate ones.
Content and Collaborative filtering systems
Content based recommendation systems start with detailed information about a product’s characteristics and then search for other products with similar qualities. Thus, content based algorithms can match people based on similarities in demographic attributes- age, occupation, location, shared interests and ideas discussed on social media.
Collaborative filtering recommendation algorithms do not focus on the product’s characteristics. These algorithms look for people who use the same products that we do. For example, two of us may not be connected on Linked In but if we have more than a hundred mutual connections, we will get a notification that we should perhaps get connected.
Algorithms also leverage the principle of digital neighborhood. One of the earliest pioneers of this principle was Google. In the late 1990s, when the internet was about to take off, the most popular online search engines relied primarily on the text content within web pages to determine their relevance. If a lot of other sites have a link to our website, then our website must be worth reading. It is not what we know but how many people know us that gets our website higher rankings. Research reveals that when Oprah Winfrey recommends a book, sales increase significantly but the books recommended by Amazon also get a significant boost. That is why digital neighborhoods are so important. For products that are recommended on many other product pages, recommendation algorithms drive a dramatic increase in sales. Spotify initially used a collaborative filter but later combined it with a content based method.
AI began with Expert systems, ie systems which capture the knowledge and experience of experts. These systems suffer from two drawbacks.
- Do not automate the decision making process.
- Can’t code a response to every kind of situation.
We can either create intelligent algorithms in highly controlled environments, expert systems style to ensure they are highly predictable in behavior. But these algorithms will encounter problems they were not prepared for. Alternatively, we can use machine learning algorithms to create resilient but also somewhat unpredictable algorithms. This is the predictability – resilience paradox. Much as we may desire fully explainable and interpretable algorithms, the balance between predictability and resilience inevitably seems to be tilting in the latter direction.
Technology is most useful when it helps human beings to solve the most sophisticated problems which involves creativity. To solve such problems, we will have to move away from predictable systems. One solution to resolve the predictability resilience paradox is to use multiple approaches. Thus, in a self-driving car, machine earning might drive the show but in case of confusion about a road sign, a set of rules might tick in.
Environmental Factors that Support
Human behavior is shaped both by hereditary and environmental factors. Same is the case with algorithms. There are three components we need to consider:
While data, algorithms and people play a significant role in determining the outcomes of the system, the sum is greater than the parts. The complex interactions among the various components have a big impact.
Many professions will be redefined if algorithmic systems are adopted intelligently by users. But if there are public failures, we cannot take user adoption for granted. We might successfully build intelligent diagnostic systems and effective driver-less cars but in the absence of trust, doctors and passengers will be unwilling to use them. So it is important to increase trust in algorithms. Otherwise, they will not gain acceptance. According to some estimates, driver-less cars would save up to 1.5 million lives just in the US and close to 50 million lives globally in the next 50 years. Yet, in a poll conducted in April 2018, 50% of the respondents said they considered autonomous cars less safe than cars driven by human beings.
Rules and Regulations
Decision making is the most human of our abilities. Today’s algorithms are advancing rapidly into this territory. So we must develop a set of rights, responsibilities and regulations to manage and indeed thrive in this world of technological innovations. Such a bill of rights should have four components:
AI systems have mastered games – now it’s time to master reality! Sometime back, Facebook developed two bots and trained them on negotiation skills. The bots were exposed to thousands of negotiation games and taught how conversations would evolve in a negotiation and how they should respond. The outcome of this training far exceeded expectations. The bots learnt how to trade items, developed their own short hand language. When the bots were made to negotiate with human beings again, the people on the other side did not even realize this!
Stay tuned…. Part 5 of this foray, we will look into the details of top 8 examples of organizations that have successfully implemented the AI/ML framework and how they are benefiting out of it.
Please feel free to review my earlier series of posts on AI-ML Past, Present and Future – distributed across 8 blogs.