Machine Learning: Similarities With Human Decision Making
Machine Learning as we know have evolved immensely in the last few decades. Thanks to the availability of huge computation power today at a fraction of the cost a decade ago. Today, ML applications can be seen everywhere, from predicting whether patterns, analyzing sentiment of user reviews, anomaly and fraud detection, self driving cars to helping farmers increase crop yield.
I would like to start with the broad definition of machine learning as put forward by Tom M. Mitchell “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks T, as measured by P, improve with E”.
While doing it a Machine Learning or AI algorithm deploys many complex mathematical and statistical concepts in its quest for a better performance. However, broadly, the fundamental principles that a ML algorithm follows is something that we as a human being applies everyday in our decision making process. Problems in ML are solved by training a model (also referred to as an algorithm) on historical data (we can refer it to the term E(Experience) as mentioned in the definition ). During training the models performs the following :
- Predicts a value for the target variable.
- Compares it with the actual value in the training set and calculate the error. In other words, calculates the cost of miscalculation.
- Adjusts the parameters and recalculates the value and see if the error is minimized.
- If the error is reduced the model continues to do that until the error stops decreasing. At this point we have the best possible prediction.
The above mentioned steps if expressed in mathematical notations will look like some complex equations only comprehensible by someone with good knowledge of algebra, calculus, probability and statistics. However, As I stated in the beginning, the same methods are being used by all of us in our day to day life without bothering about any complex mathematical equation.
Schematic comparison between human decision process and a ML algorithm:
The human basic decision making steps:
ML prediction making steps:
Lets dig a little further:
Lets consider a scenario, you are trying to travel from point A to point B in the City for a meeting. You have set a target arrival time. When planning our travel we are trying to minimize the time consumed (or expense whatever is the target value),In other words trying to maximize the probability of reaching on time. While doing that we consider some attributes, such as
- Mode of transport,
- Time of the day
- Traffic condition
- Weather condition etc
While trying to quantify the effect of each of these attributes, we unknowingly apply some mathematical principles to figure out,what route we should take, given these conditions . For instance, If the traffic condition is worse, we may consider taking a metro train as driving may lead to increased error ( Error: time consumed beyond the set target). These attributes that effect our decision making are called features and the target arrival time is called the target variable in machine learning. The model tries to find an optimum solution (or value) by learning the effects these attributes.
Our decisions are based on our past experience and the current conditions. We give more weightage to the conditions which influenced our past experience the most .Similarly, in ML from the given past dataset we first, need to find out the important features and visualize how these features actually influenced the target variable in the past. Finding these features from the historical data is called feature engineering and visualizing them to find any trend or hidden relationship is called Exploratory data analysis.
Lets consider this diagram. We are traveling from S(start) to F(Finish) and by adding the numbers in between start and finish, we can calculate the cost of a particular path.
We would want to select the path with the minimum cost.
Things don’t go as planned:
Lets say, we finalized a route thinking that we will reach the destination in N minutes say 45 minutes, however, actually took some 55 minutes. In this case (55–45) 10 minutes is our error. Next time while predicting the expected arrival time we will use this information to plan better.
In the exact same way, our ML model calculates the cost (we referred to it as error) after every iteration of training and adjusts the weightage of the features to minimize the cost in the next iteration. This process continues until we are certain that the cost is no longer decreasing. At this stage we preserve the values of the attributes and use them to make predictions on the real-life production data. The equation for calculating the cost is called the cost function and one of the primary methods of adjusting the weightage of the features is called Gradient Descent. We can visualize this process by plotting the value of the cost function against the parameter(or feature) values.
Lets plot the cost function against the parameter values.
In the above diagrams the cost function is at the most influenced by two features, while in real life the number of features can be in hundreds.
Irrespective of the size of the dataset & features the intuition and fundamental principles behind a ML algorithm have striking similarities to the way we human, take our day to day decisions. It’s just that a ML algorithm expresses the decision making method in a mathematical and programmatic way.
Concluding Thoughts:
The above similarities are significant and it can be said that machine learning algorithms are actually imitation of human decision making process.However, this is primarily confined to the logical decision making process. Emotions play the most important role in human thought process. If, we have a ML model to predict, will “Kinjal” play outside today and it is raining today, the model will predict “Kinjal will not play outside today” while in reality Kinjal may feel like dancing in rain.The model will probably say “How Emotional humans are!Its difficult to understand them”. So, it will not be wrong to say that while the emerging technologies like AI,ML will have big impact the way we live our life but end of the day our human sensitivities will be the driving force behind them.
*********************Thank You***********************
