Artificial Intelligence (AI) and Machine Learning (ML) in particular are finding widespread applications in many industries. However, many customers across industry verticals, contemplating the adoption of ML in their business and IT process automation, have unreasonable expectations about the capabilities of these learning machines. Many imagine or opine that ML can completely take over human tasks. We will take a closer look at ML, explain the learning process and set the right expectations for these customers.
To begin with, computers are dumb machines out-of-the-box. Even the most advanced computer is only useful when programmed. The definition of ML, specifically “the ability to learn without being explicitly programmed” has opened up a lingering confusion about the basic premise about machine capabilities. Self-learning and autonomous machines have become misconstrued terms in the ML world. Data scientists have tried to mimic the human learning process on machines. Therefore, understanding the way humans learn is key to understand ML. Let us begin with a few examples that illustrate how humans learn by experience:
- Companies train employees to perform certain job duties. This training includes largely solving the common expected problems on the job. Trained employees face new problems and challenges which maybe beyond their understanding. In such scenarios, they consult the relevant subject matter experts (SMEs), who provide the solution. In future, when presented with a similar problem these employees would have learned how to handle it. This is a continuous assisted-learning (assisted by SMEs).
- The English language vocabulary is extensive and even the native speakers do not have complete knowledge and command of it. When faced with a new word, we consult a dictionary to understand its meaning. In effect, this new word is learned. Next time we come across the same word, we would know how to make sense of it. Our knowledge base (vocabulary) continues to expand over time as we learn new words (quality of data). This is a continuous self-learning.
- Students appearing for a test, train (study and practice) on a given syllabus. Depending on their level of training, questions from the syllabus are generally solvable. Any questions out of the syllabus are not solvable. Consulting an external source for solution is not an option here and results in a failure to answer them. This is a non-continuous learning.
ML draws a strong similarity with the human learning approach illustrated above. ML architectural components have a learner (machine program), training material (knowledge repository or training data), a method of training (ML model) and learning tools (associated software components to process and ingest training data). ML algorithms, supervised or unsupervised, solve classification or regression problems optimally. Specifically, they minimize a parameterized cost (MMSE) or maximize a likelihood (MLSE), which is a function of input data. The larger the training input data set, the better these trained parameters are. Input data can be digitized images, audio, structured text data (database rows/columns) or unstructured text data (natural language). The trained model is employed (as a program) on computers to predict new data. Training is typically offline, and model parameters are static on new predictions. However, the model maybe periodically retrained with a larger training data including newly observed data. In addition, new data maybe fed back into the model to continuously retrain the model and update parameters real-time. This approach behaves like a closed loop feedback control system.
ML accuracy improves with more training data, higher quality of the training data and the correct model selection (method of training) for the problem at hand. We should not confuse this ML accuracy with the general high precision computing accuracy of machines. Natural language processing (NLP) of unstructured text data is an important use case in many industry applications. However, state-of-the-art NLP technology does not completely solve the knowledge mining and user intent extraction from unstructured text data. ML algorithms ingesting these processed data, as a result, do not produce high ML accuracy and often require manual intervention for continuous assisted-learning to improve it.
As a concluding remark, ML is simply a set of cognitive services (sophisticated computer programs) running on computers which to a certain extent can mimic human like learning behavior from different kinds of digital data such as audio, image and text and automate tasks done by humans which could not be addressed by the traditional rule based automation. Machines have much higher computing power, several order of magnitude faster than what humans can do. This time saving is a key tangible benefit. However, ML accuracy is still limited in many scenarios, especially when handling unstructured text data. Advances in ML and NLP technology will only improve this accuracy but a complete overhaul of humans by machines is still a long shot. Nevertheless, achieving a moderate, say 30% accuracy, in this cognitive automation can still add significant value to businesses, in terms of cutting down costs in human capital investment and improving overall productivity for the business unit.