A human baby is a voracious learning machine. With each passing moment after birth, a baby’s brain assimilates and learns from data through all its senses.
The sense organs receive a continuous stream of stimuli that are stored and processed by the brain. In infants, the brain grows rapidly in size while accumulating vast amounts of information. Research shows that the infant’s brain reaches half its adult size in the first 90 days post-birth.
As infants’ brain gathers data, they are also learning from it continuously. This experiential learning adds to the cognitive abilities of the brain so that the infant responds appropriately to similar stimuli in the future.
For example, babies start recognizing their mother very early on – first through smell, voice, and touch and subsequently sight as their vision improves.
Moving on, as infants grow into toddlers they observe patterns, objects, words, and situations (such as feeling hungry, hot or cold, and so on).
Learning and its application manifests itself in reaction to stimuli based on the vast repository of experiential data.
The brain interpolates the learning to formulate the response to a similar but new situation (new word, recognizing a different picture of the same object, and so on).
What we see is that learning and cognition need a large amount of data (experiential database) as a prerequisite to forming the cognitive function.
Machine learning and Artificial intelligence in computer systems resemble this methodology to achieve ‘intelligence’.
Learning in computer systems relies on a large amount of data. This data is used to train mathematical ‘models’ or discover meaningful patterns and insights that may not be discernable or obvious to the ‘human eye’ or be well beyond human capacity.
In essence, Machine Learning relies on Mathematical and Statistical methods to arrive at algorithms that are able to make data-driven inferences or decisions. These outcomes are then improved upon by feedback based on the performance of the algorithm or the model against new or unseen data.
Hence arriving at a solution for any Machine learning problem involves these phases.
- Gathering enough data to represent as many possible inputs (ideally all – however that is not usually feasible in real-world situations)
- Having enough known outcomes to compare the accuracy of the prediction or response
- Modeling or training of the system based on input and comparison of output with expected outcomes
In order to be able to arrive at a good model or prediction algorithm, data preparation is critical – the data representation (how the real-world data is represented in digital form), data volume (quantity), and quality (how well it reflects the real world) are critical.
Hence arriving at a solution for any Machine learning problem involves these phases.
- Gathering enough data to represent as many possible inputs (ideally all – however that is not usually feasible in real-world situations)
- Having enough known outcomes to compare the accuracy of the prediction or response
- Modeling or training of the system based on input and comparison of output with expected outcomes
In order to be able to arrive at a good model or prediction algorithm, data preparation is critical – the data representation (how the real-world data is represented in digital form), data volume (quantity), and quality (how well it reflects the real world) are critical.
Popular Machine Learning approaches
- Decision tree learning
- Association rule learning
- Artificial neural networks
- Deep learning
- Inductive logic programming
- Support vector machines
- Clustering
- Bayesian networks
- Reinforcement learning
- Representation learning
- Similarity and metric learning
- Sparse dictionary learning
- Genetic algorithms
Ref: Machine learning
Bad Data. Bad AI
Coming back to our infant analogy. It is said that the examples we set for our children make them the kind of adults they end up being. Similarly, a bad or incorrect training set for an AI system can play havoc with the results.
Microsoft learned this the hard way when its AI chatbot turned racist and started posting inflammatory posts. Its training model was based on data from Social Media Networks – guess humans are “bad parents” for AI systems!
Where AI can outshine human intelligence?
Human intelligence is unparalleled. AI systems only barely scratch the surface, when compared to capabilities of the human brain and cognition. However, there are areas where they outperform us.
For example, an AI system can be an excellent aid to Law firms. At law firms, para-legal teams spend a lot of time gathering data from past cases, judgments, and references for their lawyer teams to analyze.
This can sometimes take days, weeks, or possibly months! An AI system on the other hand can do this in a much shorter time and at a lower cost.
AI systems can minimize time spent on tedious mechanical activities so that more time can be spent on more complex decision-making and problem-solving.
It may happen, that jobs such as para-legal services or driving may be taken up by AI in the future. Yet AI isn’t ready to take over the world. Human emotion, intuition, and conscience are still something AI systems can only aspire to achieve someday.
(Artificial Intelligence Will Never Rival the Deep Complexity of the Human Mind)
Where to next…
In this series of posts, I would like to discuss
- How data gathering and preparation can be done for ML systems
- Some methodologies to arrive at prediction or model algorithms
- Limitations that ML methods have, how they may be exploited to break systems and what are the steps that may be taken to make ML robust, in order to have safer AI systems.
Next Step
Contact us if you need any help with ML/AI for software solutions.
