Framework for thinking about AI

Building AI only on data, might not be the right strategy for everyone. This is because:

1. AI built only on data, is dependent on collection of large amounts of harmonised/annotated data. This strategy works for FANG companies but not so much for startups.
2. On ground data can be dirty in terms of missing data/outliers/wrong practices. AI based only on data does not differentiate between good data and bad data. Ideally a sieve or filter should be applied on the collected data to establish its “cleanliness” and ultimately drop out data collected by “bad” or “incompetent” users/actors or just system integration issues. This process should be automated and should not involve someone looking at every instance of collected data
3. Some industries especially workflow driven industries like healthcare or logistics, lay down specific guidelines for actors/users on how to operate. These guidelines set bounds on rationality for actions by users. AI built only on data does not have any good way to incorporate these guidelines or domain expertise as part of prediction

Let’s take a step back and evaluate processes humans use to operationalise their intelligence and functionally break down the components.

Memory : Our history of perceptions of the world is stored in memory as facts. For example a fact that we all stored in our memory quite early is that touching a hot kettle can be painful.

Logic : We apply logic to deduce knowledge from facts. Logic bounds the scope of real world problem solving and learning for us. For example if you know that A is B’s son and C is B’s sister, you can apply transitive reasoning to deduce that A and C are related via aunt-nephew relationship. Note that the relationship between A and C was never stored as fact or known to you in advance. Humans are excellent at deducing knowledge about the world by evaluating facts

Learning : We are constantly bombarded with new data and new experiences. We apply logic on the new data and then push new experiences as neural pathways to memory, to be referenced again in future

Hence the ideal functional stack for building AI should look like below:

Knowledge Graph (Memory) : Build a system of representation and updation of knowledge. This is the most important step as choosing the right knowledge representation provides affordances for efficient and fast logical queries

Automated Reasoning (Logic) : Build a system that can do automated deductions or transitive reasoning (deduce relationship between entities, given no observation of direct relationship). Once this is in place, your system can drive deductions at scale across large repositories of knowledge. Pretty cool, eh !

Deep Learning and Machine Learning : This is where most of startups spend time — apply deep or machine learning to problem statements by looking at data. You will need a bridge to convert reasoning to vectorised format of data as machine learning algorithms expect vectors and matrices as inputs. Here is good way on how to implement this -> https://cs.stanford.edu/people/jure/pubs/graphsage-nips17.pdf. Ultimately as new insights are observed, they have to be pushed in the knowledge graph.

Benefits of instrumenting your AI stack this way:

1. Apply AI within the confines of bounded rationality for superior results
2. Use less data
3. Rate performance of actors based on real data and drive accountability within the organisation
4. Only use the right data for training AI

Building good AI can get lonely at times and I thought I must share my own learnings. Hope this was helpful.