C2 Montreal’s annual artificial intelligence forum was co-hosted this year by Element AI, a local unicorn co-founded by deep learning pioneer Yoshua Bengio. The theme was Taking the Next Step.
At the May 24 morning AI Road-Mapping workshop, Element AI Senior Manager of Industry Solutions Richard Zuroff told a cross-industry audience that “traditional companies were never organized for AI.” So what did Element AI learn in its journey from research to commerce?
One problem to overcome was the definition of problems. Academically speaking, deep learning-based image and language processing techniques are easy to implement. Yet businesses still need to frame their problems around these technologies. Deep learning, for example, evolves very quickly and requires enterprises to deal with unfamiliar types of data.
Zuroff posted a problem which sought to “estimate repair cost from an accident.” If an insurance company uses machine vision to do the estimation, it needs to consider the types of photos to use, compare before and after accident photos, estimate required loss reserves, while identifying potential fraud from invoice mismatches.
In this case, artificial intelligence is not solving a single pre-defined problem, but three: identify potential fraud, estimate loss reserves, and deploy investigations.
Zuroff proceeded to show how difficult it can be to actually define an AI problem. He framed problems into the subsets planning, classification, and forecasting, and illustrated with the following fill-in-the-blank worksheet:
- Framing Planning Type Problems
- We have (a set of resources such people or equipment) that we deploy to (their main task or activity). We want to (the goal of deploying the resource e.g., meet a required service level) while minimizing/maximizing (name a trade-off if there is one) but overall conditions require that (one or more requirements of your operating environments e.g., there are at least 2 nurses on duty at all time).
- If we did a better job planning how or where we deployed our resources every (a time scale such as second or day or week) then we could improve (the main business metric to measure success) and the experience of (the main stakeholder, such as employees who will benefit).
- To plan, we can collect information on data sources for the expected demand for the main task or activity, as well as data sources for the expected supply of the resource, and data sources for changes in constraints or conditions.
- Framing Classification Type Problems
- There are different types of (an entity of interest, such as customers, products or transactions) that we would like to divide into groups such as (a high-value group) and (a lower-value group). However, when presented with an example, it is difficult to do a good job assigning it to a group because (a barrier to successful sorting), which is causing the business to not (an action or decision that could be done better) as well as we would like.
- if we used (one or more datasets that are relevant to the classification) to capture characteristics like (features that are present in one group but not others that could help with sorting).
- We could better assign items to the right group, which would improve our (a business metric that would be used to measure the success of the classification) and provide a better experience to (a group of people who would be impacted).
- Framing Forecasting Type Problems
- We need to predict what (a value or status or state) will be in a (time scale such as second or week or year so that we can an action was taken based on the forecasted value, status or states). However, it is currently too (complex/data-intensive/fast) for us to rely successfully on simple prediction rules or human intuition.
- If we could make the prediction (more accurately/faster/further ahead in time) then we could better (a downstream decision that could be improved), which would improve the experience of (a stakeholder impacted by the prediction and increase our a business metric for success).
- To make the prediction we would use (name one or more numerical datasets) which is collected (a time scale such as second or week or year) and could also use (name one or more unstructured data sources such as text or images).
Founded only 18 months ago, Element AI raised US$102 million in Series A funding last year. The company expanded to 300 people in offices in Canada, UK, Singapore, and Korea. Element AI told us that it has also pinned down a business strategy for enterprises in the logistics and finance industries.
Running concurrent to Element AI’s workshop, C2 clustered partners from the Quebec Ministry of Economics, Science and Innovation (MESI), Microsoft, IBM, Montreal-based blockchain research centre Catallaxy, Invest in Canada, Montreal International, Deloitte, and local AI startup Stradigi.ai, to name a few.
C2 ran May 23-25 at the historic Arsenal building in downtown Montreal.
Journalist: Meghan Han| Editor: Michael Sarazen
0 comments on “Element AI’s Fill-in-the-Blanks AI Worksheet”