Jarvis Rising – How Google could generate a machine learning model “on the fly” to predict answers when Search can’t, and how it could index those models to predict answers for future queries [Patent]

Key points from the patent:
Similar to my last post covering different caller Google patent, I deliberation the champion mode to screen the details is to supply bullets containing cardinal points.

Generating and/or Utilizing a instrumentality learning exemplary successful effect to a hunt request
US 11645277 B2
Date Granted: May 9, 2023
Date Filed: December 12, 2017
Assignee Name: Google LLC

1. Google’s patent explains that if an reply cannot beryllium located with certainty, and the idiosyncratic submits a petition that is predictive successful nature, a trained instrumentality learning exemplary tin beryllium utilized to make a prediction.

2. For example, Google could archetypal make hunt results based connected a query, but if the results aren’t of capable quality, past a instrumentality learning exemplary tin beryllium utilized to supply a stronger predicted answer. So, the strategy tin supply predicted answers based connected a instrumentality learning exemplary erstwhile an reply cannot beryllium validated by Google.

3. Also, the instrumentality learning exemplary tin beryllium generated “on the fly”, and Google mightiness store trained instrumentality learning models successful a hunt index. Yes, Google could scale instrumentality learning models that were conscionable trained to supply predictions based connected circumstantial types of queries. I’ll screen much astir this soon.

4. The patent provided an illustration based connected the query, “How galore doctors volition determination beryllium successful China successful 2050?” If an authoritative reply cannot beryllium provided via the modular hunt results, past the query tin beryllium passed to a trained instrumentality learning exemplary to make a prediction.

5. The patent goes connected to explicate that the strategy mightiness instrumentality different years similar 2010, 2015, 2020, etc. and usage those to make a prediction (via a instrumentality learning exemplary trained connected those parameters).

6. The patent explains that trained instrumentality learning models tin beryllium indexed by 1 oregon much contented items from “resources utilized to bid the model”. And for aboriginal queries, erstwhile the strategy identifies parameters that are related to a instrumentality learning exemplary (e.g. if a consequent idiosyncratic asks a related question like, “How galore doctors wherever determination beryllium successful China successful 2040?”), the instrumentality learning exemplary could beryllium utilized to make a prediction.

7. The patent goes connected to explicate that the instrumentality learning models could beryllium stored with 1 oregon much contented items, similar entities successful a cognition graph, array names, file names, webpage names, and more. In addition, words associated with the query similar “China” and “doctors” could beryllium utilized by the instrumentality learning exemplary to make a prediction.

8. The patent goes connected to explicate that the strategy mightiness supply an interactive interface for users to prime parameters that tin beryllium passed to the instrumentality learning model. That tin beryllium a substance field, a dropdown menu, etc. Also, the effect could see a connection presented to the idiosyncratic that the effect is simply a prediction based connected a trained instrumentality learning model. So Google wants to marque definite users recognize it’s a prediction based connected a instrumentality learning exemplary versus answers provided based connected information it has indexed.

9. The trained exemplary tin past beryllium validated to guarantee the predictions are of astatine slightest a “threshold quality”. Anything beneath a definite threshold tin beryllium suppressed and not provided to the user. In that case, the modular hunt results tin beryllium displayed instead.

10. Beyond nationalist hunt results, the patent explains that the strategy could beryllium utilized connected a backstage database to assistance companies foretell definite outcomes. The patent explains, “private to a radical of users, a corporation, and/or different restricted sets.” For example, an worker of an amusement parkland mightiness ask, “how galore snowfall cones volition we merchantability tomorrow?” The strategy could past query a backstage database to recognize income of erstwhile days, upwind information, attendance data, etc., to foretell an reply for the employee.

11. The patent explains that the strategy could provide propulsion notifications from an “automated assistant” astatine immoderate point. And conscionable reasoning retired loud, I’m wondering if that could beryllium from a Jarvis-like assistant similar I explained successful my station astir Google’s Code Red that triggered thousands of Code Reds astatine publishers. 

12. From a latency standpoint, the patent explains that determination could beryllium a hold aft a idiosyncratic submits a query. When that happens, the modular hunt results could beryllium initially displayed on with a connection that “good” results are not disposable for the query and that a instrumentality learning exemplary is being utilized to make a prediction. In those situations, the strategy could propulsion that prediction to the idiosyncratic astatine a aboriginal clip oregon supply a hyperlink for users to click to presumption the instrumentality learning output.

13. Also, the patent says for immoderate situations that the idiosyncratic would person to affirm the punctual successful bid for the process to continue. For example, the strategy mightiness supply a connection stating, “A bully reply is not available. Do you privation maine to foretell an reply for you?” Then the instrumentality learning exemplary would beryllium trained lone if affirmative idiosyncratic input is received successful effect to the prompt. Like I explained earlier, I spot a transportation with the “There aren’t large matches for your search” connection that rolled retired successful April of 2020. I’m wondering if that could grow to utilize this exemplary successful the future…